https://orcid.org/0000-0002-8978-0095
Abstract
Has there been an inflation in crisis coverage in newspapers over the last centuries, and if so, what structural factors drive this change? We utilize computational text analyses along with our own signal detection algorithm to measure the presence of crisis keywords and the emergence of crisis news waves. An analysis of crisis coverage in The Times (U.K., 1785–2020, 183,239 news stories) shows that the share of coverage that uses crisis keywords has increased, though not steadily. The number and salience of crisis news waves tied to discernible events has increased at a slower pace. The hypothesized driving forces—government expansion, mediatization of politics, and the activity of crisis frame sponsors—explain the development well and allow accurate predictions even when ignoring time in the forecasting model. Crisis coverage seems to reflect not so much the problems society faces, but society’s identity, priorities, and outlook on the world.
Inflation of crisis coverage?
Public crises are exceptional situations that severely threaten something of public value and warrant intense public attention (Boin, 2009; Hase & Engelke, 2022). The public sphere (Habermas, 1989) appears to be increasingly wrapped up in crises (Hoëm, 2017; Pantti, 2018)—it seems to stumble from one crisis to the next, never or only briefly recovering into a normal state. Scholars, pundits, and policymakers have diagnosed that we live in an “age of crisis” (Saad-Filho, 2021) or in a state of “perpetual crisis” (Beckett, 2019).
There is widespread skepticism regarding the functionality of attention management in the public sphere: Scholars across disciplines agree that crises are socially negotiated and constructed (Estes, 1983; Hoëm, 2017). The liquidity of the term raises the suspicion that the abundance of crises around us may reflect an “inflation” of public crises and ensuing crisis coverage, artificially created by a political and a media system where labeling something as a crisis is a promising policy-making device (Boin, 2009; McBeth et al., 2013; Nord & Olsson, 2013). Consequently, public attention management during crisis may be headed towards dysfunctionality (Estes, 1983). Overusing crisis labels may induce issue fatigue, which in turn threatens to diminish trust in news and journalism (Gurr, 2022) by violating the expectation that news media appropriately select and signal which news stories warrant much attention (Kohring & Matthes, 2007). Nabi and Prestin (2016) also demonstrate that sensationalism can damage news’ credibility. If society is constantly wrestling with some insubstantial crisis, this will impede the ability to tackle social issues. There is a risk that the media may become the boy that cried “wolf!” (or: “crisis!”).
In an analysis of inflation of crisis coverage, it is important to distinguish two different types of inflation: (a) inflation of mere usage of crisis rhetoric about various issues without much resonance, and (b) inflation of the instances of intense waves of coverage about a specific memorable crisis that mobilize public attention, exert pressure on decision-makers, and even have the power to shape collective memory and transform society.
Motivated by recent specific crisis, there has been a hike in interest in the social construction of specific crisis in the news (e.g., Hase & Engelke, 2022; Parks, 2020). These studies provide invaluable insights into the constituent elements in public crisis discourse in specific thematic contexts. However, scholars emphasize the fractured nature of inquiry into crisis construction (Udris, 2019). Ultimately, we know very little about the general rate and frequency of public crisis discourse and its long-term trajectory, beyond subjective impressions and anecdotal evidence (Boin, 2009; Hoëm, 2017; (Holton, 1987), and research would benefit from a more holistic approach to studying crisis (Udris, 2019).
It is up to communication researchers to provide the data and the models that allow a robust reflection and discussion about the way public crisis discourse unfolds in the public sphere, debating a possible inflation of and a possible dysfunction of public crisis discourse. Such a debate would benefit from understanding which systemic factors drive the long-term trajectory. Therefore, this paper asks: How much public attention is allocated to crises, how has this developed over time, and which factors drive this development? We theorize that it can be traced back at least partially to developments in the media and/or the political system.
Analyzing historical newspaper coverage allows us to virtually go back in time and analyze expressions of the public experience of crises far beyond what survey research could offer (e.g., Page & Shapiro, 1992). Our study applies a variety of automated content analysis techniques to assess the extent of crisis coverage in The Times (London) 1785–2020. The Times is a key publication for tapping crisis discourse in the British public sphere. Our study goes beyond previous research on related questions by (a) stretching the time horizon back until 1785, (b) by moving across crisis domains, (c) by conceptually and operationally distinguishing two different modes of crisis coverage, (d) by developing a model to explain variations in crisis coverage, and (e) by building a new methodology with superior validity.
As such, our historical analysis puts into perspective how far the corona virus disease 2019 (COVID-19) pandemic (and other recent crises) are expressions of us facing an “age of crisis” (Saad-Filho, 2021) or an “unprecedented crisis” (The World Bank, 2020) in our information environment or whether we see business as usual.
Conceptualizing crisis coverage
What is crisis coverage?
Crises are one of these things that everybody thinks they recognize when they see it. In accord with many crisis scholars (Hoëm, 2017; Udris, 2019), we take the opposite view that, constrained by what can be justified given current observations and societal values, it is the observer(s) who decide what they perceive as a crisis and what not. Sometimes it appears natural to call something a crisis, and the labeling enjoys public consensus; other times, crisis labeling is subject to intense and contested social negotiation (Parks, 2020) or outright rejection. In either case, it is not a simple logical application of fixed criteria that have a pre-determined result (Boin, 2009); they are not objective givens (Hoëm, 2017; Pantti, 2018). In newspapers, authors, editors, and sources decide what they view as a crisis, second-guessing the resonance among the audience. The term “crisis” is fluid. Its usage can change, it can be negotiated, and it can be subject to crisis framing (Estes, 1983; Hoëm, 2017), rendering it an object of political instrumentalization (Balzacq et al., 2016). We are interested in how often and how intensively the public collectively experiences societal events or conditions as crises, and switches into “crisis mode.” Given the media’s power to put issues on the public agenda (Geiß, 2019, 2022a, b) and frame them in terms of crisis, we turn to crisis labels that are being used in newspaper coverage to tap crisis discourse. This has two major implications:
This study is not interested in what specific persons, professions, or organizations privately/internally view as or treat as a crisis unless it is ascribed public value. We therefore do not review studies of, for example, personal crises in developmental psychology (e.g., Robinson et al., 2017), or organizational crises in PR research (e.g., Coombs & Tachkova, 2022).
This study uses an open-ended approach to defining crisis coverage. We avoid to a priori define away meaningful public crisis discourse just because it does not match modern patterns of perception or expression. After all, notions of crisis, social values, and crisis rhetoric have changed during more than 200 years (Habermas, 1973). Rather, our working definition of crisis serves to make sure that the function of crisis coverage is equivalent across time. This functional core is that a crisis is a markedly non-normal situation and constitutes a severe threat. This justifies extraordinary attention and unusual, even extreme reactions to combat the crisis, rendering public crises a “game changer” (Loorbach et al., 2016) in policymaking. This functional definition excludes metaphorical, joking, or explicitly non-public mentions (e.g., personal crises). It includes other terms than crisis which also imply severe non-normalcy and threat (such as catastrophe or disaster) (Pantti, 2018; Winkler, 2014). Secondary criteria that tend to occur with heightened frequency in crisis discourse (but are not necessary conditions) are surprise/unexpectedness, urgency, external causation, loss of control and calls for a united reaction (Pantti, 2018; Winkler, 2014). We rely on the depiction expressed in an article rather than re-appraising the situation ourselves.
This notion of crisis coverage lends itself to identifying crisis coverage using keywords. Journalists who report about a crisis-like situation will usually in a salient way use at least one crisis keyword (e.g., “crisis,” “disaster,” “pandemic”) to draw the audience’s attention (Pantti, 2018). It is technically possible to convey the image of a highly threatening and non-normal situation without using crisis keywords, but using crisis keywords will be the almost universal way to evoke the impression that something constitutes a crisis.
