Selective Control: The Political Economy of Censorship

ABSTRACT

Alongside democratic backsliding and security threats, censorship is increasingly used by governments and other societal actors to control the media. Who is likely to be affected by it and why? We argue that censors are more likely to target outlets and journalists that provide information to politically consequential audiences, while allowing media that caters to non-pivotal audiences to report more freely. In order to test our hypotheses, we built a new dataset of around 9,000 salient censorship events and their characteristics across 196 countries between 2001 and 2015. We find strong empirical support for media market segmentation. Outlets and journalists with wide audiences, collective action coverage and domestic ownership are significantly more at risk of severe censorship actions. We also find that audience pivotality matters more than the number and diversity of outlets for censors’ strategic calculus. Our results hold across democracies and non-democracies, for government and third-party censors alike.

KEYWORDS:

• Censorship
• media capture
• political economy
• informational regimes
• transparency
• information bubbles

Introduction

Following the 1990s wave of democratization, conventional wisdom suggested that autocratic regimes censored all media indiscriminately, while democracies rarely engaged in censorship. Moreover, that blatant coercion is an obsolete strategy of information control to be replaced by media capture via inducements, such as ownership and bribery. Twenty-five years later, both assumptions have been empirically challenged. Many autocrats allowed dual information control systems to co-exist. Media catering to narrow audiences could publish or broadcast politically sensitive content, whereas outlets targeting popular audiences were systematically censored. Over the last two decades, following democratic backsliding and security challenges, old-time censorship has increased around the world, affecting both democracies and autocracies.

Who are the censors? Who is likely to be most affected by censorship and why? Are all media equally exposed to it? Building on prominent theories of media control in democracies and autocracies (Besley & Prat, 2006; Guriev & Treisman, 2020), we argue that exposure to punitive measures depends on the political threat posed by a media outlet’s audience and on the transaction costs of censorship. Irrespective of political regime type, censors are likely to allow free information in narrow, elite segments of the media market while suppressing it for larger audiences that include a critical number of voters (or selectors) to potentially act collectively upon uncensored information.¹ The paper proposes that censors economize coercion by targeting only media market segments that cater to a broad enough audience and facilitate collective action. Since currently there is no individual-level data that allow us to test empirically whether the segmentation of news consumption can lead to heterogeneous levels of censorship within a country, this paper presents a first attempt to understand selective censorship using global event data. We generated a new cross-national dataset containing around 9,000 salient censorship events and their characteristics for 196 countries, between 2001 and 2015. Our theoretical focus is on explicit coercion in anticipation of or in response to reporting, as we distinguish such strategies from inducement-based forms of media capture (bribery or collusion) for inferential precision.

We show that coercive actions against the media occur in all regimes, being committed by both governments and non-governmental actors. Our findings suggest that coercion is likely to disproportionately affect journalists and outlets that reach the broadest politically consequential audiences, confirming recent insights (Guriev & Treisman, 2020). First, we review the literature. Second, we propose a general theory. Third, after introducing our original dataset, we verify whether the hypotheses derived from our model hold empirically, followed by a conclusion.

Literature

Censorship of the media refers to formal or informal interference with the freedom of outlets and media professionals to collect information, exchange ideas and report to the public “through any media and regardless of frontiers,” as per Article 19 of the Universal Declaration of Human Rights. Interference varies widely from a lack of access to government information to the killing or disappearance of journalists. Despite the instinct to associate censorship with governments, the censors are often non-state actors, such as organized crime networks, radical religious groups, and others. What renders censorship difficult to study comparatively is its highly complex ecosystem where formal restrictions, informal pressure, and inducements mix.

The benefits of controlling information for politicians and bureaucrats revolve around being shielded from scrutiny, as non-transparency facilitates the use of public office for private gain (Adsera et al., 2003; Ferraz & Finan, 2011). For autocratic leaders, censoring the media is an essential strategy of rule that prevents contestation, the revelation of human rights infringements or corruption (Van Belle, 2000). Nevertheless, censorship is not without costs for censors. Even repressive autocrats tolerate oases of media freedom because of legitimacy concerns (Whitten-Woodring, 2009). Within autocracies and democracies alike, political competition increases the costs of censorship as media grants visibility for politicians, allowing them to pander to their constituencies (Malesky et al., 2012). The benefits of (some) media freedom are non-trivial for leaders because it also allows political principals to monitor the performance of lower tier governments, thus minimizing moral hazard problems. Investigative journalism in autocracies helps central leadership collect credible information about other officials, thus contributing to bureaucratic oversight (Egorov et al., 2009; Lorentzen, 2014). Credibility concerns and citizen perceptions are also crucial for information-manipulating autocrats and democrats alike (Shadmehr & Bernhardt, 2015; Stockmann & Gallagher, 2011). In the words of one Chinese journalist who protested a censor’s interference, “if media lose all credibility and influence, then we ask, how is the ruling Party to speak?” (Richburg, 2013). In democracies, refraining from censorship also boosts the credibility of pro-government narratives. Despite occurring daily in the White House, only a few leaks to the press are ever prosecuted. As incumbents occasionally “plant leaks” to the media, general under-enforcement allows them to appear truthful since otherwise their conspicuous exemption from punishment would lower credibility (Pozen, 2013).

Balancing costs and benefits, a selective (or segmented), as opposed to an all-or-nothing, enforcement of censorship pays off for censors. A corpus of formal literature that began with Besley and Prat (2006) seeks to model democratic decision-makers’ incentives to manipulate the media cost-effectively. Recently, Guriev and Treisman (2020) explicitly model regimes as a function of the information elite size, and of the costs of manipulation. Theoretically, we build on key insights from these two important studies to argue for a “segmentation” framework that transcends political regimes and censor types, tested with new data. Empirically, our work could be considered a first global event-level test of selective censorship occurring in modern “informational” autocracies that rely mainly on the manipulation of the public rather than on overt violence, a key feature in Guriev and Treisman (2020). We also extend the applicability of their model by showing that non-state censors respond to similar incentives, and by testing collective action threats across regimes.

