THE DILEMMA OF CRITICISM: DISENTANGLING THE DETERMINANTS OF MEDIA CENSORSHIP IN CHINA

Abstract

Recent literature claims that China censors information that has the potential to ignite collective action. This article extends this finding by arguing that Chinese censors respond differently to political challenges than they do to performance challenges. Political challenges call into questioning the Party's leading role, whereas performance challenges are directed at the failures of public goods provisions. A survey experiment of about 60 media professionals finds that censors are inclined to block political challenges and to tolerate criticism of the government's performance. However, when criticism contains both performance and political challenges, censorship is far more likely. By exploring the range of censorship activities, the results suggest that the Chinese regime's reliance on popular support constrains its censorship decisions.

INTRODUCTION

What kind of information do authoritarian regimes block from the public? The answer to this question is important insofar as it provides us with a window to understanding how regimes perceive their own strengths and weaknesses. Conventional wisdom on censorship dictates that authoritarian regimes will censor any information critical of their hold on power (Lee 1998; Marolt 2011; Morozov 2011; MacKinnon 2013). More recent literature demonstrates that censors are quite tolerant of criticism, as long as it does not increase the potential for collective action, “regardless of whether they are critical or supportive of the state and its leaders” (King, Pan, and Roberts 2014, 1).

The collective action hypothesis is compellingly well supported by a growing body of evidence, but it cannot explain the full range of censorship decisions. Specifically, the collective action hypothesis cannot account for censorship in the absence of the danger of collective action, nor can it account for the lack of censorship of material with clear signs of collective-action potential.

This article proposes that we should reconsider the effects of political criticism in a more nuanced way. It assumes that the Chinese regime tries to maximize its popular support for the Communist Party's continued monopolized rule. To make a censorship decision on a certain piece of content, censors need to consider whether this piece has potential to reconstitute an attack on the one-party rule. Thus, China's censors differentiate two types of criticisms: challenges to the Party's political leading role and those to the performance of the government—i.e. the provision of public goods to the public.

I argue that these two types of criticism generate different levels of censorship. Criticism of the Party's rule is censored more intensely because it directly challenges the standing of the Chinese Communist Party. However, criticism that targets the government's performance is more tolerated. Pacifying such criticism requires the government to respond actively and positively. Increasing censorship of performance challenges would stimulate further criticism, by showing that the government can do nothing but cover-up their failures. It would undermine the government's effort to demonstrate its responsiveness (Distelhorst and Hou 2017).

This logic generates clear empirical predictions, which I test using a survey experiment involving Chinese media professionals, who are constantly exposed to guidelines from the government censors. By doing so, I capture the censorship happening before publication and I am able to provide evidence of censorship using a lens other than social media: the professional news outlets. The results confirm that, in addition to its potential for inciting collective action, criticism of the Party's rule—or “political challenges”—increases censorship intensity. Performance challenges do not have this effect on their own, but they increase censorship intensity when the political challenges are also strong. When authoritarian regimes need to maximize popular support, they tolerate criticism that highlights the flaws of governance. By responding rather than censoring those flaws, the regime can maintain the popular support (Tang 2016). Once the criticism assails the fundamental components of the regime's rule, censorship will be exerted comprehensively.

STUDYING MEDIA CENSORSHIP IN CHINA: DIVERGENT EFFECTS OF POLITICAL AND PERFORMANCE CHALLENGES

Studies of media censorship in China follow two traditions. The first tradition is interested in the institutions, mechanisms, and strategies that the regime uses to block unfavorable information. For example, Zhang and Fleming (2005) examine the institutional channels of the regime to control media coverage of SARS. Other studies are interested in how the Chinese regime adapts to the spread of Internet service. This literature demonstrates that the regime is confident in taking advantage of this technology while minimizing its impact on the one-party rule.

A more recent line of research explores the breadth of censorship across content in China, such as political terms in social media, blog posts, and censorship directives (Bamman, O'Connor, and Smith 2012; Fu, Chan, and Chau 2013; King, Pan, and Roberts 2013, 2014; Tai 2014). Some of these studies argue that the “CCP tends to ban news that directly threatens the legitimacy of the regime” (Tai 2014, 188). In contrast, King et al., provide a collective-action explanation. Censorship is caused by the fear of collective action. The regime has the goal of “restricting the spread of information that may lead to collective action, regardless of whether or not the expression is in direct opposition to the state and whether or not it is related to government policies”(King, Pan, and Roberts 2013, 328). This logic is consistent with the idea that authoritarian regimes encourage a certain degree of freedom of speech to provide incentives to spur local bureaucrats, reduce corruption, and guide public opinions (Egorov, Guriev, and Sonin 2009; Lorentzen 2014); it explains why some critiques of the state remain uncensored.

