Speaking up or staying silent? Examining the influences of censorship and behavioral contagion on opinion (non-) expression in China

Abstract

Despite being designed to go unnoticed, censorship apparatus would occasionally manifest itself under various circumstances. In this study, we formulate four layers of censorship exposure where individual users can come across censorship. We investigate how different layers of censorship exposure influence users’ opinion expressions. Results show that people tend to stay silent when the censorship in the global environment is intensive, whereas they tend to “rebel” against censorship by voicing their opinions, when they experience censorship themselves or witness censorship occurring to their friends or reference persons. We also find community acts as a critical buffer against the influences of censorship. Outspoken crowd could shield individuals from the fear of punishment and outspoken friends could mitigate individuals’ anger against censorship. In either case, individuals can be liberated from their overconcern with censorship and be empowered to act for themselves.

Keywords

Censorship chilling effects Chinese social media opinion expression

The history of Internet regulation in China can be described as a never-ending cat-and-mouse game between authorities and ordinary Internet users (Endeshaw, 2004; Gleiss, 2015). Concurrent with the increasing social media penetration, China’s censorship apparatus has grown more and more sophisticated in order to keep information flow under control (Creemers, 2017; Jiang, 2016a; Jiang and Fu, 2018; Yang, 2004). So far China has the largest number of Internet users in the world (CNNIC, 2019), as well as the most refined and most far-reaching information control regime (King et al., 2014; Sullivan, 2014), which spans across public and private sectors.

According to our search in a national database of organizations (Supplemental materials, Note S1), at least 1196 Internet surveillance centers exist in the public sector across the country. However, online opinion, characterized by excessive volume, velocity, and variety (Kitchin, 2013), has far exceeded the capacity of public institutions (Roberts, 2018; Ye, 2019), urging the overloaded government units to outsource some tasks to private companies. Partly because of this, a new industry of censorship is booming. In our previous search, 211 companies, including state-owned enterprises, have been registered to conduct various types of opinion monitoring businesses. The leading company in this niche People.cn, which is a subsidiary of People’s Daily, has increased its net profit by 239% to 214 million in 2018 and has planned to recruit 3000 censors by 2021 (People.cn, 2018). In addition to specialized companies, all Internet content providers, from small discussion forums to national media platforms, are required to employ in-house censors in accordance with the government’s regulations (Fu et al., 2013; Zhuang, 2016).

With such concerted effort between state authorities and technology companies, censorship apparatus has been advancing unprecedentedly fast and has become an eminent issue concerning Chinese social media. However, because of the closure of Chinese Internet environment and the technical difficulty in detecting censorship (Jiang and Fu, 2018; Stockmann, 2018), our understanding of how censorship influence an individual’s day-to-day political talks is still very limited. While a number of studies have sought to reverse-engineer the machinery of censorship (Bamman et al., 2012; Fu et al., 2013; King et al., 2013, 2014; Zhu et al., 2013), these research projects mainly focus on the technical factors of the censorship system in order to unveil the government’s interest, intention, and goals (King et al., 2013). Relatively, little attention is paid to examine an individual’s reaction to censorship.

In this article, we aim to systematically investigate how censorship influences an individual’s opinion expression, while factoring in the impacts of social contagions and individual-level characteristics. We attempt to answer the following three questions. First, previous literature has suggested two opposite effects of censorship, namely chilling effect and backfire effect. To find out what gives rise to such contrasting observations, we make careful distinctions between four layers of censorship exposure: exposure to direct censorship on themselves, exposure to censorship on their friends, exposure to censorship on their reference persons, and exposure to censorship in the global environment. Based on this typology, we examine how censorship exposure in different layers influences people’s willingness to express opinions. Second, social media users are not isolated but rather networked. In light of this, we seek to explore if the exposure to outspoken public, friends, and reference persons can sway an individual’s propensity to voice his or her positions. Third, impacts of personal characteristics, such as activeness, occupation, and gender, will also be tested.

We opt for a case-based approach, studying the communication flow on Sina Weibo surrounding Hong Kong Anti-Extradition Law Amendment Bill Protest (Anti-ELAB protest) from 9 June to 31 December 2019. Anti-ELAB protest is an archetypal case of political controversy which involves sheer amounts of discussions and censorship on Chinese social media (Huang, 2019; Schiffer, 2019). Sina Weibo is chosen for three reasons. First, it is one of the most popular social media platforms in China (2018 Weibo Users Report, 2019). Second, it has profound impacts on public opinion formation, by staging citizen journalism (Bondes and Schucher, 2014; Sullivan, 2012), accommodating grassroot movements (Fu and Chau, 2014; Jiang, 2016b; Lim, 2018), and giving rise to a constellation of social media influencers. Third, it is public to all by default. Therefore, Weibo data can be more easily accessed, with relatively lower risk of violating user privacy.

The machinery of Internet censorship in China

The censorship system on Weibo is operating both before and after the publication. Before publication, the system checks on each submitted post and screens sensitive contents in three ways. First, it prohibits any post from explicitly mentioning blacklisted keywords. For example, if one submits a post containing “Liu Xiaobo (刘晓波)” or “Gui Minhai (桂民海),” regardless of intent, the user will be alerted instantly about the “violation of relevant laws and regulations” and the post cannot pass through the system. Second, the system holds up some ambiguous posts for manual review and informs the users that their posts will be released with a slight delay due to server synchronization, even though not all posts can be eventually published. According to the experiment conducted by King et al. (2014) on 100 websites including Weibo, 63% of the submissions that were held for human review never appeared on the sites. Third, the system disguises some posts as successfully published but others are not able to see them.

After publication, moderators scrutinize published posts and remove those they find objectionable. Research has unveiled that post-publication censorship is highly responsive and targets a very specific group of users. In terms of velocity, Zhu et al. (2013) and King et al. (2013) find the vast majority of post-publication censorship occurs within the first 24 hours after publication. Moreover, Zhu et al. (2013) also find that the more often one has been censored in the past, the more promptly one would be censored in the future. This indicates that the censorship might target a specific group of users and some vulnerable users are disproportionately subject to censorship. These vulnerable users, as found by Zhu et al. (2013), are characterized by showing more interest in issues related to state power. In a similar vein, Gallagher and Miller (2019) find that censorship apparatus selectively hampers the influence of opinion leaders while allowing small users greater leeway.

