Defending Stability under Threat: Sensitive Periods and the Repression of Protest in Urban China

ABSTRACT How does the elevated threat of protests during sensitive periods affect state repression in a high-capacity authoritarian regime? Drawing on a dataset of over 3,100 protests in three Chinese megacities, this study provides three key findings: first, the frequency of protests before and during national-level focal events and subsequent to national-level disruptive events is depressed, suggesting preemptive repression is taking place. Second, the likelihood of responsive repression is marginally reduced before and during local-level focal events and slightly elevated after national-level disruptive events. Third, contention is intensified when local political elites meet. Sensitive periods do not bring contention to a standstill and costly bursts of responsive repression were not observed. Stability maintenance during times of increased regime-vulnerability was thus less rigid than often assumed.

not new to scholars of contention in China, 5 such relatively free-wheeling protest in central areas of major cities during a hyper-sensitive period does not sit easily with research redolent of an increasingly muscular and repressive Chinese security state. 6 Thus, are these examples rare exceptions from the rule? Or is repression during sensitive periods less certain than frequently believed?
A sensitive period is the time prior to an event (that is known in advance) or subsequent to an event (that is not known in advance), where the event itself can serve as a coordination mechanism for contentious mobilization. During such periods, authoritarian states are assumed to perceive protests as more threatening than usual. 7 The long-accepted wisdom in research on contentious politics is that as the threat of protests increase, so does state repression. Recently, however, political scientists have observed that this does not necessarily hold when states can apply preemptive repression to stifle protests in advance. 8 Qualitative studies from China have shown that the state's capacity to prevent protests is highly developed. 9 Quantitative research has indicated that preemptive detentions of political dissidents rise before and during politically sensitive dates known in advance, and subsequent to some events not known in advance. 10 Does this pattern of temporally inflated preemptive repression extend to protests? How does the state respond when mobilization occurs nonetheless? What is the extent of preemptive and responsive repression during sensitive periods?
These questions have not yet been systematically studied in China or anywhere else. To address them, this study draws on a hand-coded dataset of over 3,100 protests in three Chinese megacities (Chongqing, Guangzhou and Shanghai) that occurred between January 2014 and May 2016. Studying repression during sensitive periods provides a prism through which we can observe how the Chinese state responds to dissent when it perceives the political order it is tasked to defend as more vulnerable than usual. How a state deploys its coercive capacity 'under threat' can also provide insights into how secure or insecure the political leadership feels and how well-oiled the coercive apparatus is.
The article first looks at research on protest, threat and repression before it examines what we know about sensitive periods and the inclination to repress in China. Based on this existing research, it generates testable hypotheses. After an introduction to the data and methods used, empirical findings on the effects of sensitivity on patterns of contention and repression are presented. The conclusion summarizes the results and discusses their broader implications.

Protest, Threat and Repression
When states are challenged, they have a strong tendency to respond with some form of repression. 11 State repression is frequently understood as actual or threatened punishments by state agents or their proxies to deter specific activities by individuals or groups whom the state perceives as threatening. 12 Contentious politics scholarship has identified what kinds of protest attributes are particularly threatening to states, and are thus more likely to be repressed, and how states' repressive inclination varies across time, place and regime type. 13 The repression examined in this literature always takes place in response to mobilization and can thus be termed 'responsive repression.' 14 Asecond line of research has begun to investigate other hidden and covert forms of repression. This type of repression typically occurs before mobilization becomes manifest and seeks to prevent protests. We may, therefore, refer to it as 'preemptive repression.' 15 Since democracies face substantial constraints in applying repression before a potentially illegal act has occurred, 16 preemptive repression is more widespread in non-democracies, including China. 17 It is well established that as the intensity of threat increases, so does repression. 18 According to this logic, an increase of protest intensity-all else being equal-leads to an increase of threat, which provokes an increase of responsive repression. Ritter and Conrad have introduced a new twist to this discussion by highlighting that observable protest, preemptive and responsive repression are intertwined. Preemptive repression acts as a filter which identifies protests that are perceived as particularly threatening by states. Thus, preemptive repression not only reduces the number of protests that become observable, it may also affect the threat that protests pose, and therefore the likelihood that state agents engage in responsive repression. Ritter and Conrad demonstrate that in authoritarian regimes, where preemptive repression is more common, the statistical link between observable protest and repression disappears once this endogeneity is controlled for, while it remains stable in democracies. 19 Building on these findings, Truex has conjectured that preemptive repression in authoritarian regimes should vary with time, because the threat of dissent is higher during certain periods than others. In a pioneering article, he demonstrated that preemptive detention of political dissidents in China was substantially augmented during periods when the state is more threatened by dissent. 20 11 See, in particular, Christian Davenport, 'State Repression and Political Order', Annual Review of Political Science 10(1), (2007)