The main drawback of previous theories is that they failed to recognize that different types of criticism may associate with different dimensions of legitimacy, which leads to different concerns for the censors. This article contends that criticism of the regime may or may not affect censorship given its different natures. I identify two types of criticism—criticism of the Party's authoritarian rule and criticism of the government's performance. Criticism of the Party goes against the Party's principle as a “vanguard party” that exercises the “leading role” over the country—the foundational ideological justification of the party-state (Joseph 2014, 13–14). I call such criticism “political challenges” because it targets the regime's political power. For example, a radical version of political challenge will claim that one-party rule should be replaced by a multi-party system. A more moderate version may criticize the party-state, claiming that it imposes to many constraints on personal freedom or political liberty. Widespread dissemination of political challenges will erode the effects of ideological control, damage popular support for the Communist Party's rule and facilitate the mobilization of its opponent's anti-regime movement. Therefore, we should expect the censors to work hard to block information that challenges one-party rule, i.e. to the Party's political leadership in the country.

Hypothesis 1 (H1)

Strong political challenges increase censorship intensity.

However, not all criticism directly aims to demolish the one-party rule. A significant portion of criticism of the government in China only involves complaints regarding failures to deliver decent public service. In this article, I call such criticism a “performance challenge.” Performance challenges emerge when the public is convinced that the government is failing at providing public services, such as environmental protection, unemployment safety nets, or education. The government can appease such challenges by improving policy outcomes. Actions such as providing general welfare, responding to natural disasters or economic crises, and enhancing responses to the public's needs will attenuate people's anger.

Censoring performance challenges will tarnish the government's image, because censoring a policy issue could discourage citizens from providing valuable information to help fix the governance problems; thus, censorship damages the capacity of the government to provide quality public services (Lorentzen 2014; Dimitrov 2013). In the long term, censoring performance challenges generates a greater damage to popular support of the regime. ¹ Therefore, the regime would rather respond to citizens challenges than block relevant information. For better policy output, the regime welcomes criticism of the government performance (Nathan 2003; Teets 2013; He and Thøgersen 2010; Dimitrov 2013).

Hypothesis 2 (H2)

Strong performance challenges have no effects on censorship.

The conceptual distinction between “political” and “performance” is also supported by previous research, such as Lipset's differentiation between “the effectiveness” and “legitimacy” of a political system (Lipset 1959, 86). Such distinction is also consistent with past studies of the legitimacy of the Chinese Communist Party. Guo Baogang distinguishes two justifications in Chinese political legitimacy:

Original justifications refers to the origin of the ruling authority, and the utilitarian justification defines the rulers’ staying power or capacity to maintain people's belief in their ruling authority. Original justification may derive from a divine being , or from a leader of moral characters or some unique quality, or simply from the will of the people … The utilitarian justification derives from the capacity of the rulers to meet people's needs, such as material well-being or physical security. (Guo 2003, 3; emphasis added)

In the context of censorship, criticism does not have to be an explicit condemnation, it could also refer to a piece of information that highlights the government's problems, provides resources and reasoning for fault-finding, and encourages disapproval. For example, a primary school student wrote a letter to advise President Xi Jinping on his weight management. This news, which was quickly censored on Chinese Internet, by no means intended to criticize Xi's body shape or embarrass the President. ² However, it can still constitute a political challenge because the letter facilitates a potential attack on Xi and thus the Party leadership. Therefore, in this article, political and performance challenges also include topics that do not directly attack the Party or the government, but that may facilitate such attack.

Finally, this article sees political criticism's effect on censorship as orthogonal to that of collective-action potential. Since collective actions cause uncertainty over social stability, the threat of collective action should elevate censorship intensity no matter what the nature of the criticism. This article thus builds on the work of King et al. and uses collective-action potential as a control variable.