Theoretical Framework and Hypotheses²

Our theory assumes that there are two distinct segments of any media market, one catering to mass audiences, and another to elite or niche audiences. The mass audience is larger than the elite audience and includes the median voter/selector. We refer to the two stylized segments as elites and masses, but these two groups can accommodate many context-specific configurations as long as they maintain their importance differential: urban versus rural news consumers, different linguistic segments of the media market, different levels of education and civic engagement, or simply two social groups of different political pivotality (i.e. importance/threat for the incumbent’s political prospects), even if the term elites is a misnomer in some cases. It is worth emphasizing that the concept of pivotality does not always equate the size of voters/news consumers, and is context and institution-specific. In some electoral systems, the number of pivotal voters is small (Mulligan & Hunter, 2003). The theoretical usefulness and flexibility of pivotality with respect to media control stems from its ability to identify “the audience that matters” for the censor with precision. For formal modeling and empirical convenience in a cross-national setup, we assume that a larger group (the mass public) is more likely to include the pivotal news consumer (median voter/mobilizer), but predict that censorship would follow a similar theoretical logic even in the case of smaller audiences deemed consequential by political incumbents or third party censors.

While elite and non-elite media markets work in similar ways, there is one fundamental difference: their importance for the incumbent’s ability to remain in office. Information control for non-elite audiences is essential for political survival, but paradoxically this is not the case with niche elite audiences. We are far from arguing that elites are inconsequential for political survival. In fact, in autocratic contexts, they have a greater capacity to directly remove the incumbent vis-à-vis the masses. Nevertheless, elites are less consequential when it comes to the incumbent’s considerations of media control. There are several reasons for this. First, the median voter in democracies (or selector in autocracies) is a non-elite member. Second, elites oftentimes have superior private information about the incumbent via formal or informal networks, and their own collective action capacity does not rely on news alone. Therefore, exposure to information via the media, as opposed to more personalized channels, is less likely to change their behavior. In autocracies, incumbents may still need to placate elites with disproportionate economic rents and/or repression, while knowing that the information sent via the media is not going to significantly augment their preexisting information and collective action potential. Conversely, mass audiences do not have private information, and, because of more significant free-rider problems owing to their larger size, have limited capacity to mobilize in the absence of the information they receive from the media. In China, for instance, powerful Communist party elites have traditionally had access to objective private and internally distributed information, whereas the stream of news available to the general public has been systematically curated and censored (Dimitrov, 2017). In democracies, policy-making and news-making are intertwined, and media indeed serves as an information provider and a crucial channel of intra-elite communication that may boost their collective action capacity (Cook, 1998; Van Aelst & Walgrave, 2016). However, the impact of the elites on the incumbent’s electoral outcome is ultimately mediated by the perceptions of the median voter located in the mass public. Earlier experimental studies also found that whereas informational exposure via the media altered significantly the views of the general public on policy priorities, its impact varied across policy elites and did not alter their rankings of issue priorities, demonstrating the resilience of private information in this group (Cook et al., 1983). Third, some media freedom in elite segments yields coalition-building opportunities in democracies, as well as legitimacy, performance signaling and elite monitoring via investigative journalism in autocracies (for instance, by pursuing divide-and-rule strategies through scandals and other compromising press revelations).

A simplifying core assumption we make is that there is no spill-over between the two segments of the media market. This is easily defensible in cases of internal media, linguistic segmentation or international news outlets, where the majority of the population would have difficulty of access (Dimitrov, 2017). There is abundant evidence that censors operate with such an assumption. When the Chinese Constitution has been modified to grant Xi Jinping extended tenure, the English-language media (CGTN) reported the event straightforwardly, while its Chinese-language equivalent avoided its direct mention. In general, news consumers evaluate narratives based on preexisting frames with the effect of limiting spillover (Iyengar & Hahn, 2009).

Based on these two assumptions, we hypothesize that censorship (when occurring) will target primarily media that reach larger audiences while bypassing niche or elite markets. Therefore, our key parameter is the pivotality of the media segment, $\hspace{0.17em}\left(\omega \right),\hspace{0.17em}$ operationalized as the likelihood that it contains the median voter in democracies or median selector/mobilizer in autocracies where formal voting or institutional checks and balances are presumably less effective for political accountability than social mobilization via extra-institutional means (i.e. protests, strikes, demonstrations). The wider the reach and popularity of an outlet, the higher the likelihood is that it will provide information to a pivotal voter in democracies or to a pivotal selector in autocracies. By arguing that the size of the audience matters because of pivotality, we implicitly theorize government censorship. Nevertheless, we emphasize that non-state censors also engage in a similar “segmentation” calculus, as they pursue goals that are threatened by broad information disclosure. Drug cartels and other extra-legal organizations, for instance, obstruct information that reveals their collusion with the state. Second, many third-party censors are directly connected to governments and excluding them would bias our understanding of principal-agent relations involved in the act of censoring. State censorship in China, for example, is largely and officially sub-contracted to private entities (King et al., 2013). Similarly, in Russia, many private media companies exercise content control by excluding critical channels from their programming packages, in close alignment with the political incumbent (Beazer et al., 2021). In all regimes, including advanced industrial democracies, government-sponsored threats against journalists are often delegated to criminal organizations. For instance, in many democratic countries, journalists covering corruption at the local level have often become victims of state-induced violence whose private perpetrators (gangs, cartels) act on behalf of local politicians (Carey & Gohdes, 2021).

Hypotheses

Our core theoretical predictions regarding censorship “segmentation” apply to individual media outlets and journalists as units of analysis. Identifying the pivotal audience segment is certainly context specific. For instance, on a micro-level, the number of pivotal counties that are politically crucial in the United States electoral races is low. Unfortunately, because of our cross-national research design, we cannot test granular micro-electoral dynamics. Instead, we assume that the median voter (in democracies) and the median selector/mobilizer (in autocracies) embodies the notion of pivotality. Among many potential dimensions, we select four characteristics that place media outlets in different categories of political threat based on the likelihood that they cater to the median voter/selector: foreign versus domestic media ownership; the size of the audience; the physical location of the outlet (in-country or diaspora); and its coverage of collective action.

We hypothesize that domestic journalism is significantly more at risk than foreign outlets as the main news supplier for a large majority of national information consumers (Shirky, 2011). Media that reaches a critical mass of the population should, according to our logic, be more prone to censorship than niche audience segments. Evidence robustly suggests that broad circulation newspapers, popular TV programs, and tabloids have been significantly more targeted by censors than outlets with smaller audiences even in repressive autocracies (Sükösd, 2000). The same logic applies to inducement-based media capture (McMillan & Zoido, 2004). Diaspora media also reaches, on average, smaller news consumer segments than in-country outlets, although in some contexts, it has been counterintuitively found to increase support for autocratic regimes (Kern & Hainmueller, 2009). Finally, our theory emphasizes the political threat factor induced by the exposure of the pivotal voter/mobilizer to accurate reporting. Whereas audience reach captures the location of the pivotal mobilizer, its static nature does not allow us to test the actual political threat. For this purpose, we additionally propose that news covering protests, demonstrations, strikes or riots, and facilitate information coordination, increase the likelihood of collective action for the pivotal mobilizer, and are likely to be subject to more intense scrutiny from censors (King et al., 2013; Roberts, 2018).