Sensitive Periods and the Inclination to Repress in China
The fact that certain periods are more sensitive than others and that the state tenses up during 'sensitive periods' is a basic fact of political survival in China. Local officials instruct their security personnel to pay close attention to 'social stability management and control work' (维稳管控工作) 21 or to 'safeguard overall social stability' (确保社会大局稳定) 22 during these times. There are thus ample reasons to assume that protests occurring during sensitive periods are-all else being equalperceived as more threatening by the state. What is not known is how this increase of threat translates into different types of repressive action, and if it leads to a discernable increase in the intensity of protest repression.
While often evoked and thought to be self-evident, what defines a sensitive period? Frequently mentioned sensitive periods in China include the anniversary of the 4th June military repression of protesters in 1989, major political gatherings such as the annual Two Meetings of the National People's Congress (NPC) and the National People's Political Consultative Conference (NPPCC) and the National Day celebration. 23 What these have in common is that they are known to everyone in advance and therefore 'might serve as a silent signal' 24 for collective action by 'unorganized interests', meaning large numbers of citizens who share common grievances. 25 Truex defined such 'focal events' as 'dates known in advance that have high salience and reduce problems of coordination for citizens seeking to act collectively.' 26 Hence, in the periods before and during such dates the perceived threat of protest is likely to be elevated.
Many other events that can trigger collective action, however, are not known in advance. In China the deaths of Premier Zhou Enlai in January 1976 and former Chinese Communist Party (CCP) General Secretary Hu Yaobang in April 1989 are prominent examples of unforeseen events that served as opportunities for unorganized interests to coordinate collective action, 27 initiating two major regime crises around the 1976 Tiananmen Incident and the 1989 Tiananmen Movement. Unanticipated opportunities for coordination can also include situations signaling division within the leadership, such as purges, or protests abroad that may diffuse domestically. 28 Likewise, natural catastrophes, accidents or scandals can trigger unrest. What such unforeseen events have in common is that they represent 'a surprising break with routine practice' and potentially disrupt existing orders. 29 The Chinese state is keenly aware of this threat. It conceptualizes such events as 'sudden public incidents' (突发公共事件) and has issued detailed response plans that instruct officials at all levels of government on how to uphold 'social stability' under such conditions. 30 Below, this type of sensitive period will be understood as a disruptive event. It is assumed that in the period after disruptive events, the perceived threat of protest by the state should be elevated.
Research from China provides evidence that suggests preemptive repression is widespread in anticipation of focal events. Aside from regime opponents being detained, 31 known petitioners and other potential protesters are warned, put under surveillance or detained, while security personnel are stationed around potential gathering sites in city centers. 32 However, some studies suggest that taking action when political leaders pay more attention to stability than usual can also provide an opportunity for extraordinary concessions. Hence, some activists specifically target their protests to focal events. 33 Evidence also suggests preemptive repression often takes place after disruptive events. Unanticipated 'trigger events' were identified as one of the four key threats to stability that the security apparatus is geared to defuse in one rural county. 34 After the Color Revolutions in the early 2000s, the Arab Spring events in 2011, and the Hong Kong Occupy Central movement, Chinese security agencies engaged in extensive preemptive repression to avoid diffusion into Mainland China. 35 However, the available literature does not provide much insight into: (a) what the extent of preemptive repression is for both focal and disruptive events; (b) what the consequence of conflicting incentives for protesters to stage, and for the state to prevent, protests before and during focal events is; and (c) how responsive repression is applied when protesters do organize protests during sensitive periods. Nonetheless, this body or research lays the groundwork for a number of testable hypotheses that can take our understanding of these issues forward.

Hypotheses
The Chinese security apparatus has a proven track record of conducting widespread preemptive repression. Since the perceived threat of protests is elevated during sensitive periods, it can be assumed that preemptive repression is intensified before and during focal events and after disruptive events. When preemptive repression becomes more intense, the frequency of observable protests should be depressed.