Besides, Bamman et al. (2012) unveil a list of sensitive keywords whose presence will lead to higher rate of post-publication deletion. Most sensitive keywords include Fang Binxing, Ministry of Truth, and Falun Gong. Fu et al. (2013) reveal another list of sensitive keywords, most of which are related to politically contentious events, such as the Bo Xilai scandal, human right lawyer event, and one-child policy. Instead of focusing on a single website, King et al. (2013, 2014) analyze censorship systems on a spectrum of websites. They find that the main goal of censorship is to curb collective action potential rather than to suppress the criticism of government.¹

However, one thing missing from this discussion is how censorship can be disclosed to the public. Although censorship is intended to go unnoticed, users can still observe censorship in various ways (Roberts, 2014, 2018). The first way is through the direct experience of censorship. When a post is censored, the author will receive a direct message from the platform informing him or her about the deletion. This is the primary form of censorship, which is very covert as the posts are removed without anyone other than the authors being notified. Nevertheless, if the posts have been reposted before being censored, the reposts may still exist after the original posts are gone. The area where the original posts once stood is hollowed out and replaced with a notice of censorship (Figure 1). We argue that despite not being deleted, such reposts are subjected to a secondary form of censorship, which is less covert and less destructive than primary censorship as people can still see “wrecks” of censored posts and can infer the intended meaning from such wrecks. People can come across secondary censorship anywhere when they are browsing posts by themselves, friends, reference persons, or anonymous others. These four targets of censorship constitute four layers of censorship exposure.

Figure 1.

Four layers of censorship exposure. Black arrows represent behavior flow from event to event. Red arrows represent social connections between individuals with arrow direction indicating the flow of influence from influencer to the influenced. Blue characters denote the names of associated variables.

To sum up, individual users can be exposed to censorship in four different layers. First layer is where users experience primary or secondary censorship on themselves. Second layer is where users are exposed to secondary censorship on their friends if their friends repost later-censored posts. The third layer is where users are exposed to secondary censorship on their reference persons, and the fourth one is where users are exposed to secondary censorship on anonymous public. People can run into censorship in any layer. In other words, censorship apparatus can exert its impacts on people in all four layers. The conceptual diagram of the four layers of censorship exposure is presented in Figure 1.

The impacts of censorship on the behaviors of opinion expressions

If censorship is observable, how do Chinese netizens react to it? In this section, we will review three different hypotheses about censorship effects—(a) chilling effect hypothesis, suggesting censorship will lead to a reduction in reading or posting messages in relation to the censored topics, (b) backfire effect hypothesis, suggesting censorship will lead to an increase in reading or posting related messages, and (c) minimal effect hypothesis, suggesting censorship will have no significant effect on people’s behavior.

The term “chilling effect” was initially coined in relation to the First Amendment of the US Constitution, describing a situation where individuals are deterred from exercising their right of free speech in presence of surveillance (Schauer, 1978; Tsui, 2003; Wacker, 2003). Many scholars have argued that after living with blatant censorship for decades, Chinese Internet users have well-taken into account the scrutiny of censors when choosing topics to read and talk (DeLisle et al., 2016; Lee, 2016; Stern and Hassid, 2012; Wacker, 2003). From the interviews with 22 Internet users, Jiang (2016a) finds that although the censorship machine is not perfect with many obvious loopholes, it has been very effective in making people aware that they are being watched, bringing out a chilling effect through self-regulation. Relatedly, Fu et al. (2013) find that the introduction of the real-name registration policy might have a chilling effect on some social media users.

The neuropsychological rationale behind the chilling effect can be understood in the light of behavioral inhibition theory (Fox et al., 2005; Gray, 1982; Yan et al., 2010). According to Gray (1982), there exist two motivational systems within the septo-hippocampal system, namely behavioral approach system (BAS) and behavioral inhibition system (BIS). BAS is responsible for processing cues associated with rewards and inducing active behaviors, while BIS is responsible for processing cues of punishments and inducing restrained behaviors. That is, when exposed to censorship, individuals might take it as a cue of punishment and inhibit their behavior of opinion expressions to avoid further punishment under the pull of BIS (Zeng et al., 2017).

In contrast to chilling effects, the minimal effect hypothesis assumes that censorship would only have marginal influence over people’s expression behavior. According to qualitative analyses by Jiang (2016a), resistant individuals who strongly disapprove of censorship are not influenced by censorship as they will anyway strive to evade censorship to express their opinions. Other studies suggest that minimal effects are not exclusive to hardcore individuals. Qin et al. (2017) find that individual citizens in China are not afraid of speaking on sensitive topics, even though there are documented cases of people being punished for doing so. Roberts (2014) analyzes the posts made by 516 Chinese bloggers and finds they do not change topics or attitudes after censorship. In a later work, Roberts (2018) finds from survey evidence that most Internet users do not report being fearful but instead being indifferent and angry after experiencing censorship.

Furthermore, a few notable studies have found censorship can backfire by angering the public and inspiring more interest in the concealed issues (Roberts, 2020). In the popular press, this backfire effect has been commonly known as “Streisand effect” (Hobbs and Roberts, 2018; Jansen and Martin, 2015). The theoretical foundation of backfire effect is based on an early psychological experiment² which unveiled censorship would increase an individual’s desire to consume more information on the censored speech (Worchel et al., 1975). The authors referred to psychological reactance theory (Brehm, 1966) to interpret such findings. They argue that the presence of censorship would constitute a threat to freedom and invoke psychological reactance to go against the will of censorship to restore their freedom.

In China’s context, from an experiment involving 150 Chinese university students, Roberts (2014) finds that Chinese censorship would trigger further interest in consuming censored topics. More broadly, Hobbs and Roberts (2018) find the sudden block of Instagram in China results in a surge in downloads of more political applications, such as the virtual private network, Twitter and Facebook. Therefore, they conclude the censorship would politicize apathetic citizens and trigger censorship backfire. However, existing evidence for backfire effects primarily lies in media consumption behaviors but not in media production behaviors.³ There is still no strong evidence to support the claim that censorship will provoke people to generate more posts on the censored topic.

Censorship will not lead to total compliance or uniform silence. Extant research has shown us a spectrum of censorship effects ranging from chilling effects to backfire effects. We assert such disparate observations could result from some confounding variables, such as how distant an individual is to actual censorship and how others are reacting to censorship.

Studies of public opinions have revealed that proximity is a critical predictor for public awareness and public attitude (Baxter and Lee, 2004; Wexler, 1996). Issues in proximity to individuals will elicit greater awareness and more concern than issues at distance. For example, “not in my backyard” oppositions which had been widely heard in grassroots movements demonstrate that many protesters rise up against the policies only if the policies are implemented in their area (Kraft and Clary, 1991; Wexler, 1996).

Following this line of thought, we hypothesize that whether people perceive censorship as a cue of punishment or as a threat to their freedom is associated with how distant they are to the censorship. Censorship on anonymous publics will more probably be regarded as a cue of punishment, which will activate the BIS and bring out a chilling effect. To its contrary, censorship on oneself poses a direct threat to one’s freedom and will activate their psychological reactance to go against the will of censorship, leading to a backfire effect. For the sake of clarity, we will consistently refer to the entire Weibo sphere consisting of anonymous publics as the global environment and refer to one’s social circle consisting of friends and reference persons as the local environment.

H1: Chilling Effect of Global Censorship. If more posts in the global environment are censored, individuals will be less likely to speak up on the censored issue.