H1a: Domestic media is more likely to be targeted by censors than foreign media because it reaches a wider segment of the population.

H1b: Media with larger audience reach is more likely to be targeted by censorship than media that reaches smaller audience segments.

H1c: Diaspora media is overall less likely to be targeted by censorship than in-country media.

H1d: Outlets and journalists covering collective action news are significantly more likely to be exposed to censorship.

Alternative Mechanisms

Our theory highlights a strategic calculus of information control. Nevertheless, it might be the case that other factors are more important for explaining censors’ decisions: the difficulty versus ease of censoring (τ), or the very need to control the media given available rents (r). We address each.

Transaction Costs of Censorship (τ)

Irrespective of their willingness to employ segmented strategies of curating information for different political audiences, censors often lack the capacity to do so. High transaction costs of content control are important factors shaping the occurrence and intensity of media capture (Besley & Prat, 2006). Transaction costs (τ) refer to the costs of controlling the media induced by technology, as well as by social, legal, and political institutions that circumscribe the use of censorship, and manifest at both country and media outlet-level. Examples of transaction costs at the country level include: (a) political costs: censors facing vocal domestic constituencies or aspiring to international legitimacy find it harder to control information without reputational consequences; (b) legal costs: in countries with a functioning judiciary the cost of censorship can become prohibitive; (c) trust-induced transaction costs: even in highly competitive media markets, entire population segments may discount “truthful” narratives from certain outlets because of partisan biases. Censorship is less likely to occur in contexts characterized by higher transaction costs. Transaction costs are notoriously difficult to operationalize empirically and can only be tested by proxy (Besley & Prat, 2006; David & Han, 2004). On a macro-level, because censors face different institutional constraints across political regime types, we can plausibly argue that the logistical, institutional and reputational costs of censorship are likely to increase in democratic settings and decrease in non-democratic contexts.

H2a: Democracies are expected to censor less than autocracies.

Transaction costs also vary significantly by types of media outlets on a micro-level, circumscribing the feasibility of segmented censorship strategies. The geographical location of the outlet imposes logistical costs to censors. Diaspora outlets and international platforms are harder to control than media located in a territory that is physically or virtually in the jurisdiction of the censor (Pan, 2017). Blatant censorship of foreign reporters or of outlets located in a different country entails higher global legitimacy costs.³ Similarly, regional and local censorship is potentially less costly for censors than more “visible” or observable events taking place in the capital city. Economically, government-owned (or influenced) media is less costly to censor and more prone to capture than independent outlets given their preexisting financial and operational reliance on the state. It might, nevertheless, be the case that compliance is secured by censors without any resort to intimidation. While their susceptibility to coercion (as opposed to economic and operational dependence) is not theoretically clear, we investigate this argument with respect to censorship as well, in order to test whether coercion and economic capture are complementary or substitutable.

H2b: Domestic media entails lower reputational costs of censorship than foreign media, and therefore is more likely to be targeted by censors.

H2c: Diaspora media is more difficult to censor because of higher technological and logistical costs of information control.

H2d: State owned or influenced outlets are more exposed to censorship.

We also take into account two context-dependent parameters that have been identified as crucial in previous models: the incumbent rents (r) and the number and diversity of media outlets (N).

Rents (r)

Previous work on censorship emphasized the role of rents, conceptualized as the proportion of public resources which can be diverted by a political incumbent to her private agenda (Besley & Prat, 2006). Governments with large and unaccountable windfall revenues from natural resources, for instance, might not need any free media segments to monitor bureaucrats or seek legitimacy, and can afford to suppress indiscriminately (Egorov et al., 2009).

H3: High rents (r) are likely to induce more intense censorship.

Media Market Diversity (N)

One of the important implications of classic theories of media capture is related to the number of outlets active in a media market. The intuition is that media plurality makes it difficult and expensive for politicians to bribe many, as opposed to few news sources, in order to control information flows (Besley & Prat, 2006). We formulate a similar alternative hypothesis.

H4: A higher number of media outlets (N) as a proxy for media diversity reduces the occurrence and intensity of censorship.

Data

To test the implications of our model, since existing global media freedom indicators are measured at country level and do not capture various media segments within countries, we generated the Global Event Dataset of Media Censorship (GED-MC), the first original repository of individual salient censorship events between 2001 and 2015, covering 196 countries (Corduneanu-Huci et al., 2020). To do so, we hand-coded factual information for the universe of censorship events identified by two main sources: the country-year reports of Freedom of the Press and global data collected by Beacon for Freedom of Expression, a project of the National Library of Norway (Sections A.2 and A.3, Online Appendix and Codebook).

The dataset contains 8,979 observations and includes a series of variables on both targets and censors, as well as many event-specific characteristics (Online Codebook). By developing our own data for individual events, we intended to work with objective measures of censorship, consistently across countries and years. Second, the data goes beyond government-led punitive actions to record censorship committed by non-government entities (Figure B11, Online Codebook). Third, unlike other datasets, our definition of censorship is not limited to extreme human rights abuses such as killing and jail. We code an expanded range of acts that count as censorship, from legal suits, fines or accreditation withdrawal, to more severe forms of punishment such as intimidation, assault, death or disappearance. Fourth, the data on censorship acts goes beyond a simple event count and records all the relevant characteristics of the media outlets and individual journalists, in order to test our main hypotheses. Fifth, our data and tests are primarily circumscribed to traditional media (television, radio, print and online news outlets). We leave aside social and digital media events or routinely occurring content review in institutionalized censorship systems because of feasibility constraints since it would be impossible to collect granular and systematic data on micro-censorship worldwide. Additionally, the very nature of digital media control often involves “porous” censorship strategies (i.e. friction that makes truthful information more difficult to obtain without completely blocking access to it, or flooding that deliberately relies on an information overload coupled with an oversupply of slogans whose de facto result is reducing the critical engagement with content; Roberts, 2018). Such strategies do not qualify as either traditional coercion or inducements, and fall beyond our theoretical scope.