ASSESSING CENSORSHIP ON POLITICAL AND PERFORMANCE CHALLENGES

The main obstacle to the study of censorship is the shortage of unbiased data. Since governments rarely publicize their censorship strategy, most studies rely on indirect observation of content deletion or, when lucky, leaked directives. This approach to data gathering leads to selection problems. Tai analyzed 1,403 censorship directives from China Digital Times, but the author admitted that it was hard to assess the accuracy of the data set. Those directives were reported by anonymous journalists, and thus the sampling process was not random (Tai 2014, 192). And in their observational study, King et al. (2013) admitted that their large-scale data set could not capture the self-censorship and censorship that occurs before they obtained the posts.

King et al. (2014) tried to fix this problem by using a participatory experiment; they used real events to manipulate “pro-government,” “anti-government” and “collective action potential” in their experimental blog posts. Inspiring as their work is, real events may confound the variables of interest for this study. For instance, it is unlikely to find two events that have the exact same level of political challenges while happening to be varying in collective-action potential. In fact, among the four collective-action events King et al. tested, the Panxu protest involved low-ranked bureaucrats and local government (i.e. it had a weak political challenge to one-party rule). The other three events, Tibetan self-immolations, Ai Weiwei, and protests in Xinjiang, all severely challenged the one-party rule because they defied the Party's ethnic policy or democratic claims (Ai Weiwei). Similarly, it is hard to disentagle the effects of performance challenges from other confounding factors in real events. Hence, using real cases cannot guarantee that only one variable is manipulated. Online observations of censorship, constrained by the complex setting of real cases, may either generate selection bias or confounding factors that problematize the internal validity of the empirical test.

Therefore, this article uses hypothetical scenarios as vignettes in a survey of media professionals. To minimize confounding factors, I manipulate the variables of interest by changing only the key terms in each scenario so that I can obtain higher internal validity. Nevertheless, this research design cannot discredit the contribution of observational studies conducted in social media that can test theories of censorship in the realistic setting and reach a better external validity. In the next section, I will discuss the operationalization strategy first and then provide more details on the experiment process.

Operationalizing Political Challenges

Table 1 summarizes the operationalization strategies for political and performance challenges. The regime's maintenance of one-party rule suggests that censorship should protect the reputation of agencies that are essential to the Party. In regard to any specific public topic, the higher the “institutional affinity” towards the central party leadership, the stronger political challenges it provokes. In this article, institutional affinity is first measured by how close an organization is to the Communist Party's core leadership. Government agencies are closer than social organizations. Military organizations are closer than government agencies. Misbehaviors of a PLA official or a party cadre will generate more challenges to the Party than a public servant in the government. The rank of cadres also determines institutional affinity. Higher-ranked officials tend to generate stronger political challenges than lower-ranked officials because they are closer to the central leadership of the Party, which implies that their behaviors and reputation are more tightly bound to the reputation and image of the Party. For instance, the party's reputation is more seriously undermined by a corruption scandal of a Politburo member than of a township party secretary. Thus, the former has higher institutional affinity.

Table 1 Operationalization of Political and Performance Challenges

Hypothesis 1.1 (H1.1)

Higher institutional affinity leads to higher censorship intensity.

Additionally, political challenges are also determined by the degree of (dis-)conformity to the Party's ideological claims. Ideology plays an essential role in maintaining popular support of one-party rule (Joseph 2014, 186). If a given topic is framed in a narrative that promotes the Party's legitimacy, its dissemination is less likely to arouse opposition to the Party. The censors will be more tolerant of ideology-conforming narratives than they will be of ones that advocate Western liberal democracy. Regarding the topic of anti-corruption, the claim that “the Party should impose stronger supervision,” generates a weaker challenge than the claim “the Party should allow free media and universal suffrage.” This is because enhancing supervision is consistent with the Party's commitment to tackling corruption problems, while free media and universal suffrage belong to the discourse of a “Western style of Democracy” (Miks 2011). To give another example, when a corrupt official was arrested, a blog post might blame either the inner-party factional struggles or the flaws of supervising institution (Jiandu Zhidu). Placing blame on the former generates a stronger political challenge. The Party's principle of reform and openness tolerates the idea of a need to improve the current institution, while factional struggles go against the fundamental claims of the Party—a claim of factional struggles erodes the image of a unified party that represents universal interest of people (Phillips 2016). Furthermore, in the international realm, criticism of ideologically close countries like North Korea will also generate stronger challenges than ideologically remote countries like Japan. Finally, news about the government's mistreatment of a minor ethnic group will also result in strong political challenges, because it challenges the Party's claim to include care for minorities in their ethnic policies.