H2: Backfire Effect of Direct Censorship. Individuals who have more posts being censored will be more likely to speak up on the censored issue.

RQ1: Effects of Local Censorship. How do individuals react if their friends’ or reference persons’ posts have been censored?

Behavioral contagion of opinion expression

Opinion expression can spread through social contacts just as many other behaviors do (Centola, 2010; Centola and Macy, 2007; Christakis and Fowler, 2007; Fowler and Christakis, 2008). When making decisions, individuals commonly and spontaneously consider other’s decisions, especially when they belong to the same group (Granovetter and Soong, 1988; Krassa, 1988; Oshagan, 1996).

In the context of opinion expression, whether an individual chooses to speak up or stay silent may hinge on how others have chosen before him (Granovetter and Soong, 1988). If only a few people have spoken of a certain issue, the individual will be very likely to fall silent as to avoid sanction or isolation. However, if the individual observes at least above minimum level of expression, he or she might feel safer to break the silence. Similar propositions have been raised by the Spiral of Silence (SOS) theory (Glynn et al., 1997; Neuwirth et al., 2007; Noelle-Neumann, 1974; Oshagan, 1996). However, SOS theory posits people will voice their opinions only if their opinion is congruent with the majority opinion, whereas behavioral contagion suggests the outspoken behavior itself is contagious regardless of the stance.

Consistent with previous observations, we postulate that outspoken people in the global environment will have positive influence over an individual’s opinion expression, and furthermore, will mitigate the chilling effects of global censorship.

H3: Global Contagion Effect. If there are more discussions about the censored issue in the global environment, individuals will be more likely to speak up on the issue.

H4: Global Contagion in Moderating Chilling Effect. If there are more discussions about the censored issue in the global environment, the chilling effect of global censorship will decrease.

Moreover, considerable research has found that the momentum of behavioral contagion is more significant among people who are closely connected through interpersonal contacts (Del Vicario et al., 2016; Glynn and Park, 1997; Krassa, 1988; Loersch et al., 2008; Oshagan, 1996; Scherer and Cho, 2003). As Scherer and Cho (2003) put it, the proximity of two actors in a social network is associated with the interpersonal influence between them. Harrigan et al. (2012) and Johnson Brown and Reingen (1987) have evidenced that strong ties are the main driver of information diffusion. People perceive those with whom they have maintained mutual bounds as more important and will be more likely to resend messages endorsed by them. Therefore, we posit that friends and reference persons are more influential than anonymous public in inspiring or curbing individuals’ opinion expressions. Besides, we expect the influence of friends will outweigh that of reference persons, as the friendship bonds are reciprocal and are stronger in emotional intensity, intimacy and trust than reference bonds.

H5: Local Contagion Effect. If there are more discussions about the censored issue in the local environment, individuals will be more likely to speak up on the issue.

H6: Friends versus Reference Persons. Friends’ expressions on the censored issue have stronger influence on individuals’ opinion expressions than reference persons’ expressions.

H7: Global Contagion versus Local Contagion. Discussions about the censored issue in the local environment have stronger influence on individuals’ opinion expressions than the discussions in the global environment.

Finally, when individuals’ posts are censored, their reaction would be tempered if their social contacts have been speaking up. To the contrary, if the person is experiencing censorship in isolation and very few of his friends or reference persons find the censored topic noteworthy, such isolation would probably enrage him, grow more radical thinking, and reinforce his motive to react against the will of censorship (Hug, 2013). Therefore, the last hypothesis summarizes this counteraction between backfire effect and local contagion in an individual’s social circle.

H8: Local Contagion in Moderating Backfire Effect. If there are more discussions about the censored issue in the local environment, the backfire effect of direct censorship will decrease.

Method

Research setting

To study the effects of censorship, Hong Kong Anti-ELAB protest is chosen as the research context. Anti-ELAB protest is a controversial political event that has been hotly discussed on Chinese social media and closely monitored by the authorities. It is deemed an appropriate case for this study.

In 2019, the Hong Kong government put forward an amendment to the Fugitive Offenders Ordinance (Cap. 503), in which a new “special surrender arrangement” between Hong Kong and mainland China was introduced. Anti-ELAB protest emerged in response to the legislation. It gained crucial momentum on 9 June 2019, where an estimated 1 million people, or one in seven of the city’s population, took to the streets (Pomfret and Master, 2019). In the following months, the protest has morphed into a pro-democracy movement demanding a broad range of changes, such as institutional reforms, investigation into alleged police brutality, and universal suffrage (Choi, 2020), until the coronavirus broke out in early 2020.

However, Chinese central government has been strongly supporting the bill and denouncing the protest as a malicious maneuver by “foreign forces” to meddle in China’s internal affairs (Chinese Foreign Ministry, 2019). State media initially responded with silence until 21 July when protesters defaced the Chinese national emblem in Hong Kong Liaison Office. Thereafter, a large amount of news articles came flooding in to condemn the protest.

Whereas, on social media, according to King et al. (2013, 2014), collective actions are always ranked as the most censorable topic in China, facing more stringent censorship than other topics. Many social media users have reported their public utterance about Anti-ELAB protest being censored (Myers and Mozur, 2019; Schiffer, 2019). Although both pro-protest and anti-protest sides were subjected to censorship, the pro-protest users were allegedly faced with disproportionately harsher censorship (Gan, 2019; Myers and Mozur, 2019).

Data

The primary dataset used in this study was collected by Weiboscope (Human Research Ethics Committee Reference number: EA260113) (Fu and Chau, 2014; Fu et al., 2013). To obtain posts related to the Anti-ELAB protest, we searched for a set of fine-tuned keywords (Supplemental materials, Note S3) between 9 June 2019 and 31 December 2019 in Weiboscope. Then, we thoroughly checked the posts and removed irrelevant ones with the help of a computer-assisted text analysis tool (see Supplemental materials, Note S4). After this, 237,818 posts were retained. We expanded the data set to include the reposts and the original posts of these 237,818 posts. Finally, we collected overall 460,731 protest-related posts from 16,520 users.

Social network

To construct the online social networks among individuals, repost relations of involved discussants (N = 16,520) during the first 6 months of 2019, a separate time period from the Anti-ELAB protest, were downloaded from Weiboscope. Repost relations are the skeletons of online social network, which are directed from originator to reposters, with directions indicating the flow of information and influence. Based on the directions, we distinguished three mutually exclusive types of social contacts: friends, reference persons, and followers. Friends of the individual A are persons who are mutually and closely connected with A. The interactions between friends are reciprocal. Therefore, we identified those who have reposted and been reposted by A as A’s friends. Meanwhile, reference persons are people to whom A refers to evaluate and adjust their performance.⁴ Information only flows from reference persons to A, not vice versa. Therefore, we identified those who have been reposted by A but have never reposted A as A’s reference persons. For simplicity, we will use the abbreviated version “reference” to denote “reference person.” Finally, followers are people who take A as their reference. So, we identified those who have reposted but never been reposted by A as A’s followers. Since individuals are not exposed to information posted by their followers, we assume followers have little influence over one’s willingness of expression.