Escalation states of crisis coverage
The open-ended and data-driven nature of our definition of crisis coverage entails that not all crisis coverage we discover corresponds to the prototypical public crisis situation that receives enormous attention and shapes collective memory, like during the COVID-19 crisis. We capture a lot of coverage where one journalist or one source calls something a crisis but without much resonance. We thus distinguish between three kinds of coverage that indicate three different escalation states of crisis discourse. Any issue, event, or condition in society (critical condition, or CC) can potentially be called a crisis and may at any time be subject to any of the three kinds of coverage, without presuming any fixed serial order, typical career or irreversibility.
Articles about a CC that include no crisis coverage at all; in operational terms, they do not mention any crisis keywords in salient spots. This set of articles represents routine or non-crisis coverage. CCs that only receive routine coverage are in a latency state. We filter out routine coverage by searching for crisis keywords.
Articles about a CC that include some crisis coverage, but often with limited resonance; most of it is not part of a major cumulation of crisis coverage about the same issue. This set of articles represents crisis labeling (CL) or crisis rhetoric coverage. If a CC only receives routine and CL coverage, it is in the CL state which is unlikely to act as a political game changer. CL indicates the attempt to frame a condition in terms of it being a crisis. The salience of CL is operationalized as the share of CL coverage relative to all coverage.
Articles that include crisis coverage and yield massive resonance by becoming part of a major cumulation of crisis coverage about the same CC (Van Atteveldt et al., 2018): a crisis news wave (CNW). This set of articles represents CNW coverage. If a CNW state lasts, it has a high potential to shape collective memory (Kligler-Vilenchik, 2011; Tenenboim-Weinblatt, 2013) and serve as a game changer. CNWs will occur rarely because they require continuous and consistent crisis coverage about an issue. To differentiate between CL coverage and CNW coverage, we identify issue-specific CNWs (Geiß, 2018; Viehmann & Geiß, 2022); all crisis coverage about an issue that occurs during an active crisis news wave about the topic is classified as crisis news wave coverage; this means that CNW coverage is a subset of CL coverage. The salience of CNW is operationalized as the share of CNW coverage relative to all coverage. In addition, since each CNW has the potential to serve as a game changer, we look at the count of CNWs in any given year, where an increasing count could indicate an inflation of crisis coverage.
A collectively experienced societal crisis that finds expression as a CNW can be a powerful political tool that strategic actors will try to use: A CNW can increase attention for an issue, mobilize resources, streamline and speed up decision-making; consonant and cumulative crisis framing can sway public opinion, public policy and help reaffirm or renegotiate the resource distribution (Loorbach et al., 2016), combining the “politics of threat […] with that of threat management”, and with performative power to “transform social reality” (Balzacq et al., 2016, p. 495), but only insofar as it remains selective.
Hence, an inflation of crisis claims in public discourse would have serious repercussions in society. To be functional in attention management, the share of CL coverage and/or CNW coverage compared to all other coverage must be low. Hence, it is useful to monitor the long-term development of the share of crisis coverage relative to all other coverage. A long-term historical perspective can inform us whether the current share of crisis coverage is unusual or not, providing a starting point for discussion and further investigation.
The three escalation states (Figure 1) reflect how broadly the notion of a crisis has permeated public discourse in a certain case at a certain time (up to the point of omnipresent publicity of a CNW). These stages also reflect how news outlets can filter out crisis claims and avoid an inflation of crisis coverage, reflecting the high threshold of public attention (Neuman, 1990) and the small issue carrying capacity of the public sphere (Hilgartner & Bosk, 1988). The distinction facilitates analysis of the negotiable and gradual nature of crisis and the politics involved in the social construction of crisis.
Drivers of crisis coverage
Distinguishing the escalation states of crisis publicity enables us to register two different kinds of crisis coverage—CL and CNW—and track their development over time relative to routine coverage. In the next step, we theorize which drivers of crisis coverage can explain these long-term trajectories: which major societal trends may inflate or deflate CL and/or CNWs? We focus on the major structural long-term trends in how a society’s public sphere reliance on CL and CNW develops. We look beyond the impacts of single authors or news outlets, single critical events and conditions, and other specific determinants that explain why a certain article includes some crisis rhetoric or is part of a CNW. By looking at developments in the long run, across various issues and contexts, and at high levels of aggregation, we reduce the noise these data induce when we look for structural factors.
Explanatory framework
We move the perspective beyond the diffuse expectation that crisis coverage has increased over time; we rather seek to identify which processes of long-term structural social change affect the extent of crisis coverage. This provides a much richer and more generalizable explanation that tells us which developments produced the change in crisis coverage in the public sphere. To find such general explanatory factors, we look to changes in the structure of the state and the public sphere rather than timelines of external events or changes in conditions. In our social constructivist perspective, any observations or records of reality are infused with values and habits of perception that limit comparability across time and societal domains. Arriving at a consistent and sufficiently valid summary indicator of the severity of CCs in a society (a measure of society’s “crisis-strickenness”) has proven impossible to us, despite having collected, for example, vast amounts of data on economic performance, epidemics, disasters, and international conflicts (Supplementary Appendix A, section 1.4).
Our explanatory model builds on four fundamental building blocks: the public arenas model (Hilgartner & Bosk, 1988), the policy dynamics model (Baumgartner & Jones, 2002), the thermostat model (Soroka & Wlezien, 2010), and the mediatization of politics phase model (Strömbäck, 2008). These theories share our main idea to focus on structures of the state and the public sphere to explain structures and dynamics in the public sphere, and treat external reality as a noisy constant. They rely on the notion that social problems are socially constructed in competitive public arenas. “[S]ocial problems are projections of collective sentiments rather than simple mirrors of objective conditions in society” (Hilgartner & Bosk, 1988, pp. 53–54). Hence, theories that rely on objective conditions fail to “[…] explain why some conditions are defined as problems, commanding a great deal of societal attention, whereas others, equally harmful or dangerous, are not” (Hilgartner & Bosk, 1988, p. 54). The same is true for being defined as a crisis or not. Various organized social interests (as well as political decision-makers from various parties) compete to prioritize “their” issues and define issues the way they prefer. These advocacy groups want the government to spend more money on specific issues, influence how the money is spent, and how the policy area should be regulated (Baumgartner & Jones, 2002). The news media are the main public arena to disseminate this elite discourse to the public (Hilgartner & Bosk, 1988).
However, the news media are not simply mirroring this discourse, either. With increasing mediatization of politics (Mazzoleni & Schulz, 1999), they apply their own logic of selection (Hilgartner & Bosk, 1988). This media logic grows the more important the more the mass media can reach a mass audience and the more political-ideological autonomy they gain (e.g., when dissolving organizational and ideational ties with parties). At full strength, policymakers and advocacy groups internalize media attention criteria as a complement or substitute for political attention criteria (Strömbäck, 2008). Political actors increasingly depend on the news media as pervasiveness of media logic in society grows.
Political autonomy renders the news media dependent on generating revenue on sales and advertising markets (Strömbäck, 2008). This economic dependency finds expression in a commercial media selection logic that seeks to maximize audiences and attention, for example, by preferring news that feature event-driven dramatic narratives that involve powerful persons and institutions, conflict and controversy, damages and threats, novel developments and surprises, so-called news criteria (Harcup and O’Neill, 2017 ; Hilgartner & Bosk, 1988).
Crisis narratives perfectly fit many of these news criteria (Pantti, 2018), promising to reliably draw intense public attention: calling something a crisis clearly communicates that a situation is threatening and non-normal, usually demanding urgent and decisive action. Given the mushy definition of crisis, strategic actors can use crisis rhetoric to stimulate CL (McBeth et al., 2013; Nord & Olsson, 2013) and sometimes even a CNW which opens a window of opportunity to re-shape government policy (Loorbach et al., 2016).