H1:
The frequency of protests is depressed before and during focal events.
H2: The frequency of protests is depressed after disruptive events.
The threat of focal and disruptive events to the state lies in their potential use as a coordination mechanism for widespread contention. Thus, although the threat of protest during sensitive periods is elevated across the board, protests that occur in highly visible and symbolically potent locations in city centers can be assumed to be perceived as particularly threatening. Preemptive repression should therefore be strongest for protests that take place in central locations. H3: The reduction in the frequency of observed protests is stronger for protests that occur in highly visible central areas.
Our priors on responsive repression are less straightforward. With regard to focal events, two contrasting scenarios are plausible. On the one hand, preemptive repression should filter out many of the most threatening protests, leading the state to be less coercive on those protests that occur after filtering. Moreover, focal events are periods of increased domestic and global attention on China's urban centers, which can translate into a higher inclination to provide concessions. 36 This, in turn, may well spill over into less responsive repression. In addition, public attention is also known to increase the likelihood of a repression backlash. 37 A well-coordinated security apparatus should thus avoid responsive repression during such periods.
H4a: Responsive repression is reduced before and during focal events.
On the other hand, however, not all potential protesters are known to the state and can thus be subjected to preemptive repression. Moreover, die-hard activists who manage to circumvent preemptive repression are particularly threatening to local leaders tasked with maintaining stability. This effect would lead responsive repression to increase.

H4b: Responsive repression is intensified before and during focal events.
With regards to disruptive events, the sensitive period begins unexpectedly and no preparation is possible. Protests which the state perceives as particularly threatening can therefore not be filtered out through preemptive repression in advance. The inclination to repress responsively should therefore be stronger than for focal events. Thus, it can be assumed that responsive repression should increase.

Data Collection
This study draws on a hand-coded dataset of 3,107 protests from three Chinese megacities (Guangzhou, Shanghai and Chongqing) that occurred between 1 January 2014 and 31 May 2016. A protest event was defined according to Tilly's definition of a contentious gathering. 38 After undergoing intensive training, three coders independently coded the data under close supervision by the author. 39 To avoid duplication of protest gatherings, the research followed the 'Dynamics of Collective Action' project manual. 40 Data were collected from four different types of sources: (1) non-domestic English and Chinese language news media, 41 (2) Chinese domestic news media, (3) Chinese online dissident media and (4) Tilly defines it as 'an occasion on which a number of people (here, a minimum of 10) outside of the government gathered in a publicly accessible place and made claims on at least one person outside their own number, claims which if realized would affect the interests of their object'. Tilly, 'Contentious Repertoires in Great Britain, 1758-1834'. 39 The 25 cases reported in non-domestic English news media were added later. These cases were coded by the author. 40 We coded gatherings for which the protest location, group and issue were the same and there were not more than 24 hours between them as one protest event. When a gap between an event and a preceding event was larger than 24 hours but the other conditions applied, the protests were recorded as separate events in a series of protests. Doug McAdam et al., 'Documentation', Dynamics of Collective Action, 2009, http://web.stanford.edu/group/collectiveaction/cgi-bin/drupal/node/3. Chinese social media. Non-domestic English news media articles were extracted manually through keyword queries from the LexisNexis database, which contains more than 1,000 news outlets worldwide. 42 Non-domestic Chinese and domestic news media were collected with similar procedures from the Wisenews database, which contains over 600 mostly Chinese language newspapers from Mainland China, Hong Kong, Macao, Taiwan and South East Asia. In addition, three major Chinese language news media platforms (BBC Chinese, Duowei News and Radio Free Asia Chinese) were screened manually or with search engine keyword queries, depending on the source. Chinese dissident media sources stem from six websites that are run by political dissidents and are well known for covering sensitive political news and protests in Mainland China (Boxun, Canyu, China Labour Bulletin Strike Map, Minsheng Guancha, Zhongguo Molihua Geming and 64tianwang). Again, these were screened manually and through search engine queries. Social media data were drawn from the Wickedonna blog. 43 It is composed of more than 200,000 posts related to more than 67,000 protest events, collected and published online by activists Lu Yuyu and Li Tingyu. The coverage begins in July 2013 and ends in June 2016, when Lu and Li were arrested. 44 The activists manually searched Weibo and other social media platforms for relevant material, mostly eyewitness accounts of a particular protest, then examined these posts for their veracity and published the results on their website. 45 Protests from the three cities and the observation period were extracted from the blog through manual inspection.
The coverage of these four source types is very uneven. By far, most of the protests were reported on social media (2,897 protests, 93.24%). The second largest source was dissident media (545 protests, 17.54%). Non-domestic English and Chinese language news media covered 95 protests (3.06%), while domestic news media covered just 39 (1.26%). 46