With this procedure, we found additional 8789 users who had been exposed to at least one protest-related message yet fell silent in the study period. Combining the outspoken and silent users, a total of 25,309 users constituted the sample of this study, who were tracked on a weekly basis over the 31-week period, yielding a data set of 784,579 observations (25,309 users × 31 weeks).⁵

Opinion expression

The outcome variable Y_ij is a count variable, whose magnitude is equal to the number of protest-related posts created by individual i in week j. Its dichotomous version is y_ij, which is equal to 1 if individual i speaks up in week j and equal to 0 if not. Besides, we differentiate two types of non-expression behaviors: silence and avoidance. Silence refers to the condition where an individual says nothing, and avoidance refers to the situation where an individual says something not related to the protest. To operationalize them, we obtained the weekly post frequency of each user, say N_ij, and labeled those with (N_ij = 0 and Y_ij = 0) as silent cases and those with (N_ij > 0 and Y_ij = 0) as avoidant cases.

As for predictors, we measured the magnitudes of behavioral contagions in global and local environments. In terms of global contagion (H3 and H4), the variable V_j denotes the overall number of protest-related posts created in week j. And for local contagion (H5 and H6), we measured three indicators: the number of posts created by one’s references/friends, the number of references/friends who speak up, and the percentage of references/friends who speaks up. Factor analysis showed that these three indicators were unidimensional for references and friends, respectively (Supplemental materials, Table S1). Therefore, we extracted the two principal components and named them as VRef_ij and VFr_ij. VRef_ij measures the prevalence of expression among the references of individual i in week j and VFr_ij measures that among the friends of individual i in week j.

Censorship: global and local

Out of 460,731 protest-related posts, 19,961 were censored by the platform (Supplemental materials, Note S5). Global censorship C_j is measured by the total count of censored posts in week j (H1 and H4). CD_ij gauges the total harshness of primary and secondary censorship on individual i in week j, which is sum of censored posts and uncensored reposts whose original posts are censored (H2 and H7). However, censorship will only fall on outspoken people. When CD_ij is 1, y_ij can only be 1. These two variables are partially redundant and using CD_ij to predict y_ij is tautological. To avoid this, we will use CD_i(j–1) instead of CD_ij as the key predictor for Y_ij, which assumes the past experience of direct censorship can influence an individual’s present behavior of opinion expression. CRef_ij denotes the secondary censorship on individual i’s references in week j, which is measured by the number of references’ reposts whose original posts were censored. In the same vein, CFr_ij denotes the secondary censorship imposed on individual i’s friends in week j. All these four variables are positively skewed. Therefore, logarithmic transformation was applied to them to make them conform more closely to normality. Since the log-transformation can only be applied to positive value, we added constant 1 to all observations to escape the zero. Compared with raw numbers, the log-transformed indicators were found to contribute more to the model’s statistical power in terms of R-squared statistics.

Control variables

Previous research has suggested individual factors such as gender (Noelle-Neumann, 1974), user influence (Jiang, 2016b; Roberts, 2018), and indigenous ideological positions (Jiang, 2016a; Lee, 2016; Stoycheff, 2016) can sway an individual’s propensity to speak out. To control for these confounding variables, we used Sina’s Open application programming interface (API) to retrieve the profile information of all 25,309 users and quantified them as follows.

Gender, gender_i, is a dummy variable equal to 1 when the individual i is male.

Activity level, level_i, is an ordinal variable, indicating official evaluation of user i’s level of activeness from the least active 0 to extremely active 48. According to Sina Weibo, this indicator is calculated on the basis of three criteria: the number of days the user has been active on Weibo, the number of posts the user has generated, and the number of friends whom the user has engaged with.

User type, type_ki, is a list of seven dummies, equal to 1 when user i has been verified as type k. Verified accounts are required to prove their identities as anyone of nine categories, namely celebrity, government, corporation, corporation under review, media, education, website, organization, and application. We dropped two infrequent categories, namely application and corporation under review, to avoid data sparsity problems, and converted the categorical code into seven dummy variables (when all seven dummies equal to 0, the user is not verified). It is worth-noting that despite sharing many commonalities, the nature of accounts in the same category still varies widely from one to another. For example, government accounts cover all kinds of public institutions from county police force to State Council and from British Embassy to the United Nation.

Besides, we also controlled for user’s ideological position by deploying a pre-trained ideal point estimator (Zhu and Fu, 2020) which maps users onto an eight-dimensional ideological space. Given the longitudinal nature of the data, we introduced three lag terms of dependent variables to alleviate their autocorrelations as well as to control for their prior levels before assessing the synchronous influences of independent variables (Burke and Kraut, 2016; Eveland and Thomson, 2006; Finkel, 1995; Fu, 2012).

Data analysis

The opinion expression is formulated as a sequential decision-making process, where individuals first decide if they want to say something on the issue and if yes, then they decide on how many times they want to speak on it in a fixed interval. Therefore, we implement a two-part model with two separate components, one determining whether an individual speaks up and one predicting the number of posts an individual will generate. Such two-part models, also known as hurdle models, are frequently adopted for modeling count data with excess zeros (Cameron and Trivedi, 2013; Fu and Chau, 2014). In this study, the binary outcome of whether an individual speaks up was analyzed by two logistic regressions. The first logistic regression (Model 1 in Table 1) analyzed the odds of opinion expression against both types of non-expression, namely silence and avoidance, whereas the second logistic regression (Model 2) analyzed the odds of expression against avoidance. If the individual was observed to speak up, his or her expression intensity Y_ij was analyzed by a Poisson regression (Model 3). To make the three models comparable and commensurable, the same set of predictor variables were used.

Table 1.

Regression models predicting opinion expression.