Crisis frame sponsor activity, public spending, and pervasiveness of media logic
Crisis frame sponsors
We consider the activity of and competition between advocacy groups for public attention as the primary (and immediately antecedent) causal mechanism behind the variation in the salience of crisis coverage: They publicly make claims that something constitutes a crisis. Crisis claims draw the attention of journalists, audiences, and policymakers. More advocacy groups will make such claims the more promising this strategy is to attain political influence and achieve the advocacy groups’ goals. Greater advocacy group activity to stimulate CNWs should leave traces in news coverage. Qualitative insights in the literature on public relations and lobbying support the idea that the lobbying and public relations industry in the U.K. has grown larger, more topically diverse, and more professionalized over the course of the 20th century (Jordan et al., 2012; L’Etang, 2004; McGrath, 2018). We argue that advocacy groups’ strategy to act as crisis frame sponsors becomes the more intense and more prevalent the more the following conditions are fulfilled: (a) Policymakers depend on popular consent; there is (b) intense and (c) diverse public spending, and there are news media with (d) high reach and (e) high autonomy.
Dependency on popular consent
The more policymakers depend on the consent of the majority, the greater the responsivity to public opinion (Soroka & Wlezien, 2010). Historically, the percentage of adults with suffrage in the U.K. increased stepwise from roughly 3% (1785) to 100% (1927 and later) with voting rights reforms 1832, 1865, 1867, 1884, 1919, and 1927 (Supplementary Appendix B). This is linked with more systematic taxation of the entire citizenry (Cronin, 2015) such that more citizens became concerned with public affairs to monitor what was done with the taxes they pay. With universal suffrage and taxation comes wider interest in public affairs, more news use (driving media penetration), and demand for spending in new—elite-neglected—policy areas (driving public spending diversity). We view universality of suffrage as a precondition for the changes in public spending and in pervasiveness of media logic rather than considering it as an independent factor of influence.
Public spending
Public spending intensity
Influencing government spending decisions is a great prize (Soroka & Wlezien, 2010) that advocacy groups will compete for (Baumgartner & Jones, 2002; Dür & Mateo, 2016). As public spending grows, competitors will invest more resources to influence the allocation of public funding, and the greater the likelihood that their crisis rhetoric pass the CL or even the CNW filters. Historically, public spending in the U.K. has increased between 1785 and 2020, both in absolute (inflation-adjusted) terms and relative to GDP. Spending relative to GDP increased primarily (and mostly linearly) between the 1880s and the 1980s, with major spending spikes during wars (Supplementary Appendix B) (Cronin, 2015).
Public spending diversity
With growing economy and tax revenue, the government can get involved in more and more thematic areas (Baumgartner & Jones, 2002). The greater the topical bandwidth of public spending, the more different interests from more different domains in society are motivated to organize and compete for public spending, and to make crisis claims to that end. The more crisis claims are made and the more professional the communication efforts, the greater the likelihood to pass CL (and possibly also CNW) filter(s). For example, if there is no welfare state and poverty is not a politicized problem, there will be few social welfare organizations with PR activity. Historically, public spending diversity mainly increased in a relatively short time span 1865–1925 in a stepwise fashion; the major wars led to a temporary strong concentration on military spending (Supplementary Appendix B).
Pervasiveness of media logic
Media penetration
The more the news media become the main channel through which the citizenry receives their political information, the more promising stimulation of crisis coverage becomes as a pathway to exert pressure on policymakers (Strömbäck, 2008) ; media become a pressure point for advocacy groups (Dür & Mateo, 2016). Historically, newspapers in the U.K. grew in terms of their aggregate circulation until reaching a saturation point in the early 1950s. Since then, political news has been universally accessible in the U.K., with a shift from newspapers to TV and Internet news (Supplementary Appendix B). (Strömbäck, 2008)
Media autonomy
The more the news media make their news judgments according to news values and expected audience attention, independently of political loyalties (Strömbäck, 2008), the more freely can and must advocacy groups compete for the news media’s attention, and using crisis rhetoric is one promising tool advocacy groups will employ to stimulate publicity. Historically, the news media in the U.K. have enjoyed high autonomy from a legal perspective (little censorship and government intervention) and regarding obvious indicators of press–party parallelism (such as a strong party press). Still, scholars point out that the U.K. has had a strong press–party parallelism in the past (Hallin & Mancini, 2004); it manifested in strong informal party loyalty of newspapers because their readers expected party loyalty from “their” newspaper. Voter party loyalty has softened considerably since the 1970s (Supplementary Appendix B), freeing news organizations from party loyalty and reducing the influence that parties exert on the news organizations (Hallin & Mancini, 2004; Seymour-Ure, 1996).
The more these conditions are satisfied, the greater the incentive for social interests to organize as advocacy groups (Jordan et al., 2012), increase their activity and competition for public attention and approval (McGrath, 2018) and professionalize their public relations activities (L’Etang, 2004). This will result in more CL (and possibly more CNWs). More and more advocacy groups see a realistic chance to exert public pressure and influence government policy such that more “tax money” is invested in their area of interest (Cronin, 2015, p. 5).
Research questions and hypotheses
Not only the widespread popular assumption, but also revisiting the major historical developments we regard as relevant, suggests an increase in crisis coverage. H1: CL salience (H1.1), CNW salience (H1.2), and CNW count (H1.3) have increased 1785-2020.
We expect that change in crisis coverage reflects three closely related developments: pervasiveness of media logic, public spending structure, and crisis frame sponsor activity.
H2: Crisis coverage (CL salience, CNW salience, CNW count) is predicted by (H2.1) the structure of public spending (intensity, diversity); (H2.2) the pervasiveness of media logic (media penetration, media autonomy); and (H2.3) the activity of crisis frame sponsors.
Our explanatory model would suggest that public spending and pervasiveness of media logic primarily affect the activity of advocacy groups to act as crisis frame sponsors, which would therefore mediate the impact on crisis coverage.
H3: Crisis frame sponsor activity/competition partly or fully mediates (H3.1) the effect of pervasiveness of media logic on crisis coverage and (H3.2) the effect of structure of public spending structure on crisis coverage (CL salience, CNW salience, CNW count).
One might argue that it is simply coincidence of historical developments that both crisis coverage and our drivers simply increase over time. To rule out this explanation, we would want to make time superfluous and figure out how well the theorized mechanisms alone account for changes in salience of crisis in news coverage.
RQ1: Can the drivers of crisis coverage predict the development of crisis coverage well enough that they can substitute time as a predictor altogether?
CL and CNW
When crisis claims have to pass CL and CNW filters, we assume that the effects of the drivers would be more immediate at the CL escalation state but only in a diminished degree translate to the CNW escalation state Passing the CNW filter is more difficult because, first, there are two filters (the CL and the CNW filter) relevant for CNW but only one filter (the CL filter) relevant for CL. Second, the CNW filter and the CL filter are calibrated differently. The CL filter allows many “false alarms” (ie crisis diagnosis that later turn out to be overstated) which are then subjected to public scrutiny. The CNW filter allows few “false alarms” (i.e., full-blown crisis news waves, where the central crisis claim later turn out to be insubstantial) because such false alarms would consume a lot of public attention, a scarce resource (high specificity filter). This would induce high social costs due to not putting more important and pressing problems into the public spotlight.
In practice, this stricter CNW filter results from competition and checks and balances between advocacy groups and between news outlets. Even if a crisis frame sponsor wants to escalate an issue to the CNW state and individual news outlets strongly use CL in the hope to produce a CNW (possibly amounting to a scoop), other news media will usually try to avoid that other news media are credited with a scoop. They have an incentive to ignore or oppose the CL, unless the situation—in conjunction with the society’s values—leaves them no other choice. If there are reasonable doubts that an issue actually is serious enough to be called a crisis, competing advocacy group will provide news media with facts and arguments to counter unjustified/contestable crisis claims.
H4.1: The over-time development suggested in H1.1-H1.3 will be more pronounced for CL than for CNWs.
H4.2: The driving mechanisms assumed in H2.1-H2.3 will be more effective in driving CL as compared to CNWs.
Figure 2 provides an overview over the hypotheses and the research question.