Hypothesis 1.2 (H1.2)

Higher ideological disconformity leads to higher censorship intensity.

Operationalization of Performance Challenges

In order to confirm that performance challenges have no effect on censorship, this article also varies the performance challenge levels in the test. First, the Party is more likely to respond to greater grievances (Lorentzen 2017). A public event that does a high degree of harm to citizens, will generate stronger performance challenges, and the regime will be more motivated to keep the discussion of the event open and transparent. The degree of harm can be measured by the number of casualties and the geographical sphere of impact. For instance, an accident involving five injured people, has a weaker degree of performance challenge than the one involving 35 injuries. An accident affecting a single city will result in a weaker performance challenge than one that affects multiple cities. This operationalization inevitably captures the scale of an event, but the degree of harm is more a measure of the delivery of public services, such as public safety or general welfare.

Second, the government is under more pressure to demonstrate quality of performance in policy areas where it has made promises than in those it has not. By fulfilling its policy promises the regime demonstrates its competence in governing—its ability to meet the people's needs. But failure to meet those needs may give social actors reason for resistance against the government (O'Brien and Li 2006). Thus, performance challenges are stronger when the public expects the government to solve a problem it has promised to solve than one it hasn't. For example, the Chinese government has promised to protect the environment, but not to report officials’ assets immediately to the public. It is committed to protecting the education of poor children, but not to granting legal marriage rights to gay couples. Thus, we can expect that criticisms of the environment and education will generate stronger performance challenges than criticisms of official assets or gay marriage rights.

Operationalization of collective-action potential

Finally, this article manipulated collective-action potential in two ways. First, at the topic level, nine out of 22 experimental scenarios are collective-action events, such as nationalist protests and environmental movements. Ten scenarios do not involve any collective-action potential, such as an official's corruption scandal or a policy issue discussion. Second, this article also manipulates collective-action potential at the framing level by adding a call for collective action to the collective-action version of a scenario. ³

Operationalizing censorship

To reduce confounding factors mingled in real cases, this article used most-similar hypothetical scenarios to manipulate the independent variables. This method, however, generated practical problems in testing their censorship intensity: it was not reasonable to conduct this experiment by posting blogs online and observing how many of them are censored. Because the scenarios were hypothetical, they may not be taken seriously by either the public or the censors. Blog posting also raises an ethical problem—posting hypothetical cases is equivalent to spreading rumors. My solution was to directly ask the professionals who have sophisticated experience with censorship at their work.

Thus, this article measures the intensity of censorship by asking Chinese journalists and editors, as a proxy of the censors, about their perception of the sensitivity of the given scenarios. I recruited journalists via email and social media (who have identified their actual profession).

The advantages of surveying media professionals are: (1) They are acquainted with censorship patterns. Journalists and editors receive propaganda instructions at work. The censorship authority asks them to follow these instructions strictly as they publish their articles. This makes them best able to approximate the propaganda cadres’ mindset. In fact, many journalists may end up being propaganda cadres in the Party system themselves. The two top leaders in propaganda affairs in the CCP in 2016, Liu Yunshan and Liu Qibao, both have work experience in the media. ⁴ (2) This measurement also captures pre-censorship or self-censorship—censorship that happens before researchers could collect data. (3) Lastly, journalists are simply more accessible than propaganda officials (the censors).

The survey asked media professionals to evaluate the “political sensitivity of publication” of different hypothetical scenarios. They answered by using a 7-point scale:

Score 1 (not at all sensitive): Any information about this event will be allowed, and journalists may report without any constraints. Blogs discussing the event online are unlikely to be deleted.

Score 3: Journalists may report, but with constraints on the in-depth investigation or follow-up stories, or within a strict narrative framework. Websites are allowed to publish the story, but may not be allowed to post it on the front page. The government will give an instruction of “no hyping” for this event. Some blogs on this topic are blocked when they are related to more sensitive information.

Score 5: Journalists are not allowed to do interviews on their own. Media are only authorized to publish the standard article from Xinhua Agency, CCTV or People's Daily. The negative comments on websites are all blocked.

Score 7 (extremely sensitive): No newspaper publishes it. Even Xinhua Agency, CCTV or People's Daily are trying hard not to respond to it. No blogs or articles related to this event are allowed to survive online once the censors find them.