Odds/Regressand	Model 1 (Logit)		Model 2 (Logit)		Model 3 (Poisson)
Regressors	$\frac{P (e x p r e s s i o n)}{P (s i l e n c e o r a v o i d a n c e)}$		$\frac{P (e x p r e s s i o n)}{P (a v o i d a n c e)}$		#expression (where #expression > 0)
	ß	Wald	ß	Wald	ß	Wald
Censorship effects
C_j (overall censorship)	–.31*	6.06	–.56***	17.19	.37***	29.05
CD_i_(j−1) (direct censorship last week)	2.35***	688.30	2.26***	570.25	1.89***	94064.07
CRef_ij (censorship on references)	.90***	759.36	.93***	769.50	.19***	1452.30
CFr_ij (censorship on friends)	.23**	8.71	.41***	26.84	.35***	1621.80
Contagion effects
V_j (overall discussion)	1.25***	231.10	1.34***	234.76	1.80***	1311.70
VRef_ij (references’ expression)	.16***	1187.00	.13***	730.39	.13***	5192.52
VFr_ij (friends’ expression)	.26***	4346.94	.22***	3060.26	.15***	23719.44
Interaction
C_j × V_j	.02	.30	.04	1.67	–.20***	142.42
CD_i_(j–1) × VRef_ij	–.36***	35.87	–.32***	25.89	–.05***	338.22
CD_i_(j–1) × VFr_ij	–.13**	10.20	–.03	.41	–.19***	14663.48
Control terms
gender_i (male = 1)	.29***	770.23	.38***	1242.90	.22***	2814.93
level_i	.02***	834.62	.00***	56.89	.00	.47
type₁ celebrity	–.09***	60.02	–.20***	303.16	–.11***	765.02
type₂ gov	–.07**	9.32	–.15***	42.04	–.07***	44.55
type₃ corporation	–.77***	519.75	–.91***	698.92	–.55***	615.42
type₄ media	.47***	769.97	.22***	166.17	.50***	8769.26
type₅ education	–.15^*	4.19	–.35***	23.02	–.77***	241.55
type₆ website	.22**	7.71	–.06	.61	.65***	1009.67
type₇ organization	–.75***	132.26	–.94***	201.02	–.87***	263.64
ideological position	Yes	–	Yes	–	Yes	–
1st, 2nd and 3rd lags	Yes	–	Yes	–	Yes	–
N	784,579		526,016		79,967
Pseudo R²	.40		.39		.41

p < .05; **p < .01; ***p < .001.

Results

Table 1 presents the results of three regressions. Since the main objective of this study is to understand why people will or will not speak up on the censored issue, we will put more emphasis on Models 1 and 2.

Evaluation of H1 (chilling effect of global censorship)

The C_j term has significantly negative coefficients in Model 1 and Model 2, with the odds ratios at 0.73 and 0.57 respectively. It means that if C_j term increases by one unit (the global censorship increases to 10 times its value plus nine), the odds of expression to non-expression will be reduced to 73%, while the odds of expression to avoidance will be reduced more prominently to 57%. Therefore, increasing censorship in the global environment can significantly decrease an individual’s propensity to speak out and its influence is found to be greater in inducing people to avoid rather than to stay silent.

Nevertheless, despite elevating the hurdle for speaking up, global censorship would backfire among those who have already crossed the hurdle. As indicated by the significantly positive coefficient of the C_j term in Model 3, once an individual has decided to break the silence, harsher global censorship will lead them to post more contents on the censored issue. Yet, given that only 10% of the studied cases crossed the hurdle, the backfire effect is overshadowed by the chilling effect. Therefore, the chilling effect assumption about global censorship is still largely supported.

Evaluation of H2 (backfire effect of direct censorship)

The analyses also support H2. The term CD_i_(j−1) demonstrates significantly positive relationships with the odds of expression to non-expression (β = 2.35, p < .001), the odds of expression to avoidance (β = 2.26, p < .001) and also the expression intensity (β = 1.89, p < .001). The direct experience of censorship will motivate individuals to speak on the controversial topic to a greater extent. Besides, the first coefficient is greater than the second. It means if an individual experiences direct censorship, his or her chance of being totally silent will be reduced more prominently than the chance of avoidance. After being censored, people will have a strong impulse to say something, no matter what it is.

Evaluation of RQ1 (censorship on references and friends)

RQ1 concerns how users react to the censorship on their references and friends. As indicated by the positive coefficients of CRef_ij and CFri_ij, when people observe their friends or references being censored, they will be more emboldened to speak up. In other words, censorship will trigger backfire and encourage more expressions as long as it enters one’s social circle. Moreover, the relative effect sizes of CRef_ij, as measured by the Wald statistics, are greater than CFri_ij and CD_i(j−1) in Models 1 and 2 but not in Model 3. That is, the censorship on references can induce people to break the silence, while the censorship on oneself and on one’s friends can increase the outspoken people’s expression intensity.

Evaluation of H3 and H4 (global contagion effect)

Since the coefficients of V_j in all three models are significantly positive, H3 is substantiated. Consistent with conventional wisdom about crowd psychology, we find the more anonymous people speak up in the global environment, the more likely the individual is to speak up. Furthermore, global contagion can diminish the chilling effects of censorship as indicated by the positive coefficients of the C_j × V_j terms in Models 1 and 2. However, the coefficients are not statistically significant. So, we do not have sufficient evidence to accept H4.

Evaluation of H5–H8 (local contagion effect)

We expect the local contagions among friends and reference persons have positive influences over an individual’s opinion expressions (H5) and the influence of friends would presumably outweigh that of reference persons (H6), while both will outweigh that of anonymous public (H7). Table 1 confirms that both VRef_ij and VFr_ij have positive relationships with opinion expression in all three models and their Wald statistics are greater than that of V_j, in support of H5 and H7. Also, the standardized effect sizes of VFr_ij, as indicated by the Wald statistics, are greater than that of VRef_ij, which means one standard deviation increase in the friends’ expressions can more prominently boost people’s tendency to speak out than the one standard deviation increase in their reference persons’ expression. H6 is therefore supported.

In terms of moderation effects, the coefficients of CD_i_(j–1) × VRef_ij and CD_i_(j–1) × VFr_ij are consistently negative in all models. Given that the coefficients of CD_i_(j–1) are positive, these two interaction terms will counteract the backfire effect caused by CD_i_(j–1). When local contagion gets larger, the backfire effects caused by direct censorship will decrease. However, the coefficient of CD_i_(j–1) × VFr_ij in Model 2 is not statistically significant and H8 is therefore only partially confirmed.

Generally speaking, the Wald statistics of contagion-related main effects are larger than censorship-related terms within each model, implying that the behavioral contagion is more decisive than censorship in determining opinion expression.

Evaluation of control variables

The regression coefficients of all control variables demonstrated consistent relationships with dependent variables across two models. In line with previous studies (Noelle-Neumann, 1974; Wells et al., 2017), male, highly active, media, and website accounts are found to be more outspoken on the sensitive issue. To the contrary, female, less active users, and all verified types except media and websites are more restrained in disclosing their opinions.

Notably, the media, bearing the responsibility as the watchdog to inspect anomalies in the society as well as being more knowledgeable to comment on sensitive topics, is the only type of verified accounts that is neither silent nor avoidant but significantly straightforward in discussing this sensitive topic. Except for the media, high-profile verified users have higher propensity to stay silent than non-verified users. This might result from the selective censorship strategy that disproportionately crackdown on elite users on Weibo (Gallagher and Miller, 2019; Zeng et al., 2017). Or, it might generally reflect the “insufficient hub” hypothesis which posits high-profile users are overloaded with inputs and are less responsive to viral messages and socially contagious transmission is mainly driven by less overloaded, more average, individuals (Barabási, 2002; Harrigan et al., 2012).