Method
Media sample and time period
The study draws on a full-text archive of newspaper coverage in The Times (London) between 1785 and 2020 (Gale 1785–2014; Factiva 2015–2020). Few other news outlets provide a fully digitized archive of their historical news coverage, fewer cover a comparably long period of time, and even fewer play a comparable role in society. Given the The Times’ influential position close to political power in the U.K. (Bingham, 2013; Liddle, 2016; Nevins, 1959), it will allow us to reconstruct the major crisis discourses that occupied British politics and the British public, much like (Jucker and Berger, 2014) used The Times to reconstruct habits of quotation in broadsheet newspapers. Its coverage is not necessarily representative of the newspaper landscape of the U.K. at large (Hobbs, 2013); but that is not the goal of our study. Any extension of the newspaper sample would substantially shorten the period of study or lead to a convenience sample of newspapers (any sample of historical newspapers will be skewed in the direction of newspapers that are considered high in prestige and importance, precious enough to archive). We rely on The Times for all main analyses and hypothesis tests. We carefully consider the historical development of The Times (see Supplementary Appendix C) to identify potential idiosyncrasies that reflect our design. For exploring generalizability to the prestige press, we consult keyword searches in The Guardian (left-leaning prestige newspaper) (Supplementary Appendix D); to map even broader generalization potential, we analyze full-texts in The Economist (weekly, U.K., 1843–2014), Neue Zürcher Zeitung (NZZ, 1780–1996, Switzerland), and Washington Evening Star (U.S., 1853–1962) (Supplementary Appendix E).
Total coverage sample
The total coverage of the Times was not downloaded but the volume of coverage per year was recorded based on searches with a blank search term, returning all articles included in the archive.
Crisis coverage sample
We started with only the search string “cris!s”, expanding and validating the search string further in a systematic procedure (see Supplementary Appendix F). The following resulting search string was used to find articles that label conditions as a crisis or a related critical status: “cris!s” OR “disast*” OR “catastrophe*” OR “pandemic*” OR “epidemic*” OR “recession*” OR “breakdown*” OR “collapse*” OR “debacle*” OR “emergency” OR “emergencies”. Articles were included in the The Times crisis corpus if any of the terms occurred in the headline or in the first 100 words of the article. This resulted in a total of 187,573 hits. After removing 4,334 articles due to incomplete metadata (e.g., missing date) or because they were duplicates, we arrived at 183,239 articles in the crisis corpus with CL. We assessed the performance of the keyword search against two human coders who rated the intensity of crisis rhetoric in 50 articles (25 from the crisis corpus and 25 from a random sample or articles from The Times), resulting in a good F1 score of .889. The classification performance metric is constant across different 50-year time slices (Supplementary Appendices F and G).
Text analysis procedures
Corpus preprocessing
After standard preprocessing routines (removal of stopwords, special characters, and punctuation; spellcheck; see Supplementary Appendices H and J), we built two different corpora: (a) a core corpus that includes only the words that appear in at least 100 different documents was used for Structural topic modeling and (b) a full corpus that includes the entire vocabulary was used for named entity recognition (NER).
Preparing CNW detection: structural topic modeling
The purpose of the structural topic model (STM) is to create topic-specific time series that allow finding issue-specific CNWs (and prevent a CNW concerning one topic to be compensated by a simultaneous decline in crisis coverage about other topics). We ran the STM (Roberts et al., 2019) on the core corpus of crisis coverage (183,239 documents with a 5,293 word dictionary) with the year as a covariate. Checking the diagnostics for STMs (see Supplementary Appendix K, section 10.1), we concluded that extracting 250 topics would be sufficient; this rather high number of topics was expected given the 235-year time frame of the study and the wide range of topics the corpus covers. We compared manual coding of topics from the crisis corpus to the topic modeling results (making them comparable by aggregating both results to 20 broader topic areas), yielding satisfying to good F1 scores: (a) economic crises (.713), (b) disasters and accidents (.754), and (c) epidemics (.803). The rare occurrence of the other topic areas did not allow assessing F1 scores for them (Supplementary Appendix K, section 10.4).
Crisis news wave identification and extraction
We constructed day-by-day time series for all 250 STM topics (days without any reporting were scored as “0”) and ran a time series analysis that detected and extracted phases of extraordinarily high salience for that topic in crisis coverage. If such a spike in attention surpassed a threshold (defined relative to the total amount of coverage in the given year), it was recorded as a CNW (each of which belongs to one of the 250 topics). All coverage that belongs to the topic and falls into the time frame of the CNW is interpreted as belonging to that crisis news wave (Supplementary Appendix L, sections 11.1–11.4). A human coder used a list of words typical of the CNW, headlines characteristic of the CNW, as well as the dates and topic of the CNW to assign a random sample of 300 CNWs to specific real-world events, succeeding in 287 cases (96%). There was high agreement between two coders who labelled the same 50 CNWs (96% agreement), and the two coders agreed that almost all CNWs were linked to a major societal crisis (Supplementary Appendix L, section 11.5).
Preparing the advocacy group indicator: Named Entity Recognition (NER)
We use the crisis corpus to construct a measure of crisis frame sponsor activity, as no external data with a sufficiently long time frame and temporal resolution exist; we took care to make sure that this measure is disentangled from the dependent variables and risk of circularity (common dependency of the dependent and independent variables) is minimized (Supplementary Appendix M). We ran named entity recognition on the full corpus using the spacyr package in R (Benoit & Matsuo, 2020), and extracted per document all entities that were recognized as “organizations” (or ORG entity). We then removed all entities that occurred only once, which removed many of the OCR errors that were falsely assumed to be named entities.
For validation, five coders manually coded organizations mentioned in 242 articles from the crisis corpus and coded how the organization contributed to crisis discourse (agreement in assigning actors to crisis-related roles: κBrennan&Prediger = 0.95). We found that 78% of the organizations mentioned in news coverage were engaging in crisis discourse. The diversity of organizations participating in crisis discourse increased in 1950–2020, compared to 1785–1949 (Supplementary Appendix G). We aggregate the number of unique organizations that appear in the crisis coverage of an entire year. If a more diverse set of organizations appears in a year, probabilistically, there will also be a greater and more diverse set of organizations that engage in crisis discourse. Aggregation will eliminate the noise resulting from the fact that not all organizations that appear in crisis coverage also contribute to crisis discourse.
Independent variables
Pervasiveness of media logic
We use the share of respondents from the British Election Study without any party identification as a proxy indicator of media autonomy (Supplementary Appendix B). Media penetration is measured as the highest of three media penetration scores that range between 0 (no penetration) to 1 (full penetration): Newspaper, TV penetration, and Internet penetration (see Supplementary Appendix N for calculation and full sources).
Government spending in the U.K.
Data on government spending are available from official sources (Central Statistical Office, Public Expenditure Statistical Analysis) and from a historical statistics compilation (Mitchell, 2011). The data were compiled by and retrieved from a private website (Chantrill, 2023) whose data we spot-checked against the official data repositories and found to be accurate. The GDP data were retrieved from Ryland and Williamson (2021).
Government spending intensity
Government spending intensity was calculated by dividing the total amount of spending by the total GDP in year; this indicates how voluminous state activity can be regarded as compared to the total economic activity (Supplementary Appendix B).
Government spending diversity
We calculated the yearly share of each budget category (total government budget equals 100%), computed the Gini coefficient for the budget composition of each year, and subtracted it from 1. The resulting indicator would be 1 when spending is equally distributed across all budget categories; it would be 0 if spending is concentrated in just one spending category (Supplementary Appendix O).
Unique organizations in crisis coverage per year
From the NER annotation table, we extracted all organizations that were identified in each article, listed all organizations that appear in that year and thereby created a count of the number of unique organizations that appeared in crisis coverage each year. We removed all unique organizations that appeared only once (this was mainly done to remove text fragments that erroneously were categorized as organizations by the algorithm), resulting in the number of organizations mentioned in a crisis context in a year (organizations count). The density of organizations mentioned in a crisis-context was calculated as: 1,000 times the number of unique organizations mentioned in CL that year divided by the total number of articles. This measures the density of organizations involved in crisis in that year (unique actors per 1,000 articles), many of which will be advocacy groups engaged in crisis framing.