Variables and Measurement
The number of protests. Preemptive repression has the effect of preempting protest events that would otherwise take place. Ergo, it cannot be directly observed in protest event data. However, the number of protest events that occurred in a city on a given day (city-day) can nonetheless provide insights into the effects of repressive state power on mobilization. The underlying assumption is that if potential confounders are controlled for, the observable association between a sensitive period and the number of protest event per city-day reflects the impact of preemptive repression. However, activists may also 'voluntarily' refrain from mobilizing due to the expectation of responsive repression. While it is impossible to disentangle one from the other, in both scenarios protests do not take place as a result of concern over state punishments.
Thus, all else being equal, a reduced number of protest during a sensitive period is an indication that preemptive repression or the fear of responsive coercion have depressed the observable number of protests. If the frequency of events is not reduced or even higher than usual, it indicates that concern over state sanctions did not have a robust deterring effect. This provides valuable insight into how state behavior and citizens' expectations of it affect contentious action.
Responsive repression was tapped with three commonly used binary variables: policing (whether or not police was reported to be present at the protest site), arrests (whether or not arrests at the protest site were reported) and coercion (whether or not coercive methods were used). These were recoded into a categorical variable with three values: 0) no policing, 1) policing (without coercion or 41 Non-domestic Chinese language news media includes outlets from Hong Kong and Taiwan. 42 For keywords used, please refer to Appendix A. arrests) and 2) policing with coercion or arrests. This coding assumes that the very presence of the police and the threat it poses to protesters is a form of repression. It also allows for a more finegrained observation of reactive state measures. If alternatively, only coercion or arrests were used as a measure of repression, the indicator would not distinguish between protests that have been policed (without coercion or arrests) and those that have not been policed. Evidently, the presence or absence of police during a protest in an authoritarian regime is an important event attribute that should not be overlooked.
Extra care was taken to ensure that insufficient information did not lead to the recording of 'false negatives' for repression. Arrests were recorded as absent (0) only when there was no textual evidence confirming the presence of the variable, and textual information was sufficient to render a judgement. Policing and coercion were recorded as absent (0) only when there was no textual or visual evidence indicating the presence of the variable, and textual or visual information was sufficient to render a judgement. If there was no evidence of the presence of one of these variables, but the information was insufficient to render a judgement, they were recorded as missing values. 47 This led to 38.45% missing values for variable measuring responsive repression.
Focal and disruptive events were captured as follows. A research assistant was tasked to go through a list of sources to search for events nationwide and in the three cities between 1 December 2013 and 30 June 2016. 48 Once the initial list was exhausted, sources were expanded until searches did not provide any new information. The resulting list of events was screened by the author and reduced to those that, based on the available literature and background knowledge, were deemed to have the highest political salience. 49 Based on the distinction between focal events known in advance and disruptive events not known in advance, events were distributed into these two categories. Finally, a distinction was made between national and local focal and disruptive events. National events fall within the political purview of the central government and local events are primarily relevant at the local level. It was assumed that local leaders are subject to more pressure from above to avert or stop protests when events of national significance take place.
National focal events include the dates of the annual meetings of the NPC, the CCP Central Committee, the anniversary of the crackdown on the student movement on 4 June 1989, and National Day on 1 October. It also includes the dates of visits of central government leaders or foreign leaders (limited to heads of government or state of G-20 member countries or international governmental conferences) in one of the three cities. 50 Local focal events include the meetings of city and provincial (in Guangzhou) or municipal (in Chongqing and Shanghai) People's Congresses and CCP committees. National disruptive events include the dates when a purge of a national leader (members of the CCP central committee) was publicized and when a pro-democracy protest in Taiwan and Hong Kong began. 51 Local disruptive events include the dates when purges of local leaders (members of the relevant city/provincial/municipal CCP committees) were publicized, when central discipline inspection teams visited the city and when a high profile, non-protest event with 47 For the information sufficiency conditions, see Appendix B.) 48 Three types of events were defined by loosely following Truex' terminology as focal, leadership division events and governance shock events, with examples provided. Truex, 'Focal Points, Dissident Calendars, and Preemptive Repression'. A category of major protests movements in Greater China was added. The three initial sources were: a BBC chronology of main events in China (https://www.bbc.com/news/world-asia-pacific-13017882), Baidu Encyclopedia annual chronologies for China 2014-16 (https://baike.baidu.com) and all articles with relevant city names in the headline or lead published in the South China Morning Post. 49 Instructions for data collection are available upon request. Events judged to have weak political salience were, for example, visits of foreign leaders outside of G-20 countries, Political Consultative Conference meetings at all levels, or the openings of trade fairs or amusement parks. 50 Although these events happen at the local level, they are of national significance and thus fall within the political purview of the central government. 51 While restricted circles inside the state have advance knowledge of such purges, it was assumed that these circles are fairly small and are even smaller for those with knowledge of the precise timing of a purge. relevance to public security occurred (such as a stampede on New Year's Eve 2014/15 in Shanghai and a knife attack in a Guangzhou rail station in 2015).
Please refer to Appendix C for the detailed measurement of the sensitive periods and to Appendix D for the measurement of all variables used in the analysis, including inter-coder reliability measures for hand-coded variables.