Discussion and conclusion

This study draws on a large-scale social media dataset to systematically investigate people’s reaction to censorship. To overcome the technocentric and decontextualized problems (Jiang and Fu, 2018), we endeavor to preserve the social context by accounting for the influence of social contacts, overall climate of outspokenness, and individual characteristics. The main findings are fourfold.

First, in contrast with previous research which assumes censorship’s influence is largely monolithic, either solely encouraging or solely discouraging opinion expression (Hobbs and Roberts, 2018; Roberts, 2014, 2018), we argue that the impact of censorship is actually multifaceted and it can bring out different reactions in individuals through four layers of exposure. Censorship at distance is more likely to cause chilling effects, while the censorship in “one’s backyard” is more likely to backfire. People tend to stay silent when the global censorship is intensive. But when they experience censorship themselves or witness censorship really occurring to their friends or reference persons, they are inclined to “rebel” against censorship by voicing their opinions. The chilling effect is weaker, while the backfire effect is more limited in scope as it is concentrated on people who have come across censorship in their social circles.

In light of behavioral inhibition theory (Yan et al., 2010) and psychological reactance theory (Brehm, 1966; Worchel et al., 1975), such findings suggest that censorship at distance is more likely to be perceived as a cue for punishment, which can invoke fear and induce the inhibition of expression, while censorship close to individuals is more likely to be perceived as a threat to freedom, which can provoke people to restore freedom by intentionally reacting against censorship.

This theoretical framework provides a possible integration of seemingly mixed results derived from previous studies and offers a coherent model to inform future research. For example, as noted by Jiang (2016a), two interviewees felt hesitant about publicly discussing sensitive topics and attributed their reluctance to the mistreatment of innocent people who wanted to speak up. Whereas, in Roberts’ (2014) experiment and observational study (Roberts, 2018), censorship directly imposed on the Internet users can backfire, provoking angry backlash against the censors and censored topics. The first case is consistent with our finding of the chilling effects of censorship at distance, while the second is congruent with the backfire effects of direct censorship.

In addition, we find the behavioral contagion has more profound influence than censorship exposure. People tend to follow in other’s footsteps to speak up, even in face of censorship. This indicates that people fear social isolation more than the punishment of censorship.

Furthermore, behavioral contagion can alleviate the impacts of censorship. Global contagion can feebly counter the chilling effects of global censorship and local contagion can significantly relieve the backfire effects of direct censorship. In other words, the outspoken crowd could shield individuals from the fear of punishment and the outspoken friends and reference persons could soften individuals’ anger against direct censorship. In either case, the behavioral contagion can counteract the influences of censorship, liberate individuals from the overconcern with the censorship and empower them to act for themselves.

Finally, all types of verified users, except for media and websites, have a stronger tendency than non-verified users to stay silent or avoidant. So, even though elite users have a large number of followers and their followers are very defensive about them, elite users’ lower propensity to speak out prevents such contagion effects and backfire effects from rippling through the network. Note that, despite earlier reports that Weibo imposed selective censorship strategy to disproportionately crack down on elite users (Gallagher and Miller, 2019), we find in this study censorship “hits” ordinary users more frequently than elite users (Supplemental materials, Table S2). Nevertheless, in an absence of immediate threat of punishment, skittish elites still remain restrained, whose self-censorship undermines the resilience to censorship.

To summarize, despite being increasingly potent, censorship is not always effective in stopping people from speaking up. Instead, it is a double-edged sword, which can either silence people or backfire. Meanwhile, community acts as a critical buffer against the stress of being watched. If opinion expression has been popularized in the community either globally or locally, people are less concerned about censorship and more inclined to speak out. To its contrary, if opinion expression is rare anywhere globally and locally, it adds on extra concerns about censorship and exaggerates both chilling effects and backfire effects. Therefore, it is reasonable to infer that the timing of censorship is more important than its scale. Censoring an already popular topic will not curb the diffusion of information and will set the censors on fire by prompting backlashes from the outraged users. Whereas, censoring a nascent and undermentioned topic will be more detrimental as it can curb the information diffusion without risk of backfire. This also implies that whistle-blower or early-discussant hunting will be more destructive than the mass purge for the development of public discussion on Weibo.

Some limitations are worth noting. The findings are based on an analysis of the Weibo discussions on the 2019 Hong Kong social movement. Even though it appears to be a relevant case for censorship research, further study is warranted to test hypotheses in other settings. The user samples of Weiboscope are built by a combination of self-selected high profile and randomly selected users. Although they can be considered fairly representative, they may still deviate from the whole user population of Weibo. Finally, this study does not distinguish users’ political stances. Results discussed earlier demonstrate a common pattern shared by different opinion groups, which means, just like their pro-protest counterparts, anti-protest users will also be silenced by distant censorship and will be angered by proximal censorship. Indeed, we do find some anti-protest influencers backlashed censors and belittled Sina Weibo as traitorous platform after their posts being censored. However, political stance might sway people’s reaction in some other respects. For example, according to the SOS theory, if people find themselves in alignment with the majority opinion, they will be more willing to disclose their stances (Neuwirth et al., 2007; Noelle-Neumann, 1974). Future research might investigate whether censorship still triggers backfire when the individuals disagree with the censored opinions, and whether the individuals are still inclined to follow their friends and references to speak up when the individuals disagree with the majority opinion expressed by their friends and references.

Supplemental Material

NMS_Supplemental_Materials_0717 – Supplemental material for Speaking up or staying silent? Examining the influences of censorship and behavioral contagion on opinion (non-)expression in China

Supplemental material, NMS_Supplemental_Materials_0717 for Speaking up or staying silent? Examining the influences of censorship and behavioral contagion on opinion (non-)expression in China by Yuner Zhu and King-wa Fu in New Media & Society

Footnotes

Authors’ note

All authors have agreed to the submission. The article is not currently being considered for publication by any other journal. Correspondence concerning this article should be addressed to King-wa Fu.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Yuner Zhu

King-wa Fu

Supplemental material

Supplemental material for this article is available online.

Notes

Author biographies

Yuner Zhu is currently a PhD candidate at the Journalism and Media Studies Centre, The University of Hong Kong. Her research interests primarily lie in public opinion formation, political opinion expression, computational social sciences, social media use, and media framing. Her online CV: .

King-wa Fu is an associate professor at the Journalism and Media Studies Centre, The University of Hong Kong. His research interests cover information control policy in China, political participation and media use, computational social sciences, health and the media, and younger generation’s Internet use. He was a visiting associate professor at the MIT Media Lab and Fulbright-RGC Hong Kong senior research scholar in 2016–2017. His online CV: .

References

Bamman

O’Connor

Smith

(2012) Censorship and deletion practices in Chinese social media. First Monday 17(3): 234.

Barabási

(2002) Linked: The New Science of Networks. New York: Plume.

Baxter

Lee

(2004) Understanding expressed low concern and latent concern near a hazardous waste treatment facility. Journal of Risk Research 7(7–8): 705–729.