Year
We calculated the number of years elapsed between 1785 and the article’s date.
Dependent variables
From the news article count, we created a yearly time series of how many articles appeared in The Times. From the crisis corpus, we created a yearly number of articles which contain crisis labeling (CL articles). The CL salience indicator is defined by CL article count divided by the total article count in that year. From the crisis corpus and the register of CNWs, we created a yearly time series of full-articles-equivalents that belong to any crisis news wave (CNW articles). The CNW salience indicator is calculated as the share of total articles that are CNW articles. From the register of CNWs, we created a yearly time series of the number of CNWs observed in the given year. As there were count data, we logarithmized the variable to enable linear regression analysis.
Data analysis
We test H1 by analyzing the development of crisis coverage over time. To test H2, we regress crisis coverage on the structure of public spending (H2.1), pervasiveness of media logic (H2.2), and crisis frame sponsor activity (H2.3) as drivers in the development of crisis coverage over time (see Supplementary Appendices P and Q for more details on these analyses). Our tests for autocorrelation and non-stationarity show no statistically significant autocorrelation or non-stationarity (Supplementary Appendix R). To test H3, we combine the models of crisis coverage (CL salience, CNW salience, and CNW count) with a model that explains the potential mediator, crisis frame sponsor activity. Based on these models, we conduct causal mediation analysis (Imai et al., 2010; despite its name, the procedure requires experimental designs to establish causality; we employ it to establish indirect relationships rather than making causal inferences) to get estimates of direct/unmediated as well as indirect/mediated effects of structure of public spending and pervasiveness of media logic. H4 is tested based on the models of CL salience and CNW salience. However, we use different ways of standardizing the coefficients and computing effect sizes. The statistical procedure for exploring RQ1 is explained in the results section because it is iteratively linked to the results. Robustness tests using four other newspapers are conducted for H1.1–H1.3 (and for all analyses with CL salience as dependent variable (with The Times and The Guardian)). Between observed values in the time series, missing values are imputed using last observation carried forward; for observation points before the time series starts, we impute values by next observation carried backward, unless stated otherwise.
Results
Development of crisis coverage (H1)
CL salience in The Times increased over time. Using a linear model (in which intercepts below 0 were not allowed), we obtain a yearly increase of 0.012 percentage points per year (SE = 0.0003; t(235) = 45.5; p < .001). The estimated CL salience was 0% in 1785 while it was 2.8% in 2020. The joint results for the The Times and The Guardian leads to the same result, reflecting the synchronicity in the CL salience time series of the two newspapers (Supplementary Appendix D). Even repeating the analysis with a set of four diverse newspapers (The Times, Washington Daily Star, NZZ, The Economist) leads to similar results (Supplementary Appendix E). Overall, the data strongly support H1.1.
CNW salience in The Times increased over time. Our linear model estimates a yearly increase of B = 0.0014 percentage points per year (SE = 0.00018; t(235) = 7.79; p < .001). The estimated CNW salience was 0% in 1785 while it was 0.33% in 2020 (Figure 3). The results do not replicate when repeating it with a set of four diverse newspapers; this analysis estimates no yearly increase (Supplementary Appendix E). H1.2 adequately describes the trajectory in The Times but does not seem to capture a broader pattern.
Three stages of crisis publicity.
Drivers of crisis coverage.
Development of crisis coverage 1785–2020: (A) Share of coverage (%) with crisis labeling (top panel); (B) Share of coverage (%) that belongs to a crisis news wave (center panel); (C) Count of crisis news waves (bottom panel), the years are printed such that larger numbers represent th years with extremely many crisis news waves—this shows the face validity of recognizing years with memorable major societal crises such as COVID-19 (2020), the World Financial Crisis (2008), and the First Oil Crisis (1973).
The log count of CNWs increased over time. The linear model displays a yearly increase of CNW count of 0.004 (0.4 percentage points) (SE = 0.001; t(235) = 4.87; p < .001). The CNW count was 1.95 in 1785; it was 5.47 in 2020 (Figure 3). The analysis replicates when using a set of four diverse newspapers (Supplementary Appendix E). This supports H1.3 and suggests the long-term growth of CNWs per year might be a broader pattern in different contexts.
Additional observations relevant to the analysis of inflation of crisis
Visual inspection of Figure 3 reveals that the past two decades have seen several extreme outliers in terms of massive CL salience, CNW salience, and CNW count; 2008 (global financial crisis) and 2020 (COVID-19) were in the top three years with highest CNW counts. This is part of a change in patterns of CNW count (and to a lesser extent, CNW salience, but not CL salience). Until the 1930s, there were long phases of relatively many CNWs per year (1870–1914) and phases of relatively few CNWs per year (1820–1870, 1915–1930). Since the 1930s, this changed to a low count of CNWs per year in most years, but some exceptional outlier years with extremely many CNWs that coincides with crises that shaped collective memory (1931, 1938, 1973, 2008, 2020; this is why we plot the years as data labels). This may add to the perception that current crises are exceptional: the contrast between crisis years and non-crisis years has become stark in the 20th and 21st centuries.
Drivers of crisis salience (H2)
Structure of government spending and crisis coverage (H2.1)
When adding government spending indicators as predictors, they significantly predict CL salience and logged CNW count, and adding them leads to significant model improvement (M1 → M3: F(2, 231) = 16.558; p < .001; ΔR2 = .042 and M9 → M11: F(2, 231) = 13.11; p < .001; ΔR2 = .098). The structure of government spending does not impact CNW salience (M5 → M7: F(2, 231) = 1.859; p = .158; ΔR2 = .014). H2.1 is supported when it comes to CL salience and logged CNW count while government spending does not seem to affect CNW salience (Table 1). In those cases where government spending structure has predictive power, diversity of government spending has a consistent positive effect whereas the impact of intensity of spending varies: positive for CL salience, and negative for CNW salience. The joint analysis of The Times and The Guardian (only CL salience), further corroborates H2.1 and finds a positive effect of both government spending intensity and diversity (Supplementary Appendix D). Overall, H2.1 is partially supported.