Data Validity
All protest event data generated in the absence of a highly valid 'ground truth' have to confront possible biases. Most fundamental here is selection bias, the likelihood that a protest which occurred is covered by the sources used. 52 How biased are the media-elicited data underlying this study? Source diversity is the first important safeguard against bias. 53 To the author's knowledge, this dataset is the most source-diverse protest event database from China. Moreover, media penetration through the Internet, mobile phones and news media bureaus is another safeguard. 54 All these are maximized in large and developed cities. The three megacities from which data were collected therefore provide optimized conditions for protests to be picked up by media sources in China. Likewise, the social media data that has contributed the largest part of the events in this study have emerged as the source with by far the largest coverage of events and the fewest relative biases in data from China. 55 Censorship introduces another layer of potential bias. The following attributes of the data, however, minimize this effect. First, only social media and domestic news media are affected by direct censorship. The other sources in the database are not under the control of the Chinese censorship system. Second, research has revealed that censorship of protests reported in social media is much less pronounced than often assumed. The censorship system targets collective actionrelated posts that become viral, not the large majority of posts that do not get significant attention. 56 Posts that do attract attention are likely about protests which are particularly large, of long duration, or violent. Such protests, in turn, have a much higher chance to be covered by dissident media or non-domestic news media. 57 Again, the diversity of sources used here provides a distinct advantage.
This study faces the additional problem that the intensity of censorship may be temporally increased during the sensitive periods under scrutiny. 58 However, a serious distortion of the findings due to this potential effect is unlikely for the following reasons. First, if the indicator for preemptive repression were to be affected by a temporally increased censorship bias, the observed reductions in protest frequency would only be present in the directly censored sources, or at least be much more pronounced there compared to the uncensored sources. The findings below show that this is not the case. Second, even if observed reductions in protest frequency were to some extent augmented by temporally increased censorship, another observation, that the reduction of protest frequency through preemptive repression is limited in scale, is not affected by such concerns. Hence, the present findings represent an absolute maximum of protest frequency reductions. Third, it is very unlikely that the indicator for responsive repression is affected by a temporally increased censorship bias. For one, such a bias would necessitate that repressed events are singled out for censorship to a larger extent than during other periods. Research into social media censorship, however, has found that censorship is nowhere near such a fine-grained selection of content, but rather sweeping. 59 For another, the findings below show that responsive repression is slightly elevated for one type of sensitive period.
Hence, although the possibility of selection bias in general, and selection bias via censorship in particular, can never be fully excluded in media-elicited protest event data from China, the likelihood of this bias sabotaging the validity of the findings is minimized in this study.

Results
The following analysis proceeds in two steps. First, the effects of sensitive periods on the number of protests per city-day are examined. To do this, the protest data were reshaped to the 2,646 city-days (882 per city) as the unit of analysis. In a second step, the analysis moves to the intensity of responsive repression during sensitive periods. Here the unit of analysis is the protest event, and the dependent variable is whether a protest was subject to a form of repression or not. In both steps, bivariate results are reported before the regression models. Since no firm indications on the duration of sensitivity are available, one-, two-and three-week windows are used for the four types of sensitive periods. Figure 1 displays the number of protests per city-day over sensitive period type and one-to threeweek windows and provides a first impression of the observable patterns. Across all city-days in the sample, an average of 1.17 protests was recorded per day (indicated by the dashed line in Figure 1). For national focal events, national disruptive events and local disruptive events, the number of protests is below average, at up to 1.06 (national focal events), 0.80 (national disruptive events) and 1.01 (local disruptive events) protests per day. By contrast, for local focal events, the intensity of contention is noticeably elevated at up to 1.49 protests per day.