Bondes

Schucher

(2014) Derailed emotions: the transformation of claims and targets during the Wenzhou online incident. Information Communication and Society 17(1): 45–65.

Brehm

(1966) A Theory of Psychological Reactance. New York, NY: Academic Press.

Brown

Reingen

(1987) Social ties and word-of-mouth referral behavior. Journal of Consumer Research 14(3): 350–362.

Burke

Kraut

(2016) The relationship between Facebook use and well-being depends on communication type and tie strength. Journal of Computer-Mediated Communication 21(4): 265–281.

Cameron

Trivedi

(2013) Regression Analysis of Count Data, vol. 53. Cambridge: Cambridge University Press.

Carver

Humphries

(1981) Havana daydreaming: a study of self-consciousness and the negative reference group among Cuban Americans. Journal of Personality and Social Psychology 40(3): 545–552.

10.

Centola

(2010) The spread of behavior in an online social network experiment. Science 329(5996): 1194–1197.

11.

Centola

Macy

(2007) Complex contagions and the weakness of long ties. American Journal of Sociology 113(3): 702–734.

12.

Chen

(2012) The impact of ignoring a level of nesting structure in multilevel mixture model: a Monte Carlo study. SAGE Open 2(1): 1–10.

13.

Chinese Foreign Ministry (2019) Chinese foreign ministry spokenman Geng Shuang hosts daily conference, 10 June. Beijing, China. Available at: https://www.fmprc.gov.cn/web/wjdt_674879/fyrbt_674889/t1670837.shtml

14.

Choi

SYP

(2020) When protests and daily life converge: the spaces and people of Hong Kong’s anti-extradition movement. Critique of Anthropology 40(2): 277–282.

15.

Christakis

Fowler

(2007) The spread of obesity in a large social network over 32 years. New England Journal of Medicine 357(4): 370–379.

16.

CNNIC (2019) Forty-Third Statistical Report on Internet Development in China. Beijing, China: CNNIC.

17.

Creemers

(2017) Cyber China: upgrading propaganda, public opinion work and social management for the twenty-first century. Journal of Contemporary China 26(103): 85–100.

18.

DeLisle

Goldstein

Yang

(2016) The internet, social media, and a changing China. In: DeLisle

Goldstein

Yang

(eds) The Internet, Social Media, and a Changing China. Pennsylvania, PA: University of Pennsylvania Press, pp. 1–27.

19.

Del Vicario

Vivaldo

Bessi

, et al. (2016) Echo chambers: emotional contagion and group polarization on Facebook. Scientific Reports 6: 37825.

20.

Endeshaw

(2004) Internet regulation in China: the never-ending cat and mouse game. Information and Communications Technology Law 13(1): 41–57.

21.

Eveland

Thomson

(2006) Is it talking, thinking, or both? A lagged dependent variable model of discussion effects on political knowledge. Journal of Communication 56(3): 523–542.

22.

Finkel

(1995) Causal Analysis with Panel Data (Sage University Paper Series on Quantitative Applications in the Social Sciences). Thousand Oaks, CA: SAGE.

23.

Fowler

Christakis

(2008) Dynamic spread of happiness in a large social network: longitudinal analysis over 20 years in the Framingham Heart Study. BMJ 337: a2338.

24.

Fox

Henderson

Marshall

, et al. (2005) Behavioral inhibition: linking biology and behavior within a developmental framework. Annual Review of Psychology 56(1): 235–262.

25.

Chau

(2014) Use of microblogs in grassroots movements in China: exploring the role of online networking in agenda setting. Journal of Information Technology and Politics 11(3): 309–328.

26.

Chan

Chau

(2013) Assessing censorship on microblogs in China. IEEE Internet Computing 17(3): 2–10.

27.

(2012) Selecting online videos from graphics, text, and view counts : the moderation of popularity bandwagons. Journal of Computer-Mediated Communication 18(1): 46–61.

28.

Gallagher

Miller

(2019) Who not what: the logic of China’s information control strategy. Working Paper. Available at: http://blakeapm.com/

29.

Gan

(2019) Chinese lawyer Chen Qiushi, censured over Hong Kong social media posts, vows to keep speaking out. South China Morning Post, 16 October. Available at: https://www.scmp.com/news/china/politics/article/3033215/chinese-lawyer-chen-qiushi-censured-over-hong-kong-social-media

30.

Gleiss

(2015) Speaking up for the suffering (br)other: Weibo activism, discursive struggles, and minimal politics in China. Media, Culture and Society 37(4): 513–529.

31.

Glynn

Park

(1997) Reference groups, opinion intensity, and public opinion expression. International Journal of Public Opinion Research 9(3): 213–232.

32.

Glynn

Hayes

Shanahan

(1997) Perceived support for one’s opinions and willingness to speak out: a meta-analysis of survey studies on the “spiral of silence.” Public Opinion Quarterly 61(3): 452–463.

33.

Granovetter

Soong

(1988) Threshold models of diversity: Chinese restaurants, residential segregation, and the spiral of silence. Sociological Methodology 18: 69–104.

34.

Gray

(1982) Précis of the neuropsychology of anxiety: an enquiry into the functions of the septo-hippocampal system. Behavioral and Brain Sciences 5(3): 469–484.

35.

Harrigan

Achananuparp

Lim

(2012) Influentials, novelty, and social contagion. Social Networks 34(4): 470–480.

36.

Hobbs

Roberts

(2018) How sudden censorship can increase access to information. American Political Science Review 112(3): 621–636.

37.

Huang

(2019) After a Chinese flag was thrown into the sea, Hong Kong celebrities flex their patriotic muscles. Quartz, 8 August. Available at: https://qz.com/1682950/hong-kong-celebrities-vow-to-protect-chinese-flag-on-weibo

38.

Hug

(2013) The role of isolation in radicalization: how important is it? Master’s Thesis, Naval Postgraduate School, Monterey, CA.

39.

Jansen

Martin

(2015) The Streisand effect and censorship backfire. International Journal of Communication 9: 656–671.

40.

Jiang

(2016a) Managing the micro-self: the governmentality of real name registration policy in Chinese microblogosphere. Information Communication and Society 19(2): 203–220.

41.

Jiang

(2016b) The coevolution of the internet, (un)civil society, and authoritarianism in China. In: DeLisle

Goldstein

Yang

(eds) The Internet, Social Media, and a Changing China. Pennsylvania, PA: University of Pennsylvania Press, pp. 28–48.

42.

Jiang

(2018) Chinese social media and big data: big data, big brother, big profit? Policy and Internet 10(4): 372–392.

43.

Kang

Zhang

Wang

(2015, June) Analysis of sentiment transition of retweeting messages on Sina Weibo. In: Proceedings of the 2014 international conference on management, information and educational engineering (MIEE 2014), Xiamen, China, 22–23 November 2014, p. 105. CRC Press.

44.

King

Pan

Roberts

(2013) How censorship in China allows government criticism but silences collective expression. American Political Science Review 107(2): 326–343.