Several caveats are in order regarding the limitations of our dataset. We are aware that the universe of censorship events worldwide is much larger than the “tip of the iceberg” we capture. Global rapporteurs are more likely to record salient or “highly visible” episodes of censorship rather than small friction or less observable, yet routinely occurring, events. Actions affecting journalists and outlets located in the capital city, for instance, are more reported upon than those taking place in rural or remote areas. So are terrorist attacks compared to other triggering events, since they have been prioritized by international observers. This data-generating process by which some event types are more likely to be recorded than others could be related to underlying factors (i.e. platform-specific, temporal or geographical), and might raise concerns related to the statistical representativeness of our sample with respect to the entire population of actually occurring censorship episodes. Despite the fact that saliency has intrinsic value, this potential limitation could in principle affect the generalizability of findings. Therefore, we assessed such potential “observability” biases in two contexts for which detailed crowd-sourced micro-data allowed us to do so, and examined how the dataset fares in comparison to established cross-national measures of media freedom. Figures A1 and A2 in the Online Appendix illustrate the parameters of generalizability. Even more challenging, institutionalized censorship systems that rely on routine automated and/or bureaucratic review and engage in systematic micro-censorship affecting millions of content entries are impossible to observationally unpack at event level in a cross-national set-up, and extant indicators are available only as country-year variables.⁴ Because the main purpose of this paper is to understand the variation in censorship strategies across types of targets within countries, we can only adopt a second-best approach by controlling with country-level variables capturing institutionalization of censorship, and focus instead on individual salient coercive events that reactively or pro-actively target the media. Moreover, previous research on institutionalized censorship rests on hypotheses and findings that share similarities with our own theory (Guriev & Treisman, 2019; King et al., 2013).

We also confine our conceptualization (and operationalization) of salient censorship events to cases of individuals and outlets that meet two criteria of inclusion: 1) they were directly and 2) punitively or coercively targeted by censors because of their media work. Other scenarios not meeting these two conditions, such as indirect actions (deaths caused by crossfire, fatal incidents while being on assignment, etc.), or information control by positive inducement actions (i.e., bribery and collusion) are excluded. The original dataset that we have designed according to these rules is, to the best of our knowledge, the only source that clearly distinguishes coercion from inducement-based media capture at the micro-level. We defend this choice on two grounds. Empirically, salient censorship events are ceteris paribus the most visible (or “observable”) actions that shed light on when censors react and trace the red lines of reporting in the future. Unfortunately, inducement-based media capture actions (i.e., individual bribes or collusion) are not systematically observable at the micro-level, and therefore less consistently captured by global media monitors. For the limited individual-level inducement data we coded (on outlet ownership), the strict analytical separation of coercion from inducements allows us to test the complementarity or substitutability of control strategies.

Empirical Investigation

Studying censorship is methodologically challenging for several reasons. In the case of observational data, the underlying binomial distribution is not observable. One can record systematic data on what has already been censored, but cannot observe all the news items that could have been censored but were not. This problem is particularly insurmountable in cross-national work since even basic information on numbers of media outlets is rarely available. Experimental methods are far better suited for causal inference, but rarely used in the study of censorship because of political sensitivity and difficulty of access to censors’ internal decision-making process (King et al., 2013).

Within these constraints, we opted for a pragmatic operationalization. In addition to estimating a country-level probability of censorship occurring or not (Occurrence of censorship), by collecting all the salient censorship events that have been recorded by media freedom reports, we chose to also construct a dependent variable that records the Severity of censorship based on a theoretically-informed coding scheme that assigns ordinal values according to the magnitude of consequences intended by censors and signaled to targets: legal and financial consequences [lawsuits (1), fines (2), eviction (3)]; short-term obstacles to the production and distribution of information [such as temporarily blocking websites, broadcasts or print] (4) and denying accreditation (5)]; long or medium term obstacles to the production and distribution of information [license removal (6), bans (7), intimidation (8)]; threats of physical violence (9); assault (10); incarceration (11), death (12) and other measures including disappearance (13)].⁵ Two aggregate sub-categories of this variable also separate moderately punitive measures (lawsuits, fines, evictions, blocks, accreditation denials, license removals, bans, intimidation, threats of violence) from severe censorship events that entail assault, incarceration and death.⁶ We argue that censorship severity is conceptually more informative and operationally more discerning than the mere occurrence of a censorship act because of its simultaneous roles of retrospective punishment and prospective signaling. A caveat: it is possible that events that we assign a lower severity score to could be consequential for a media ecosystem. For instance, the cancelation of a broadcast license might constrain news production and distribution to a larger extent than a physical attack. Theoretically, however, we focus on the signals that censors intend to send rather than on overall consequences for media systems that are harder to predict (i.e. the physical intimidation of a journalist intends to deter, but might not lead to actual deterrence). From this “signaling” perspective, a scale that scores violence as more severe than a legal ban or a fine is justifiable.

Accordingly, our dependent variables of interest are: (1) a dichotomous variable of censorship occurrence; (2) a theoretically informed ordinal measure assessing the severity of censorship incidents with thirteen categories of actions, grouped by consequences for targets, and ranging from lawsuits and fines to jailing, death and disappearance; and (3) counts of country-year censorship events, by media platform type.

At country-level, we test four main independent variables hypothesized to shape censorship incentives according to some of the theoretical mechanisms we have discussed: political regime type, the magnitude of rents, legal and economic media market constraints, and level of development and population. For identifying political regime types, we use the Hadenius-Teorell indicator, computed as the mean of the Freedom House (FH) and Polity IV scales converted to a 0–10 range. Following a conventional cutoff point, democracies score values higher, and autocracies lower, than 7.5 (Hadenius & Teorell, 2007).