I use the scores above to construct the dependent variable “censorship intensity,” ranging from 1 to 7. Score 2 to Score 6 indicate that news articles could be deleted and blogs are blocked, but relevant information is still accessible. Accessibility reduces when the score is higher. Score 7 indicates that the censors do not want to disclose any information. This design of censorship intensity score is able to capture the diverse censorship strategies adopted by the Chinese authorities.

The survey of media professionals allows us to observe the patterns of censorship in professional media settings, where censorship guidelines are usually kept as a secret. The survey experiment design also allows better internal validity. However, it also has several limitations. Although hypothetical scenarios can better discern the effects of variables, this is at the expense of limited external validity, compared to studying real cases. In addition, surveying journalists provides an indirect measure of censorship, increasing the potential errors compared to direct observations of social media. After all, media professionals are not censors, although I would argue that they are the best available respondents. If the decisions made by censors are an accurate representation of censorship, then the response of media professionals may potentially have errors. Since this is a survey of experts, the respondents were not randomly selected. The majority of respondents come from non-state-affiliated media, in which the censorship standard may be different from the state-affiliated ones. In the analysis, I controlled the respondents’ workplace and found that results are consistent. I will discuss how these weaknesses may impact the results in the next section.

THE EXPERT SURVEY ON CHINESE MEDIA PROFESSIONALS

Experiment Design

The survey experiment was conducted from January 26 to May 12, 2015 via Qualtrics survey service. In the survey, I created 22 experimental scenarios, each containing two versions with one manipulated variable or four versions with two manipulated variables. ⁵ Respondents were randomly assigned to only one version per scenario. Fourteen scenarios have two versions and eight scenarios have four versions. Therefore, there are in total 2¹⁴ × 4⁸ combinations of versions of scenarios to finish the survey. No participants had read the exact same two combinations of materials. For each scenario, I made the versions as similar as possible, and only manipulated the key value of political challenges, performance challenges, or collective-action potential. All independent variables were coded as either 0 or 1. For example, one scenario testing political challenges was about the corruption scandal around a public official. There were two versions. The wording of the two versions were almost identical, but in Version One the official is a director of a provincial department (weak challenge, coded 0) whereas in Version Two the official is a minister in the central government (strong challenge, coded 1). ⁶ I also randomized the order of questions to prevent any biased priming effects generated by the fixed order of scenarios or by one respondent exposed to two or more versions in the same scenario.

The survey was distributed via both email and snowballing on social media. According to the report from Qualtrics, the email recruitment message reached 354 inboxes, 141 of which were read. Forty-one surveys were started while only 18 were completed. The snowballing recruitment via social media reached around 100 journalists, of which 40 completed the entire survey. In total, I collected 68 responses that answered at least one question (1,390 questions answered); 59 of them finished most questions, including two demographic questions at the end of the survey. Since I need to include demographic variables as the control, I only use these 59 responses (1,293 questions answered) in the regression analysis. The results remain similar with or without control. ⁷

I used the respondents who read my invitation for the survey as the denominator. The response rate is around 0.25 (59 out of 241). This rate is acceptable since the content of survey is sensitive in China. Respondents have reported an average of 8.6 years (Std.Dev = 6.1) of work experience in Chinese media. ⁸

One potentially influential factor in the respondents’ estimation of censorship intensity, might be who they work for. This survey categorizes their workplace by how closely the type of media is supervised by the party-state. Figure 1 shows that 56 percent of the respondents work for marketized media, while 12 percent work for private news websites/social media. These media outlets have a looser connection with the Party and have more flexibility in publication (coded 0 in the state-affiliated variable). Such media is the product of media marketization, which provides “incentives for media practitioners to overstep boundaries of news reporting” (Stockmann 2013). The rest of respondents (about 30 percent) are from TV or Radio, Party Newspaper, and state-owned websites—media outlets that the Party controls more directly (coded 1 in the state-affiliated variable).

Figure 1 Respondents’ self-reported employer

This sample is by no means representative, and I expect it will generate social desirability bias, selection bias on respondents’ background, and non-response bias. As I am disinterested in the personal effect on the estimation of censorship intensity, I used several methods to reduce the bias generated by individual preference. First, I randomly assigned respondents to scenarios and allowed respondents to answer only one version per experimental scenario. Two individual-based variables—respondents’ workplace and working experience—are balanced across three treatments: political challenges, performance challenges, and collective action. ⁹ Since respondents did not have a reference from another version of the same scenario, they could not intentionally underestimate or overestimate the effect of performance and political challenges within each scenario.