45.

King

Pan

Roberts

(2014) Reverse-engineering censorship in China: randomized experimentation and participant observation. Science 345(6199): 1251722.

46.

Kitchin

(2013) Big data and human geography: opportunities, challenges and risks. Dialogues in Human Geography 3(3): 262–267.

47.

Kraft

Clary

(1991) Citizen participation and the NIMBY syndrome: public response to radioactive waste disposal. Western Political Quarterly 44(2): 299–328.

48.

Krassa

(1988) Social groups, selective perception, and behavioral contagion in public opinion. Social Networks 10(2): 109–136.

49.

Lee

(2016) Surviving online censorship in China: three satirical tactics and their impact. China Quarterly 228: 1061–1080.

50.

Lim

(2018) China’s #MeToo censorship bypassed through netizens’ creative use of language. South China Morning Post, 2 March. Available at: https://www.scmp.com/magazines/post-magazine/short-reads/article/2134847/chinas-metoo-censorship-bypassed-through

51.

Loersch

Aarts

Payne

, et al. (2008) The influence of social groups on goal contagion. Journal of Experimental Social Psychology 44(6): 1555–1558.

52.

Myers

Mozur

(2019) China is waging a disinformation war against Hong Kong protesters. New York Times, 13 August. Available at: https://www.nytimes.com/2019/08/13/world/asia/hong-kong-protests-china.html

53.

Neuwirth

Frederick

Mayo

(2007) The spiral of silence and fear of isolation. Journal of Communication 57(3): 450–468.

54.

Newcomb

(1950) Role behaviors in the study of individual personality and of groups. Journal of Personality 18: 273–289.

55.

Noelle-Neumann

(1974) The spiral of silence: a theory of public opinion. Journal of Communication 24(2): 43–51.

56.

Oshagan

(1996) Reference group influence on opinion expression. International Journal of Public Opinion Research 8(4): 335–354.

57.

People.cn (2018) 2018 annual report of People.cn. Available at: http://vip.stock.finance.sina.com.cn/corp/view/vCB_AllBulletinDetail.php?stockid=603000&id=5202592

58.

Pomfret

Master

(2019) Hong Kong pushes bill allowing extraditions to China despite biggest protest since handover. Reuters, 10 June. Available at: https://www.reuters.com/article/us-hongkong-extradition-idUSKCN1TB08W

59.

Qin

Strömberg

(2017) Why does China allow freer social media? Protests versus surveillance and propaganda. Journal of Economic Perspectives 31(1): 117–140.

60.

Roberts

(2014) Fear or friction? How censorship slows the spread of information in the digital age. Working paper, pp. 1–34. Available at: https://pdfs.semanticscholar.org/f10f/d77e13c31c75ba4c4290ad89582f04aab266.pdf

61.

Roberts

(2018) Censored: Distraction and Diversion Inside China’s Great Firewall. Princeton, NJ: Princeton University Press.

62.

Roberts

(2020) Resilience to online censorship. Annual Review of Political Science 23: 22.1–22.19.

63.

Schauer

(1978) Fear, risk and the first amendment: unraveling the chilling effect. Boston University Law Review 58(5): 685–732.

64.

Scherer

Cho

(2003) A social network contagion theory of risk perception. Risk Analysis 23(2): 261–267.

65.

Schiffer

(2019) WeChat keeps banning Chinese Americans for talking about Hong Kong. The Verge, 25 November. Available at: https://www.theverge.com/2019/11/25/20976964/chinese-americans-censorship-wechat-hong-kong-elections-tiktok

66.

Stern

Hassid

(2012) Amplifying silence: uncertainty and control parables in contemporary China. Comparative Political Studies 45(10): 1230–1254.

67.

Stockmann

(2018) Toward area-smart data science: critical questions for working with big data from China. Policy and Internet 10(4): 393–414.

68.

Stoycheff

(2016) Under surveillance: examining Facebook’s spiral of silence effects in the wake of NSA internet monitoring. Journalism & Mass Communication Quarterly 93(2): 296–311.

69.

Sullivan

(2012) A tale of two microblogs in China. New Media & Society 34(6): 773–783.

70.

Sullivan

(2014) China’s Weibo: is faster different? New Media & Society 16(1): 24–37.

71.

Tsui

(2003) The panopticon as the antithesis of a space of freedom: control and regulation of the internet in China. China Information 17(2): 65–82.

72.

2018 Weibo users report (2019) Available at: http://data.weibo.com/report/reportDetail?id=433

73.

Wacker

(2003) The internet and censorship in China. In: Hughes

Wacker

(eds) China and the Internet: Politics of the Digital Leap Forward. London: Routledge, pp. 58–82.

74.

Wang

Jin

Yang

, et al. (2017) An emotion-based independent cascade model for sentiment spreading. Knowledge-Based Systems 116: 86–93.

75.

Wells

Cramer

Wagner

, et al. (2017) When we stop talking politics: the maintenance and closing of conversation in contentious times. Journal of Communication 67(1): 131–157.

76.

Wexler

(1996) A sociological framing of the NIMBY (not-in-my-backyard) syndrome. International Review of Modern Sociology 26: 91–110.

77.

Worchel

Baker

Carolina

(1975) The effects of censorship on attitude change: the influence of censor and communication characteristics. Journal of Applied Social Psychology 5(3): 227–239.

78.

Yan

Dillard

Shen

(2010) The effects of mood, message framing, and behavioral advocacy on persuasion. Journal of Communication 60(2): 344–363.

79.

Yang

(2004) The co-evolution of the internet and civil society in China. Asian Survey 43(3): 405–422.

80.

(2019) 主流媒体引导力, 可否这样实现 [Can mainstream media achieve the power of orientation in this way?] People.cn, 11 February. Available at: http://media.people.com.cn/n1/2019/0211/c192370-30617943.html

81.

Zeng

Chan

(2017) How social media construct “truth” around crisis events: Weibo’s rumor management strategies after the 2015 Tianjin blasts. Policy and Internet 9(3): 297–320.

82.

Zhu

Phipps

Pridgen

, et al. (2013) The velocity of censorship: high-fidelity detection of microblog post deletions. In: Proceedings of the 22nd USENIX security symposium, Washington, DC, 14–16 August, pp. 227–240. Berkeley, CA: USENIX Association.

83.

Zhu

(2020) Mapping ideological landscape of social media: a neural network approach. In: ICA (International Communication Association), 69th annual ICA conference. Virtual conference 20–27 May 2020. Washington, DC, United States. Available at: https://juniorworld.github.io/YunerZhu_Ideology_underreview.pdf

84.

Zhuang

(2016) China’s top internet regulator lays down the rules on live-streaming programmes. South China Morning Post, 4 November. Available at: https://www.scmp.com/news/china/policies-politics/article/2043012/chinas-internet-regulator-tightens-restrictions-live

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.28 MB