We also conduct robustness checks with several other proxies for political institutions, as well as with two specific variables capturing explicit interference with media work, Political and Legal constraints (Freedom of the Press), to control for institutionalized censorship systems while being cautious of potential tautology triggered by their partial overlap with our dependent variable (Online Appendix, sections A.4 and A.5). The analysis operationalizes rents as Oil production. As a macro-control of market concentration and financial independence, we use the variable Economic constraints (Freedom of the Press) whose higher values indicate oligopolistic or monopolistic information markets, as well as a V-Dem measure of Media corruption (Table A1, Online Appendix). To test the specific hypotheses of censors’ selective control of different media market segments based on the audience pivotality and transaction costs at the micro-level, the independent variables of interest aim at operationalizing several dimensions: (1) whether or not the targeted media outlet is foreign or domestic, with these categories going beyond ownership per se and aiming to primarily capture the locus of control over news content, dichotomous (Foreign ownership)⁷; (2) whether the media outlet is located in the country or abroad, dichotomous (Media located abroad); (3) whether the media outlet distributes information to a narrow or large audience, ordinal (1–4, low to high values) (Media reach); (4) whether the media or journalist covered collective action events when censored (Collective action reporting); and (5) whether government owned or influenced media was the target of censorship, dichotomous (State owned or influenced media).⁸ All the event-level variable have been constructed and cross-validated by human coders following a strict coding protocol (Table A.1 in Online Appendix and Codebook).

For our two key proxies of pivotality and political threat, Media reach and Collective action reporting, we generated two variables. To assign meaningful values on an ordinal scale for the size of the audience, we used available Pew Global Attitudes and Trends surveys (2001–2009) asking respondents what type of media they obtain their national and international news primarily from.⁹ Between 2001 and 2015, as captured by Table A20 and Figure A9 (Online Appendix), on average, television led worldwide as the primary source of news, followed by radio, print, and remotely, by online sources. This empirical distribution confirms previous findings on platform-specific news consumption. Accordingly, based on this aggregate data and theoretical insights, we created an ordinal variable, Media reach, ranging from 1 to 4, by coding all the individual censorship events that targeted each specific type of media. Lower values indicate a smaller audience reachable by the platform [Online (1); print (2); radio (3); and TV (4)]. We acknowledge that from a current perspective, the data we use is outdated since in many contexts online media has equaled or bypassed traditional media as the primary news supplier over the last decade. Further research is needed to test how well recent trends respond to our theoretical predictions. However, for the period selected for this dataset (2001–2015), our media reach scale is supported by extant polls on news consumption.

For Collective action reporting, we exploited a categorical variable in our dataset (Event triggering censorship) that assigns values (1–7) to the following categories of events if sources explicitly mention coercion as a direct response to the coverage of a particular episode: International sport competitions (1); Elections (2); Ethnic/racial tensions (3); Protests, demonstrations, rallies, strikes (4); Terror attack (5); Conflict (6); and Other events (7) (Online Codebook, Figure B9). For capturing censorship in response to (or in anticipation of) collective action, we constructed a dummy variable taking a value of 1 for protests, demonstrations, rallies, strikes, and 0 for all other types (Table A1 and Table A20 for summary statistics). A typical value of 1 corresponds to the following type of content and its many variations: “Reporter (…) was severely beaten by district security forces while covering an (opposition) rally in (…).”

Econometric Specification

The dependent variables are thus the occurrence of censorship (as event probability and number of events), as well as its severity. For Hypotheses 1–4, the baseline models [Logit (1), Ordered probit (2) and Poisson (3)] have the following specifications:

(1) $p\left(Even{t}_j=1\right)=X{\beta }_j^{\prime }+W{\gamma }_{ijt}^{\prime }+{\theta }_t+{\delta }_i+{\epsilon }_j$(1)

(2) $Severit{y}_j=X{\beta }_j^{\prime }+W{\gamma }_{ijt}^{\prime }+{\theta }_t+{\delta }_i+{\epsilon }_j$(2)

(3) $Event{s}_{it}=\alpha {\vartheta }_{it}^{\prime }+W{\gamma }_{it}^{\prime }+{\theta }_t+{\delta }_i+{\epsilon }_{it}$(3)

where X is a vector of hypothesized event-specific characteristics – transaction costs of individual censorship (state owned media, foreign media and media located abroad) and political pivotality (media reach and collective action); $W$ is a vector of exogenous country-year variables serving as proxies for macro-level transaction costs and rents (economic and legal constraints, democracy, GDP pc, population, oil rents). ${\theta }_t$ is a time fixed effect, ${\delta }_i$ is a country fixed effect, and ${\epsilon }_j$ is the error term. $\alpha$ is the platform-specific audience size. Given the obvious selection effects involved in censorship (i.e., censorship severity is observed only when censorship occurs), we also employ a Heckman Selection Model (HSM) with the following specification:

(4) $Y_j=Y_j^{*},Zijt*0Y_j,Z_{ijt}^{*}=0$(4)

(5) $Z_{ijt}^{\ast }=W{\gamma }_{ijt}^{\prime }+{\theta }_t+{\delta }_i+v_{ijt}$(5)

(6) $Y_j^{\ast }=X{\beta }_j^{\prime }+{\epsilon }_j$(6)

where $Z_{ijt}^{\ast }$ is the incidence or occurrence of a censorship event j being observed in country i and year t. $Y_j$ captures censorship severity of individual censorship events conditional upon a censorship event being observed. ${\epsilon }_j$ and $v_{ijt}$ are error terms. The conditional expectation then is:

(7) $E{Y}_j\text{|}Zijt*0=X\beta j+\rho {\sigma }_{{\epsilon }_j}\lambda W\gamma ijt$(7)

where $\lambda$ is the inverse Mills ratio; ${\sigma }_{{\epsilon }_j}$ is the standard deviation of ${\epsilon }_j$ and $\rho =cor\left(\nu ,\epsilon \right)$.

Identification Strategy and Placebo Tests

In the case of observational data, the risk of endogeneity and omitted variable bias is high without proper identification. In the case of censorship occurrence, the audience size might be endogenous to information control, as sophisticated consumers discount government propaganda and censorship, and are more likely to shift to other sources, thus reducing the platform-specific audience size (Knight & Tribin, 2019). Similarly, the independent segments of the market are also a byproduct of media capture strategies employed by governments (Gehlbach & Sonin, 2014). We opted for two additional identification strategies in addition to the Heckman selection model to mitigate such concerns. First, our analysis uses severity as one of the main dependent variables. Irrespective of the dynamics and size fluctuations of media segments and audience sizes, severity captures intentionality and arguably transcends media platforms and their level of political independence. Simply put, as occasional acts of professional journalism happen across all outlets independently of their level of capture, the reaction of the censor in terms of severity of punishment should still be able to proxy strategic behavior and intentionality, and is thus less directly affected by the dynamic equilibrium in the media market. Nevertheless, high severity censorship against certain outlets as a red line marker could still decrease the probability that they will get censored again in the future (for instance, violent acts in year t against a TV station might manage to intimidate the target who will refrain from broadcasting “truthful” information in year t + 1, therefore reducing the overall need for censorship). Acknowledging this problem, we also use an instrumental variable (Audience size) that measures country-year shares of main news sources, based on Pew Research Center’s Global Attitudes and Trends surveys (2001–2009). For each country and year, our instrument captures the percentage of survey respondents who indicated the platform that they primarily get their everyday news from (TV, Radio, Print and Online). For instance, a 0.8 value on our TV specific Audience size variable, compared to 0.2 for Radio indicates that in year t, 80% of country i’s population relies primarily on television news whereas only 20% on radio outlets. Figure A9 (Online Appendix) plots this time series instrument by panels, showing that there is significant variation by both country and year.¹⁰ The intuition for choosing this instrument to predict censorship occurrence is quite simple, in line with our theory. In first stage, we use the platform-specific (m) audience size (${\alpha }_m$) to predict whether a platform-specific censorship event is likely to occur or not. For instance, the audience size for TV (country i, year t) instruments censorship events targeting TV (country i, year t), and so on:

(8) $p\left(Even{t}_{jm}=1\right)={\alpha }_m{\vartheta }_{it}^{\prime }+W{\gamma }_{it}^{\prime }+{\theta }_t+{\delta }_i+{\epsilon }_{it}$(8)

Since in our hypothesis, the broader the reach of the media, the harsher the consequences, in the second stage of our Two Stage Least Square (2SLS) model, contingent upon censorship occurrence, we estimate the level of platform specific severity of censorship:

(9) $Severit{y}_{jm}={\phi }^{2S}\widehat{p\left(Even{t}_{jm}\right)}+W{\mathrm{\partial }}_{it}^{\prime }+{\theta }_t+{\delta }_i+u_{it}$(9)

We believe that the exclusion restriction of our Media reach instrument is satisfied because of how our sample is constructed. In order for an observation to take values on the Severity of censorship scale, the censorship event by media type (Media Reach) needs to occur in the first place. In other words, it is mathematically impossible for our instrument (Audience size) to affect the dependent variable (Severity) through any other channel than Media Reach that captures platform-specific censorship occurrence, therefore being exogenous.

We also designed a placebo test in order to ascertain whether it is indeed the concept of pivotality or “threat” proxied by the size of the audience that is driving the occurrence of censorship rather than other platform related idiosyncrasies. Within the print platform, the placebo tests distinguish further between daily newspapers and weekly or bi-weekly magazines, based on the assumption (backed by our Gallup survey data) that significantly more people access dailies as a primary source of news compared to lower frequency magazines. If our theory is correct, controlling for the number of print publications and other factors in country i year t, the audience size increase should induce more censorship of dailies, but not of weekly or bi-weekly publications that have a significantly lower readership. The next section discusses our findings.

Estimation of Results

Hypotheses 1a-1d disentangled the main “pivotality” argument by individual outlet characteristics (Media reach, Collective action reporting, Foreign ownership, and media located Abroad). The micro-level findings are in general consistent with our comparative statics, statistically significant, substantively strong and remarkably stable across model specifications. The outlet-specific characteristics that capture pivotal audiences – Collective action reporting (H1d) and Media reach (as a proxy for the location of pivotal voter/mobilizer) (H1b) as well as, to some extent, Foreign ownership (H1a) – are by far the most robust factors that explain censorship occurrence and severity across regime types and in the full sample. Figure 1 plots the coefficients of the main variables when regressed on the severity of censorship.

Figure 1. Standardized coefficients for dimensions of media market segmentation (dependent variable: severity of censorship).

Note: Robust standard errors, clustered by country; Ordered probit models with country and year fixed effects

We also present a Heckman Selection Model (Table 1) with a full range of the Media reach variable (small to broad audiences) and include multiple alternative specifications and reduced form results in the Online Appendix (Tables A6, A7).¹¹

Table 1. Determinants of the occurrence (stage 1) and severity of censorship events (stage 2), Heckman selection model (full variable range)

	(1)	(2)	(3)
VARIABLES	Model 1 (full sample)	Model 2 (democracies)	Model 3 (non-democracies)
Heckman selection equation	Censorship occurrence	Censorship occurrence	Censorship occurrence
Stage 1
Country level transaction costs $\left(\tau \right)$ and rents (r)
Economic constraints $\left(\tau \right)$	−0.0200	0.147**	−0.104***
	(0.0288)	(0.0574)	(0.0335)
Legal constraints $\left(\tau \right)$	0.0165	0.0561	−0.0215
	(0.0244)	(0.0418)	(0.0328)
Democracy $\left(\tau \right)$	−0.182**	−0.484	−0.159*
	(0.082)	(0.390)	(0.0896)
GDP per capita $\left(\tau ,r\right)$	−0.857***	−0.607*	−0.507***
(log)	(0.189)	(0.337)	(0.182)
Population $\left(\tau \right)$	0.323*	0.750***	0.413*
(log)	(0.166)	(0.277)	(0.232)
Oil production (r)	0.00314	−0.102	−0.00866
(log)	(0.0344)	(0.0724)	(0.0325)
Country FE Year FE	Yes Yes	Yes Yes	Yes Yes
Stage 2 equation	Severity of censorship	Severity of censorship	Severity of censorship
Individual Level Transaction Costs $\left(\tau \right)$ & Pivotality ($\omega$)
State owned or influenced media $\left(\tau \right)$	0.148	0.207	0.109
	(0.115)	(0.256)	(0.130)
Media located abroad $\left(\tau ,\omega \right)$	−0.0870	0.158	−0.130
	(0.0807)	(0.266)	(0.0848)
Foreign ownership ($\tau ,\omega$)	−0.174***	−0.112	−0.211***
	(0.0557)	(0.147)	(0.0605)
Collective action reporting $\omega$	0.342***	0.468***	0.282***
	(0.0772)	(0.150)	(0.0903)
Media reach $\omega$	0.0244***	0.0351**	0.0256**
	(0.00940)	(0.0178)	(0.0111)
Constant	4.567	−2.023	1.757
	(3.228)	(5.599)	(3.819)
LR chi2	9.36	3.14	8.87
	(0.002)	(0.08)	(0.003)
Wald	45.55	14.78	40.39
	(0.000)	(0.01)	(0.000)
Observations Countries Years	4,633 169 12	1,456 87 12	3,177 101 12

Standard errors in parentheses; *** p < 0.01, ** p < 0.05, * p < 0.1.