Predicting crisis coverage from pervasiveness of media logic, structure of government spending, and crisis frame sponsor activity
| Models explaining the development of salience of crisis coverage | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CL Salience | CNW Salience | Log CNW Count | ||||||||||
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 | |
| Predictors | B | B | B | B | B | B | B | B | B | B | B | B |
| (Intercept) | 0.260*** | 0.297*** | −0.025 | −0.113* | −0.018 | 0.049 | −0.032 | −0.054 | 0.924*** | 0.976*** | 0.821*** | 0.716*** |
| Pervasiveness of media logic | ||||||||||||
| Media Penetration | 1.203*** | −0.155 | −0.301 | 0.034 | −0.034 | −0.071 | 0.091 | −0.402 | −0.578 | |||
| Media Autonomy | 8.526*** | 8.335*** | 2.345*** | 2.201*** | 2.110*** | 0.581 | 3.013# | 2.404 | 4.784*** | |||
| Structure of public spending | ||||||||||||
| Public Spending Intensity | 1.632*** | 1.216# | 1.344*** | −0.111 | −0.225 | −0.192 | 1.917*** | −1.455 | −1.301 | |||
| Public Spending Diversity | 5.133*** | 4.095*** | 1.563*** | .840*** | 0.567 | −0.080 | 3.516*** | 3.887** | 0.848 | |||
| Crisis frame sponsor activity | 0.311*** | 0.079*** | 0.373*** | |||||||||
| Model performance | ||||||||||||
| R2 | .667 | .591 | .709 | .911 | .096 | .058 | .111 | .202 | .032 | .120 | .130 | .436 |
| ΔR2 (media logic first) | .667*** | .042 *** | .202*** | .096*** | .015 | .091*** | .032* | .098*** | .306*** | |||
| ΔR2 (public spending first) | .591*** | .118 *** | .202*** | .058*** | .053** | .091*** | .120*** | .010 | .306*** | |||
| Models explaining the development of salience of crisis coverage | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CL Salience | CNW Salience | Log CNW Count | ||||||||||
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 | |
| Predictors | B | B | B | B | B | B | B | B | B | B | B | B |
| (Intercept) | 0.260*** | 0.297*** | −0.025 | −0.113* | −0.018 | 0.049 | −0.032 | −0.054 | 0.924*** | 0.976*** | 0.821*** | 0.716*** |
| Pervasiveness of media logic | ||||||||||||
| Media Penetration | 1.203*** | −0.155 | −0.301 | 0.034 | −0.034 | −0.071 | 0.091 | −0.402 | −0.578 | |||
| Media Autonomy | 8.526*** | 8.335*** | 2.345*** | 2.201*** | 2.110*** | 0.581 | 3.013# | 2.404 | 4.784*** | |||
| Structure of public spending | ||||||||||||
| Public Spending Intensity | 1.632*** | 1.216# | 1.344*** | −0.111 | −0.225 | −0.192 | 1.917*** | −1.455 | −1.301 | |||
| Public Spending Diversity | 5.133*** | 4.095*** | 1.563*** | .840*** | 0.567 | −0.080 | 3.516*** | 3.887** | 0.848 | |||
| Crisis frame sponsor activity | 0.311*** | 0.079*** | 0.373*** | |||||||||
| Model performance | ||||||||||||
| R2 | .667 | .591 | .709 | .911 | .096 | .058 | .111 | .202 | .032 | .120 | .130 | .436 |
| ΔR2 (media logic first) | .667*** | .042 *** | .202*** | .096*** | .015 | .091*** | .032* | .098*** | .306*** | |||
| ΔR2 (public spending first) | .591*** | .118 *** | .202*** | .058*** | .053** | .091*** | .120*** | .010 | .306*** | |||
Note. Models 1, 2, 3, 4 are the same as M1.2, M1.3, M1.5, and M1.6 from Supplementary Appendix Q (and the model numbers from Supplementary Appendix Q are used in all appendices). Models 5, 6, 7, 8 are the same as M2.2, M2.3, M2.5, and M2.6 from Supplementary Appendix Q. Models 9, 10, 11, 12 are the same as M3.2, M3.3, M3.5, and M3.6 from Supplementary Appendix Q. The two different entries and tests for R2 change result from a model series where media logic is entered first, then public spending is added, and finally crisis frame sponsor activity is added (“media logic first”). The other series enters public spending first, adds media logic in step two, and finally adds crisis frame sponsor activity.
Predicting crisis coverage from pervasiveness of media logic, structure of government spending, and crisis frame sponsor activity
| Models explaining the development of salience of crisis coverage | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CL Salience | CNW Salience | Log CNW Count | ||||||||||
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 | |
| Predictors | B | B | B | B | B | B | B | B | B | B | B | B |
| (Intercept) | 0.260*** | 0.297*** | −0.025 | −0.113* | −0.018 | 0.049 | −0.032 | −0.054 | 0.924*** | 0.976*** | 0.821*** | 0.716*** |
| Pervasiveness of media logic | ||||||||||||
| Media Penetration | 1.203*** | −0.155 | −0.301 | 0.034 | −0.034 | −0.071 | 0.091 | −0.402 | −0.578 | |||
| Media Autonomy | 8.526*** | 8.335*** | 2.345*** | 2.201*** | 2.110*** | 0.581 | 3.013# | 2.404 | 4.784*** | |||
| Structure of public spending | ||||||||||||
| Public Spending Intensity | 1.632*** | 1.216# | 1.344*** | −0.111 | −0.225 | −0.192 | 1.917*** | −1.455 | −1.301 | |||
| Public Spending Diversity | 5.133*** | 4.095*** | 1.563*** | .840*** | 0.567 | −0.080 | 3.516*** | 3.887** | 0.848 | |||
| Crisis frame sponsor activity | 0.311*** | 0.079*** | 0.373*** | |||||||||
| Model performance | ||||||||||||
| R2 | .667 | .591 | .709 | .911 | .096 | .058 | .111 | .202 | .032 | .120 | .130 | .436 |
| ΔR2 (media logic first) | .667*** | .042 *** | .202*** | .096*** | .015 | .091*** | .032* | .098*** | .306*** | |||
| ΔR2 (public spending first) | .591*** | .118 *** | .202*** | .058*** | .053** | .091*** | .120*** | .010 | .306*** | |||
| Models explaining the development of salience of crisis coverage | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CL Salience | CNW Salience | Log CNW Count | ||||||||||
| M1 | M2 | M3 | M4 | M5 | M6 | M7 | M8 | M9 | M10 | M11 | M12 | |
| Predictors | B | B | B | B | B | B | B | B | B | B | B | B |
| (Intercept) | 0.260*** | 0.297*** | −0.025 | −0.113* | −0.018 | 0.049 | −0.032 | −0.054 | 0.924*** | 0.976*** | 0.821*** | 0.716*** |
| Pervasiveness of media logic | ||||||||||||
| Media Penetration | 1.203*** | −0.155 | −0.301 | 0.034 | −0.034 | −0.071 | 0.091 | −0.402 | −0.578 | |||
| Media Autonomy | 8.526*** | 8.335*** | 2.345*** | 2.201*** | 2.110*** | 0.581 | 3.013# | 2.404 | 4.784*** | |||
| Structure of public spending | ||||||||||||
| Public Spending Intensity | 1.632*** | 1.216# | 1.344*** | −0.111 | −0.225 | −0.192 | 1.917*** | −1.455 | −1.301 | |||
| Public Spending Diversity | 5.133*** | 4.095*** | 1.563*** | .840*** | 0.567 | −0.080 | 3.516*** | 3.887** | 0.848 | |||
| Crisis frame sponsor activity | 0.311*** | 0.079*** | 0.373*** | |||||||||
| Model performance | ||||||||||||
| R2 | .667 | .591 | .709 | .911 | .096 | .058 | .111 | .202 | .032 | .120 | .130 | .436 |
| ΔR2 (media logic first) | .667*** | .042 *** | .202*** | .096*** | .015 | .091*** | .032* | .098*** | .306*** | |||
| ΔR2 (public spending first) | .591*** | .118 *** | .202*** | .058*** | .053** | .091*** | .120*** | .010 | .306*** | |||
Note. Models 1, 2, 3, 4 are the same as M1.2, M1.3, M1.5, and M1.6 from Supplementary Appendix Q (and the model numbers from Supplementary Appendix Q are used in all appendices). Models 5, 6, 7, 8 are the same as M2.2, M2.3, M2.5, and M2.6 from Supplementary Appendix Q. Models 9, 10, 11, 12 are the same as M3.2, M3.3, M3.5, and M3.6 from Supplementary Appendix Q. The two different entries and tests for R2 change result from a model series where media logic is entered first, then public spending is added, and finally crisis frame sponsor activity is added (“media logic first”). The other series enters public spending first, adds media logic in step two, and finally adds crisis frame sponsor activity.
Pervasiveness of media logic and crisis salience (H2.2)
When adding pervasiveness of media logic indicators as predictors, they significantly predict crisis coverage in terms of CL salience (M2 → M3: F(2, 231) = 46.885; p < .001; ΔR2 = .118) and CNW salience (M6 → M7: F(2, 231) = 6.84; p = .001; ΔR2 = .053); however, they do not significantly improve predictions of logged CNW count (M10 → M11: F(2, 231) = 1.40; p = .249; ΔR2 = .010). H2.2 is supported regarding CL salience and CNW salience, but there is no evidence for H2.2 regarding CNW count (Table 1). Media autonomy has a consistent stimulating effect on CL salience and CNW salience across models, while media penetration has no consistent effect across models. The joint analysis of The Times and The Guardian (only CL salience), also corroborates H2.2 but finds a positive effect only for media autonomy, not for media penetration (Supplementary Appendix D). Overall, H2.2 is partially supported.