However, respondents could still bias the absolute censorship intensity by scoring all answers lower or higher than that they actually believed. Since I did not inform respondents which variable I tried to test in the survey, such bias was unlikely to correlate with the independent variables. In addition, respondents did not intentionally concentrate their answers to either “not at all sensitive” or “extremely sensitive.” The responses to all scenarios are approximately normally distributed (Mean = 3.73, Standard Deviation = 1.94, Skewness = 0.18).

Also, I use the answer to one scenario per respondent (i.e. one question per respondent) as my unit of analysis and add fixed effect to each scenario, so that I can compare both their average answer scores across different versions of a single scenario and the average of each scenario across different scenarios. This procedure allows my results to be less dependent on respondents’ personal answers. I also controlled for a respondent's work experience (measured by self-reported working years) and workplace, and clustered the standard errors at each respondent, in order to capture error dependencies at the individual level.

Experimental Results

Table 2 shows a tabulation of the mean censorship score between political and performance challenges without controlling collective actions or including fixed effects for each scenario. Cases manipulated to have stronger political challenges were estimated to have higher censorship intensity. The difference is 1.00 when collective-action potential is strong, 1.72 when collective-action potential is weak. Strong performance challenges, on the other hand, do not stimulate higher censorship intensity. The average censorship intensities of high performance challenges are actually less than the low ones across high and low collective-action potential. In the absence of strong political challenges, strong performance challenges have inconsistent effects on censorship. However, when combined with strong political challenges, strong performance challenges contribute to increased censorship intensity. The simple tabulation is consistent with Hypothesis 1 but not Hypothesis 2.

Table 2 Tabulation of Experiment Result

Note: The upper panel shows the results when collective-action potential is low, the bottom panel shows the results when it is high. Frequencies are in parentheses. N = 1390

Figure 2 displays the coefficient plot of the OLS regression analysis. The baselines are low political and performance challenges as well as low collective potential. Political challenges and collective-action potential significantly increase censorship intensity by 1.26 points and 1.50 points (in the range from one to seven). If we assume that nothing is censored when censorship intensity equals 1 while everything is censored when censorship intensity equals 7, strong political challenges could censor 21 percent more relevant content and high collective-action potential will increase censorship by 25 percent. As the theory predicts, strong performance challenges have no significant effects on censorship intensity. Working experience and workplace are two respondent-level variables and they do not affect censorship intensity. ¹⁰

Note: CI = 95%. Unit of analysis is the answer to per scenario per respondent. Dependent variable is 7-point scale censorship intensity. Baseline is when all variables manipulated are ‘low’ (coded as 0). All models have fixed effect on each experimental scenario and standard errors clustered at individual respondents.

Figure 2 Coefficient Plots of Regression Analysis

However, the experiment has limited scenarios. It is difficult to manipulate variables of interest in a perfectly random way across all scenarios, while still maintaining the scenarios as realistic. ¹¹ Therefore, I cannot guarantee that all combinations of variables have equal likelihood to be assigned to respondents. Therefore, I use nonparametric conjoint analysis to examine the robustness of the results (Hainmueller, Hopkins, and Yamamoto 2014b). ¹²

Robustness Check

Table 3 shows the results of conjoint analysis estimated using R package “cjoint” (Hainmueller, Hopkins, and Yamamoto 2014a). The first model is the replication of regression analysis by using the conjoint estimator. ¹³ The results are similar: political challenges and collective actions have positive effects on censorship intensity while performance challenges’ effects are minimal and insignificant. The second model interacts political challenges and performance challenges in the analysis. The effects of political challenges and performance challenges remain unchanged, while performance challenges have effects that seemingly violate Hypothesis 2. When political challenges are weak, performance challenges, as expected, have no impact on censorship. However, when political challenges and performance challenges are both strong, censorship intensity increases. Model 3 adds collective action into the mix, but none of the interactions between collective action and the two challenges generate a significant effect. On the other hand, the interaction effects between political and performance challenges found in Model 2 still exist in Model 3.