Since we argue that media market segmentation should carry a similar logic across regimes, we find that audience pivotality is the only consistent determinant of censorship severity in both democracies and autocracies.¹² In non-democratic regimes only, domestic outlets as the primary source of information for a majority of the population are also more severely censored on average than foreign media (Table 1 and Figure A8, Online Appendix). Both variables capturing audience pivotality remain statistically significant despite all possible variations in regression models, and the magnitude of the effect is substantial. Keeping other covariates at mean, in both democracies and non-democracies, news platforms with smaller audiences (online) are 15% less likely than traditional platforms and journalists with broader audiences (TV and radio) to be severely censored, as opposed to receiving less harsh punishments.¹³ Impact-wise, coefficient differences across platforms are somewhat more pronounced in democratic regimes. The predictive margins for collective action reporting suggest that outlets and journalists covering protests, riots or demonstrations are around 30% more at risk of being the targets of threats and violence than those who report on other issues (Figure 2).

Figure 2. Predictive margins of media reach and collective action reporting (DV: severity of censorship) by political regime type.

Note: Ordered probit models with country and year FEs and additional controls [GDP pc (log); Population (log); Economic constraints; Legal constraints; Democracy; State owned or influenced media; Media located abroad; Foreign ownership]. Severity of censorship compressed to two values: 1 for Punitive censorship (values 1–9 of the Severity variable) and 2 for Severe censorship (values 10–12: Assault; Jail and Death)

As discussed in our econometric specification section, the paper also employs a 2SLS identification strategy out of endogeneity concerns. Table 2 shows the first and second stage results for Media reach instrumented with Audience size. Since our Media reach variable is ordinal and in its truncated form assigns values 1–4 to censorship events on a theoretically informed scale from low reach (Online; Print) to broader audience media (Radio; TV), we also decompose the Media Reach variable into its four constitutive dummies for robustness. The high F-statistic values show (with the notable exception of Radio) that the instruments are strong. The coefficient magnitude and signs confirm our leading hypotheses (TV is more severely censored than all other types, while Online and Print – less severely), and suggest that the ordinal scale we use for Media reach is not arbitrary.

Table 2. Instrumental variable (2SLS) and first stage estimation of severity of censorship by media reach and individual media type

	(1)	(2)	(3)	(4)	(5)
VARIABLES	Severity of censorship	Severity of censorship	Severity of censorship	Severity of censorship	Severity of censorship
Media reach (low to high)	0.346*** (0.138)
TV censorship (dummy)		0.775** (0.310)
Radio censorship (dummy)			−51.53 (287.8)
Print censorship (dummy)				−1.060** (0.432)
Internet censorship (dummy)					−3.053** (1.265)
Population(log)	−1.063 (3.374)	−1.017 (3.418)	−7.471 (43.84)	−0.612 (3.515)	−1.803 (3.233)
GDP per capita(log)	0.937 (0.850)	0.947 (0.859)	5.860 (28.32)	0.778 (0.850)	1.144 (0.870)
Democracy	−0.033 (0.143)	−0.0407 (0.144)	−0.0480 (0.644)	−0.0639 (0.146)	0.0267 (0.139)
First stage of 2SLS: DV- Media reach and Censorship events by media type
Instrument Audience size	0.028***	0.013***	−0.0001	−.009***	−.003***
	(0.001)	(0.0008)	(0.001)	(0.0006)	(0.0004)
Observations	1,499	1,499	1,499	1,499	1,499
Country FE	YES	YES	YES	YES	YES
Year FE	YES	YES	YES	YES	YES
Number of countries	52	52	52	52	52
Model	2SLS	2SLS	2SLS	2SLS	2SLS
F-stat	624.54	226.43	0.030	197.83	65.84

Robust standard errors in parentheses, clustered by country; *** p < 0.01, ** p < 0.05, * p < 0.1.

The second set of hypotheses (H2a-d) referred to the transaction costs of censorship and combined both micro and macro-level variables. Overall, we found weaker support for them (Table 1). The level of democracy (H2a) is significant with respect to occurrence, but not in terms of Severity of censorship. H2b-d proposed three micro-characteristics of outlets related to the difficulty (or micro-transaction costs) of censoring outlets and journalists, and our results lend significant support only to Foreign media (Figure A8, Online Appendix).

Finally, H3 and H4 proposed two classic hypotheses in the literature on media capture: the role of rents and of media diversity. In terms of generic economic rents, a higher GDP per capita is accompanied by a lower probability of censorship occurrence (Table 1), whereas Oil production as a country-level proxy of rents seems to be close to significance only in some models and types of political regimes, but not in the pooled sample (Table A18 in Online Appendix). Hypothesis 4 suggests that the number of outlets renders censorship less likely (N). This test is empirically hampered by the limited data coverage on the number of platform-specific outlets (TV, Radio and Print) as a proxy for diversity. Nevertheless, acknowledging such limitations, when we examine the relative importance of the number of outlets and the pivotality of audiences, it is the latter that stands out (Figure 3). While the number of outlets is not robustly correlated with the number of censorship events, the audience size (a proxy for pivotal audiences) predicts censorship.

Figure 3. Coefficient plot for platform specific drivers of censorship (audience size versus number of outlets) Note: Panel negative binomial models of censorship events by media type (count) with country and year fixed effects and following controls: Economic constraints, Legal constraints, Political constraints.

Table A7 (Online Appendix) uses alternative variables on traditional and online diversity (media perspectives) and corroborates this result.

Robustness of Results

We have included in the Online Appendix several additional tests that aim at ascertaining the robustness of our main results. First, since our baseline model pools the entire sample, we also show statistical results disaggregated by type of censor (government or third party). The coefficients are in general strong and stable (Table A8). Second, one potential critique of our key pivotality variable (Media reach) might be that it is constructed on an ordinal scale, therefore artificially inducing variation in the dependent variable. To address such caveats, Tables A10-A11 also present findings from Logit and Negative Binomial models, regressing platform-specific audience sizes for traditional media (TV, Radio, Print and Online) on platform-specific censorship occurrence and event numbers. Additionally, Table A13 shows that the severity of censorship is, as expected theoretically, significantly lower for online platforms compared with more frequented news consumption sources (i.e., TV). As detailed in the placebo test section, we also zoom into one platform only (print news) and find that an increase in the print audience size leads to a higher probability of print censorship occurrence in the case of daily newspapers, but does not affect magazines with less frequent and less widespread circulation (weeklies or bi-weeklies), confirming that pivotality is the driving theoretical mechanism (Table A15).