Crisis frame sponsors and crisis salience (H2.3)
When adding crisis frame sponsor activity to the model—the potential mediator (see H3.1 and H3.2)—model fit improves for CL salience (M3 → M4; F(1, 230) = 521.20; p < .001; ΔR2 = .202), CNW salience (M7 → M8; F(1, 230) = 26.36; p < .001; ΔR2 = .091), and CNW count (M11 → M12; F(1, 230) = 124.61; p < .001; ΔR2 = .306). The density of organizations that participate in crisis discourse in news coverage positively predicts CL salience, CNW salience, and logarithmized CNW count across the board. This supports H2.3. The joint analysis of The Times and The Guardian (only CL salience) corroborates H2.3.
Mediated impact of public spending and mediatization (H3.1, H3.2)
The diversity of public spending has an indirect effect on crisis coverage (as predicted by models M4, M8, M12) via crisis frame sponsor activity (model M4.3 in Supplementary Appendix Q). This is true for CL and CNW salience as well as (log) CNW count. In contrast, spending intensity has no significant indirect effects via crisis frame sponsor activity (Figure 4, left-hand side, two bottom panels, “indirect effects”). H3.1 is supported for spending diversity but not for spending intensity.
Greater media autonomy indirectly stimulates crisis salience via crisis frame sponsor activity. This is true for CL salience, CNW salience and (log) CNW count. In contrast, media penetration exerts no indirect effects. H3.2 is supported for media autonomy but not for media penetration (Figure 4, left-hand side, two top panels, “indirect effects”). The mediation analysis results are similar if CL salience of The Times and The Guardian are analyzed jointly (Supplementary Appendix D). Sensitivity analyses (Supplementary Appendix S) show that the indirect relationships prove robust even if moderately strong unobserved confounding variables were present.
Mediation analysis and effect size comparison. (A) Mediation analysis of media logic’s and public spending’s effects on crisis coverage. (B) Comparing the strength of effects on CL share and CNW share.
Note. Unstandardized coefficients give the percentage point change in CL salience and CNW salience, respectively. Since CNW salience has a much lower share of the total news hole and a lower variability, this possibly exaggerates the effect size on CL relative to CNW. Min-max standardized coefficients consider that the range of CL and CNW salience are different and standardizes their values to the range that empirically occurs 1785–2020. This is an intermediate solution that may give the fairest estimate of which effect size is greater. Finally, z-standardization equalizes the averages of both distributions to 0 and their standard deviation to 1. This may exaggerate the effect size for CNW relative to CL. Reading example (A): For public spending diversity (bottom panel), we observe significant indirect effects (“indirect”, purple circles) mediated by crisis frame sponsor activity on all three dependent variables (left-hand panel: CL salience, center panel: CNW salience, right-hand panel: CNW count). Reading example (B): The coefficients of public spending diversity’s effect (bottom panel) on CL salience (purple circle) are significantly greater than coefficients of its effect on CNW salience (green triangle) across coefficient standardization methods (left panel: unstandardized; center panel: MinMax standardized; right panel: z-standardized).
Can drivers substitute time (RQ1)?
Can our drivers actually substitute time such that we can ignore time in a model altogether, and still arrive at accurate forecasts? To estimate the predictive capacity of models with and without time, we “play dumb,” act as if we had measured only part of our data and want to make predictions for the remaining data (which we pretend to not know). To that end, we divide our period of study into a training period to which the model is fitted, and a heldout period for which we use the now-trained model to make predictions. We compare the predictions with the measured values to check the accuracy of predictions. We vary the training and heldout periods in different scenarios, and vary the model architecture to check how much training is necessary under different conditions. The results are reported as robustness checks in Supplementary Appendix U in greater detail. For RQ1, two take-aways are important:
It is not necessary to use time (and run ARIMA models) to make accurate forecasts. Ignoring time altogether is possible if all driver components are considered and the training period includes the second half of the 20th century. Adding the ARIMA components creates more conservative and stable predictions which smooth out some of the inadequacies of poorly specified or insufficiently data-saturated models, but provides little benefit if models are complete and data-saturated. The answer to RQ1 is: It is possible to substitute time for the theorized set of drivers, if the model is properly specified and sufficiently trained.
Something in the way the drivers affect crisis coverage has changed around 1970 (and this is why including the second half of the 20th century into the training data is so important). This leads to extremely inaccurate predictions of non-ARIMA models whose training period stops in 1975 (when the new mechanism becomes observable, but its boundaries are not yet visible in the data). This coincides with the strongest increase in media autonomy and corresponding mediatization of politics.
Comparing impacts on CLs and CNWs (H4.1, H4.2)
To compare sizes of drivers’ effects on CL salience and CNW salience, we contrast three differently standardized coefficients to be sure to not fall victim to a standardization artefact (Figure 4): (a) Unstandardized coefficients (which might exaggerate the effects on CL salience), (b) z-standardized coefficients (which might exaggerate the effects on CNW salience), and (c) Min-Max-standardized coefficients (which are a middle way). If coefficients for CL salience are greater than those for CNW salience across standardization techniques, it would strongly support the hypotheses (see Supplementary Appendix T for details).
The effect of time on CL salience is significantly greater than on CNW salience (Figure 4, right-hand side, top panel). This supports H4.1. Also for the drivers, coefficients of effects on CL are always greater than effect sizes for CNW (see Figure 4, right-hand side, lower five panels). For media penetration, government spending intensity and diversity, the confidence values also do not overlap at all. For media autonomy and crisis frame sponsor activity, confidence intervals overlap when using z-standardized coefficients. Overall, the stronger stimulating effects of drivers on CL salience than on CNW salience supports H4.2.
Discussion and conclusion
Insights and their implications
The main insights from the present study can be summarized in seven points:
The recent two decades were extraordinary in terms of public crisis coverage and, most likely, crisis perception (side finding). The Coronavirus Pandemic (2020) and the Lehman Brothers Crisis (2008) were in the top 3 years according to all three indicators of crisis salience we used. This justifies recent diagnoses of extraordinary crisis salience in the past years (Beckett, 2019; Saad-Filho, 2021).
There has been a general growth in CL salience, CNW salience (partly), and CNW count. This trend does not seem to be limited to The Times or Great Britain but may be a broader pattern found in several Western countries and across different types of outlets (H1).
The growth of crisis salience is not a simple linear trend. Periods of growth dominate, but there have been intermittent phases of decrease, stagnation, or up and down; time alone does not account for growth (H1). Narratives of incessant growth of crisis coverage (Beckett, 2019) ignore nuances which are informative to understand the phenomenon.
The growth in crisis salience is strongest and most universal for CL salience, which we interpret as a pronounced increase in crisis frame building efforts and crisis rhetoric (Estes, 1983; Holton, 1987; Parks, 2020). In CNW salience and CNW count, which we interpret as full-blown public crisis situations, growth is weaker and more context-dependent. The CNW filter more strictly keeps the gates, filtering out much crisis rhetoric and reflecting the limited carrying capacity of the public sphere (Hilgartner & Bosk, 1988; Neuman, 1990). It is of vital importance to differentiate between CL and CNW when investigating crisis coverage (or to specify which of the two one investigates) (H1, H4). Disentangling the different processes that influence CL and CNW coverage, respectively, facilitates highlighting that there are different kinds of crisis coverage with different implications for society.
A pattern of alternating longer “crisis eras” with many CNWs and “normal eras” was prevalent before the 1930s. After the 1930s, this pattern has been replaced by a more volatile pattern of singular “crisis years” with an enormous number of CNWs, after which normalization occurs abruptly (side finding). This may also lead to the impression that crises have become dominant in public discourse (Saad-Filho, 2021): the pulsating appearance–disappearance pattern makes them stand out and they dominate discourse temporarily.
There is a temporal coincidence between drivers of crisis salience (media autonomy and government spending diversity) on the one hand, and greater CL salience on the other hand (H2). The drivers have stronger and more consistent relationships with CL than with CNW (H4). The data fit a mediating relationship where greater media autonomy (reflecting mediatization of politics; Strömbäck, 2008) and greater public spending diversity (reflecting state expansion; Soroka & Wlezien, 2010) are associated with greater crisis frame sponsor activity that in turn stimulates crisis coverage (H3).