Table 3 Robustness Check: Conjoint Analysis on Interactions

Note: The baseline of all models are weak political challenges, weak performance challenges, and weak collective-action potential. Working Experience is recoded into binary by its mean (8.6). All models are estimated with fixed effects on each experimental group (not displayed in this table). All models have robust standard errors clustered on respondents. *<0.05, **<0.01

In summary, the conjoint analysis confirms that the positive effects of political challenges and collective-action potential are robust predictors of censorship. It also shows that strong performance challenges do not stimulate censorship independently. Only when strong performance challenges coincide with high political challenges, do they increase censorship concerns. Although this finding is not fully consistent with the theoretical prediction, it has implications to realistic censorship scenarios. For example, if a leader of the Communist Party fails to fulfill his promise on a policy issue, criticisms of the policy may finally spill over onto the Party leader, which generates distrust of the Party; and this may “induce demand for systemic (regime) changes.”(Li 2011, 291) This mechanism could be illustrated by China's strict censorship of the news of the Panama Papers scandal (Forsythe and Ramzy 2016). Although the regime has committed to curbing corruption, unconstrained media reports about the Panama Papers may draw the public attention to the fact that many high-ranked party leaders are involved in the scandal. Since high-ranked leaders are supposed to represent the Party's image of defending the public rather than private interest, the Panama Papers may expose the leaders’ corruption and impair the Party. Thus, the regime had to censor the story.

Moreover, strong performance challenge may also draw enormous public attention, making it easier to find flaws in the propaganda of the Party's rule. For instance, the ethnic conflict in Xinjiang, in 2009, which resulted in at least 197 deaths and a thousand injured, led to the blockage of Twitter, the shutdown of Fanfou, and the disconnection of Xinjiang Autonomous Zone from the Internet for 10 months (Ward 2009; Wong 2010). In this case, the high degree of harm generated a strong performance challenge as well as tremendous public attention. It also questioned the Party's rule by revealing the severity of ethnic hostility, which deviates from the Party's propaganda that all ethnic groups under the Party's rule live peacefully together.

One alternative explanation of censorship's indifference to pure performance challenge is that performance challenges usually generate greater public attention. Thus, things that are too hard to hide are less prone to censorship. However, the regime also has a record of censoring material that received too much attention. For example, the viral environment documentary “Under the Dome” was censored after being viewed hundreds of millions of times. ¹⁴ Political challenges with potential to generate tremendous public attention, such as the Panama Papers story, are also heavily censored. In the case of events that have attracted too much public attention, censors may also order media not to conduct further reporting. ¹⁵ In general, censors seem unlikely to relax their guard when the focal event generates too much public attention.

The Potential Bias from Respondents

Both the regression and the conjoint analysis show that respondents’ individual backgrounds do not affect the variables of interest in this article. This makes sense, because even if survey measures suffer from social desirability or other forms of respondent bias, the randomized treatment design ought to filter these errors out of the main effect estimates.

Randomization, however, does nothing to deal with issues of representativeness and non-response bias. But even if such sampling biases are present it is not obvious how they undermine the analysis. If some professionals refused to take the survey because they are have no experience of censorship, missing their answers will not compromise the results with bias. Even if they had answered the survey, they would only add more noise rather than more truth. Only if non-respondents (1) are more familiar with the censorship patterns and (2) have a uniquely different perception of censorship patterns could the survey results be seriously biased. These non-respondents could be experienced media professionals as well as censors.

Given that there is no reliable way to measure how well this sample identifies censorship patterns, the sampling bias may affect the accuracy of the scale of effects. For example, my respondents could generally be more optimistic than the “truth group” and may thus estimate a lower censorship intensity score than the true one. However, the main purpose of this article is to compare the censorship intensities at two types of challenges and collective-action potential, all of which I manipulated via experimental design. Therefore, respondents’ optimism (or pessimism) about censorship is not likely to correlate with these variables. The sampling problem may thus lead to an inaccurately estimated censorship intensity without biasing the effects of manipulated variables. ¹⁶

Nevertheless, since this survey is already sensitive, it did not include questions about respondents’ other personal information such as political beliefs, age, or party membership. Further studies are needed to study how individual judgment may affect the censorship perception.

CONCLUSION: DUAL EFFECTS OF POLITICAL CRITICISM

This article adopts a survey experiment to explore the causes of media censorship in China. The experimental design with hypothetical cases minimizes the confounding factors and respondents’ bias that appears in the analysis of actual cases. The survey of journalists also captures not only the censors’ standard but also pre-censorship—censorship that stops information from being published and the self-censorship of media professionals. Different from previous works, which focus on social media, this article adds new evidence of censorship in professional media outlets. It indicates a new way to collect censorship data in future studies.