Third, we also demonstrate that when the events are weighted according to a theoretically informed scale that records the scope of censorship according to how pervasive events are and how many targets it impacts (One, Several, Many and All journalists and outlets), the results are stable and strongly significant (Table A9 in the Online Appendix). Fourth, since audience sizes exhibit time trends, additional concerns might stem from the non-stationary nature of the time series used for this analysis. Structural breaks (for instance, online technology becoming more widespread in selected countries) might endanger the stability of coefficients. We ran augmented Dickey-Fuller tests for unit roots and co-integration (Table A12), as well as Chow tests, and found that despite structural breaks (Tables A13 and A14, Online Appendix), the coefficients of our main specification appear to be stable. Platform-specific audiences determine the occurrence of platform censorship independently of structural breaks. Sections A.4 and A.5 of the Online Appendix also include additional sensitivity analyses, including separate sub-sample specifications for outlets and individuals to mitigate the potential concern that these two types of actors are targeted by distinct and non-overlapping censorship actions (Table A19, Online Appendix).

Conclusion

This paper asks whether the censorship strategies employed by censors vary across individual targets. Our leading hypothesis is that censors engage in a cost–benefit calculus of punitive action against the media based on the political threat posed by the audience. This balance leads to economizing censorship, by confining it only to segments of the media market that reach politically important constituencies and facilitate collective action that might hurt the survival prospects of the incumbent censor. These empirical results support the main logic underpinning recent insights on informational regimes (Guriev & Treisman, 2019, 2020). To test our theory of market segmentation, we constructed the first global dataset of salient censorship events affecting individual media outlets and professionals, in 196 countries between 2001 and 2015. The main results suggest that the risk of being censored is strongly correlated with audience pivotality and collective action coverage. Platforms that appeal to large audiences and report on collective action events are significantly more likely to be the targets of severe forms of censorship than media accessible to narrower segments of the market. The results hold across democracies and non-democracies, for government and third-party censors. They also confirm the persistent co-existence of information bubbles on media markets.

Open Scholarship

This article has earned the Center for Open Science badge for Open Data. The data are openly accessible at https://doi.org/10.1080/10584609.2022.2074587.

Acknowledgments

The authors would like to acknowledge two grants from the School of Public Policy at Central European University and the World Bank’s Development Data Group (generously supported by the UK Department for International Development). We are grateful to Giang Vu and Balint Nemeth for providing excellent research assistance. Eva Bognar, Cait Brown, Silvia Chocarro, Marius Dragomir, Craig Hammer, Macartan Humphreys, Kimuli Kasara, Staffan Lindberg, Ellen Lust, Jennifer Pan, Dean Starkman, Simon Rippon, as well as the participants of workshops and panels held at WZB Berlin, Central European University and the American Political Science Association annual meetings provided valuable insights and feedback on previous versions of the manuscript. The views and opinions expressed in this article are those of the authors, and do not necessarily represent those of the UK Foreign, Commonwealth, and Development Office or of the World Bank.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Supplementary Material

Supplemental data for this article can be accessed on the publisher’s website at https://doi.org/10.1080/10584609.2022.2074587.

Data Availability Statement

The data described in this article are openly available in the Open Science Framework at https://doi.org/10.7910/DVN/G442E0.

Additional information

Funding

This work was supported by the Central European University [School of Public Policy Research Grant] and World Bank's Development Data Group.

Notes on contributors

Cristina Corduneanu-Huci

Cristina Corduneanu-Huci (Ph.D, Duke University) is an Associate Professor at the Department of Public Policy of the Central European University, and a research affiliate of the CEU Democracy Institute. Her research interests include government transparency, international development and autocratic politics.

Alexander Hamilton

Alexander Hamilton (D.Phil. (Ph.D), University of Oxford) is an economic adviser at the UK Foreign, Commonwealth, and Development Office. His research interests include political economy and population economics, with a particular focus on fragile and conflict affected states.

Notes

1. We assume that the political importance of an audience depends on whether it includes the median voter (in democracies) or selector/mobilizer (in autocracies).

2. The Online Appendix (section A.1) includes a stylized formal model informing our theory.

3. Empirically, it is difficult to disentangle what is driving the propensity to censor foreign and diaspora outlets – audience pivotality (ω) or transaction costs (τ). Therefore, in our tests, we acknowledge the inseparability of the two dimensions in these cases.

4. The extended database includes a variable that records the scope of censorship events and assigns weights to One, Several, Many and All targeted journalists and outlets (Table B1, Online Codebook).

5. Figure B10, Online Codebook, defines and summarizes these categories, and the Online Appendix, section A.4 discusses alternative operationalizations.

6. Figure 2 in the empirical section shows results regarding this highly stylized typology of censorship events.

7. The concept of ownership (as shareholder control) is complicated and dynamic. For operational purposes, we code outlets and subsidiaries as domestic if there is no explicit reference in our global monitoring reports to editorial or managerial decisions taken elsewhere. Our coding protocol labels individual journalists as “foreign” if they explicitly report on a different country for a foreign/transnational outlet.

8. We expanded the government-owned media category to also include government-influenced outlets and journalists as identified in international media freedom reports.

9. Platform audience size is only an indirect way to account for “pivotal” elite or mass news consumers – the key parameters of our theory – but no other data is available to directly proxy these groups at event level in a cross-national format. Therefore, we use the relative size of audience segments across all traditional media platforms for which we have country-year data (TV, radio, print and online).

10. For instance, unlike European TV preferences, in many Sub-Saharan African countries radio is a primary source of news. Chronologically, in several advanced industrial democracies online media consumption increased sharply after 2008, while TV audience sizes decreased.

11. The full coded range of the Media reach variable is: 1 Other; 2 Citizen journalist; 3 Freelancer; 4 Blogger; 5 Online news channels; 6 Print; 7 Radio; 8 TV (Figure B1, Online Codebook). Most analyses of the paper truncate the data to include only traditional media platforms (5–8: Online, Print, Radio and TV).

12. As a cutoff, we use the conventional 7.5 value of the Hadenius-Teorell Democracy scale.

13. The statistical effects of punitive and severe censorship go into opposite directions as predicted by our theory. Since these are essentially two ends of the same ordinal Severity scale (< 10, and ≥10), when the probability of severe censorship increases, the punitive sub-category of events mathematically decreases.