Our model can make predictions of CL salience without the use of time as an explicit variable (Magin & Geiß, 2019). Relying on time alone is clearly the inferior modeling solution compared to relying on the drivers of crisis coverage. Our model captures key components in the long-term change of CL and CNWs (RQ1).
Contributions to scientific and public debates
Besides these clear-cut insights, there are additional important and noteworthy claims that are linked to the findings but require interpretation and invite contention.
Mediatization or state expansion?
Our analyses suggest that mediatization and state expansion jointly stimulate crisis coverage through heightened crisis frame sponsor activity and competition. How exactly can this mechanism be imagined? We reconstruct that the Industrial Revolution set into motion a chain of events that stretched out over many decades: Economic growth, the increased potential for taxation, and the novel kinds of inequality it created triggered the movement towards expansion of voting rights. Government expansion in scale and scope became necessary to address the demands of the newly enfranchised voters; more government spending became possible due to steady economic growth and greater tax revenue potential; the newly enfranchised voters started to care about public affairs and consumed news (monitoring what happened to their tax money), widening the public sphere and its topical spectrum (Estes, 1983; McBeth et al., 2013; Parks, 2020); the circulation of newspapers increased to satisfy the demand. Expansion of wealth and education gradually weakened voter-to-party ties and set news media free from informal party allegiance, pressuring politicians and parties to compete for news attention and public approval. Framing problems in terms of crisis would be one strategy for highlighting newsworthiness of policymaking (Boin, 2009; Olsson and Nord, 2015). This proposed explanation adds facets to the debate on the roots and dynamics of mediatization of politics that only come into view from a extremely long-term perspective.
Future of crisis coverage?
Will the trends that we outline continue or do we see saturation or a trend reversal? We can approach this question by looking at the likely trajectories of the drivers: How will public spending and pervasiveness of media logic develop, and how will crisis frame sponsor activity respond? We currently see a stagnation of some of these indicators. In the U.K., there has been stagnation in the share of government spending as a share of GDP and in spending diversity. Media penetration hit a ceiling in the 1950s (though we have change in the typical quantity and quality of news use). Regarding media autonomy, there are contradictory predictions as to whether party affiliation will increase or not. Partisan polarization on the one hand and political apathy and news avoidance on the other hand are opposing trends whose relative strengths are not yet clear. The flip side of media autonomy from politics was greater dependence on revenue from advertising. Here, digital media environments resulted in intensified competition and this ongoing commercialization will most likely further pose incentives for extensive crisis coverage. The two major spikes in 2008 and 2020 also could indicate continued growth. Overall, some trends hint at continued expansion of crisis coverage while others suggest a stagnation. Time will tell how crisis salience will develop and whether our theory proves robust. Mapping the relationships between these long-term trends and more recent observations of (increasing) populism in media and politics, spread of disinformation and conspiracy beliefs could help us extrapolate future developments of the structure of the public sphere and the prominence of crisis coverage therein.
Inflation or justified growth?
Is the pattern of growth of crisis salience—normatively—to be viewed as an unjustified inflation without real substance? Or is the growth justified by the changed eventscape or the changed (but improved) standards of news selection and crisis labeling? On average, the human condition in the U.K. and around the world has improved in the past 235 years in terms of welfare and security. Therefore, “traditional” disasters might become more seldom (Boin, 2009). However, the greater population density and the higher complexity of society (in terms of interconnectedness and danger of spillover across social domains and across borders) can give rise to more critical situations that warrant public attention, despite greater wealth. The better availability and easier flow of information can contribute to more crises being discovered and covered (Supplementary Appendix A, section 1.5.1). Overall, there is no obvious trend that society would be more crisis-stricken today than in the past.
For now, the main drivers that we can isolate are factors relating to the structure and endogeneous dynamics of the public sphere. The idea of an inflation of crisis coverage is the best description of the empirical patterns. According to our models, they reflect an expanding state that covers and regulates more areas of social life, and a greater pervasiveness of media logic in society. We presume a supply–demand spiral where the greater demand for crisis creates a greater supply of crisis. More diverse interests are represented publicly, leading to a more diverse set of crises. The story of crisis salience is for now not so much a tale of the problems societies face, but of a society’s identity, priorities, and outlook on the world.
Limitations and outlook
Measuring crisis frame sponsor activity
There is some co-dependency concerning CL salience and CNW salience on the one hand and crisis frame sponsor activity on the other hand. They all reflect the increasing number of articles with CL over time. To reduce the risk of spurious relationships, we validated against small-scale manual coding, which suggests that the thematic diversity of organizations mentioned actually increased, which is also consistent with the historical analysis of the development of PR (Miller & Dinan, 2008) and lobbying (McGrath, 2018) in the U.K. (Supplementary Appendix V).
Detecting CNWs
When it comes to detecting CNWs, the additional criterion of intermedia consensus—which tends to be very high in crisis coverage (Hase & Engelke, 2022; Geiß, 2013)—could not be considered as we only analyzed one newspaper. However, the same logic still applies: A newspaper would (usually) not present an issue as a “crisis” and follow it intensively unless it resonates with their audience. This may occur in single cases but will not be a systematic pattern.
Causality
Our study mostly presumes a causal direction from drivers to crisis coverage based on a plausible theoretical argument. By probing time-lagged relationships and Granger causality (Supplementary Appendix R), alternative explanations (Supplementary Appendices A [sections 1.4–1.6] and C) and by pointing at the intricate relationship of social change at various levels that find expression in the drivers of crisis coverage (paragraph on “Mediatization or State Expansion?”), we illustrate the plausibility, nuance, and robustness of our explanatory framework.
Generalizability
Obviously, The Times is a special case, but it is also an especially interesting and influential case, given its traditional closeness to the centers of power. Analyzing other newspapers may reveal very interesting nuances in which kinds of crisis salience. Still, we argue that the major structural developments that have shaped the salience of crisis coverage in The Times have affected other outlets in similar ways; this is corroborated by additional data; our analyses suggest that some of these trends are even transnational (Supplementary Appendices D and E). The media we have analyzed (The Times, The Guardian, The Economist) are influential among the population, the British media, and British politicians; such that systematic changes in their crisis coverage have significant implications, and per se indicate structural change. Even beyond the U.K., it appears to be a plausible working hypothesis that other Western countries may have experienced similar mechanisms that drive crisis coverage that, however, lead to divergent temporal patterns (as public spending and pervasiveness of media logic change at different rates and at different times). Exploring such a time-lagged crisis inflation processes in different countries would be an interesting next step to explore the generalizability of our results. Crisis coverage has many additional facets beyond salience to explore, such as crisis framing (including debates on responsibility, guilt, and blame), the geography of crisis news and other news in comparison, the difference between crisis and other social problems.
Historical newspaper analysis
Over 230 years, we have to reckon with substantial language change, social change, political change and media change. During the period of study, the U.K. has been through the Industrial Revolution, the Digital Revolution, the Silent Revolution, the Napoleonic Wars, and both World Wars. This raises the question to whether CL and CNWs have the same meaning or significance today as they used to. Linguistically, we checked and can confirm that crisis keywords we used then and now describe the same basic idea of an existential and non-normal threat (Supplementary Appendices F and G). At the same time, we view the evolution of the public sphere and the evolution of public crisis as closely interrelated and use it as a window to study social, political, and media change. We introduced a methodology for studying crisis coverage in the long run that can be applied to other news outlets or even to other research areas than crisis coverage (e.g., the study of scandalization). In an even broader view, it represents an approach for analysis of historical newspapers using fully automated content analysis methods.
Supplementary material
Supplementary material is available online at Journal of Communication online.
Data availability
The higher-order data we use in the analyses and plots are publicly available on a GitHub repository (https://github.com/stefangeiss/inflation) and the OSF repository for the study (https://osf.io/kuhzx/). We are not legally permitted to publish the raw text data.
Conflicts of interest: None declared.