The survey experiment reconfirms that the regime is concerned about information regarding collective action, as King et al. (2013, 2014) argue. Furthermore, the findings show that different types of criticism towards the regime can also affect censorship intensity. This article differentiates two types of criticism: political challenges and performance challenges. The theory predicts that political challenges provoke more censorship than performance challenges. The empirical results, however, suggest that strong performance challenges have no impact on censorship when political challenges are weak. When political challenges are strong, performance challenges magnify the severity of political challenges and increase censorship intensity. Such a finding is consistent with anecdotal evidence about the harsh repression of the Panama Papers discussion and the termination of Internet service during Xinjiang Riot. Nevertheless, more work should be done to theorize why and when performance challenges deteriorate the effects of political challenges.

This article's framework shares common ground with information theory of censorship that regards the autocrat's (the center's) tolerance of constrained watchdog journalism as a way of monitoring local bureaucrats (Lorentzen 2014, 403). Indeed, by tolerating performance challenges, the regime facilitates its access to local information so that it can monitor its agents more easily. However, while information theory considers that the regime's goal is to prevent revolt, this article argues that the regime aims to maximize popular support. Moreover, information theory treats all negative news as a threat to the regime's rule, but I argue that a portion of negative news—performance challenges—do not produce a direct threat to the regime. Even if citizens take to the streets to protest against the government's failure to fulfill its responsibility, the regime can still pacify them and win popular support back by offering compromise and punishing the responsible officials.

Admittedly, the regime cannot offer compromises or punish local officials in every case. There are two possibilities under such circumstances. First, the regime may not be capable of solving the performance challenges, even though it hopes to. Under such circumstances, censorship is still unnecessary unless the social activists use such failure to criticize one-party rule. In other words, unless political challenges increase with such performance challenges, the regime does not need to censor. The interaction analysis in Table 3 demonstrates this scenario—censorship intensity increases when strong performance challenges are accompanied by strong political challenges.

Second, the regime does not compromise when it is faced by a political challenge. For instance, when a local protest breaks out against a policy issue, offering compromises may weaken the regime's control of public affairs. Likewise, punishing officials may disrupt the internal stability of the bureaucratic system and damage local officials’ loyalty. In such cases, the regime will choose to censor relevant information because responding to the protesters weakens one-party rule. In contrast, if the center believes that punishing officials will not disrupt the bureaucracy, it will not censor its punishment of officials. In summary, in both circumstances, political challenges are the determinants of censorship intensity, not performance challenges.

This article adopts the survey experiment design, which provides a better measurement of different dimensions of political criticism and thus enhances its internal validity. However, its external validity is restricted, given the limited variety of topics it was able to cover. The sample of respondents was also small and not representative. The respondents mainly came from traditional media, which may limit the theory's applicability to other forms of media. Exploratory as it is, this article should inspire the studies of censorship to several promising directions: first, a greater variety of topics and respondents are needed to test the external validity of the framework, especially in other authoritarian countries relying on popular support. Second, the experiment only statically models regime censorship. A dynamic model is necessary to fully estimate whether censorship intensity may change over time. Finally, this framework expects that censors send the same censorship guideline to all types of media platform. The main empirical evidence in this study comes from the professional media, whereas previous work has focused on social media. Although anecdotal evidence suggests that strong political challenges are censored in both media channels, it is likely that the regime's censorship intensity changes with different media platform. ¹⁷ This limitation needs to be further explored in the future.

Beyond identifying patterns of censorship, the findings of this article also contribute to our understanding of China's state–society relationship. As long as the authoritarian regime's survival relies on popular support, its excessive use of censorship could be counter-productive. The regime realizes that censorship has dual effects on popular support—it is a “necessary evil” to maintain the regime's legitimacy, but abusing it will also endanger the regime's rule. To maximize popular support, the autocrat needs to tolerate criticism and increase responsiveness as long as the criticism does not spill over onto its fundamental legitimacy—the one-party rule. The constraints on the regime's censorship may create space for social actors to expand their activism. For example, although the regime regards collective actions as a threat to stability, it will become more tolerant if these actions make only moderate challenges to political issues. If the actions have strong challenges on performance issues, the regime is compelled to respond efficiently; it means that social actors could make the state hear them while staying safe from oppression by framing their claims strategically. Such tactics are consistent with the previous findings on peasants’ “rightful resistance” (O'Brien and Li 2006; Chen 2012).