Sensitive Periods and the Number of Protests
Among the national focal events, NPC meetings exert the strongest effect (a reduction down to 0.91 protests; not displayed). The protest frequency prior to 4 June is reduced, but not dramatically so (to 1.02 protests per day for the two-week window; not displayed). This is fairly surprising but mirrors the anecdotes presented at the outset of the article, indicating that this presumably 'hyper sensitive' period does not lead to a standstill of contention. Moreover, protest activity before National Day and Central Committee Meetings are even slightly elevated (up to 1.38 and 1.47, respectively; not displayed). Among national disruptive events, the observable reductions in protest frequency are most pronounced for pro-democracy protests in Hong Kong and Taiwan. During the first week after the beginning of such protests, protest activity in the three Mainland cities is reduced to 0.44 protests per city-day (not displayed). This finding mirrors a robust increase of dissident detentions during protests in Hong Kong, found by Truex. 60 In terms of the elevation of protests before local focal events, the most important are Local People's Congress (LPC) meetings. Before and during these occasions, the daily number of protests rises to up to 1.75 during the one-week window (not displayed). Thus, even if preemptive repression takes place or protesters fear increased repression, congregations of local political elites seem to provoke more protests than usual. For local disruptive events, the strongest protest-reducing effect is observable after high profile, non-protest public security events (up to 0.80 protests per day).
The next step of the analysis moves to Poisson regression models to control for potential confounding factors. 61 The analysis controls for the share of workers' protest. This type of protest is not only the most frequently occurring form of contention in China, it also displays a very distinct seasonality, with peaks in the weeks before Chinese New Year, when workers seek wage payments before the holidays. 62 Since the annual LPC meetings typically fall in this period, it is necessary to test whether peaks of protests during local focal events are largely the result of this seasonality. Moreover, the Occupy Central Movement (28 September, 2014) and the Mong Kok Riots (8 February, 2016) began immediately before or during major national holidays. This may affect the findings for national disruptive events. Hence, the analysis also controls for public holidays. The models further control for periods of dissident repression, city, day of the week, a lagged variable measuring the number of protests on the previous day, as well as a month counter-variable to control for potential trends over time.
The negative and mostly significant coefficients for national focal events and national disruptive events in Table 1 confirm the bivariate pattern. Transformed into marginal effects, the two-week model estimates that these two periods of sensitivity decrease the number of protests by 9% and 21%, respectively. Investigating selected events separately further reveals that (these models are not displayed): NPC meetings lead to predicted protest reductions of up to 14% (two-weeks window). The 4th June anniversaries do not lead to statistically significant protest frequency decreases. By contrast, social movements in Hong Kong and Taiwan lead to a predicted reduction of protest intensity of up to 42% (one-week window; not displayed).
The positive but not significant coefficient for local focal events indicates that the elevated number of protests during this sensitive period is to a substantial extent the result of the seasonality workers' protests. When the share of workers' protests is not controlled for, the coefficients for local focal events become highly significant and the number of protests is predicted to rise by up to 26% (one-week window; not displayed). Evidently, workers are not deterred-and possibly even enticed -by the political sensitivity of this period. Lastly, the protest reduction after local disruptive events does not hold with control variables.
Hypotheses H1 and H2, which assumed depressed protest frequencies for focal and disruptive events, are confirmed for events of national significance but are disconfirmed for local events. Although preemptive repression and protester self-censorship may still have taken place around sensitive periods of a primarily local relevance, these effects were not robust enough to meaningfully reduce contentious activity. This supports the assumption that pressure from the top spurs local officials to repress harder when the central government's interests are at stake. At the same time, activists may expect tougher repression under these conditions and therefore voluntarily refrain from protesting.
Hypothesis H3 assumed that preemptive repression is stronger for protests that take place in highly visible city centers, because contention in symbolically potent areas should be perceived as particularly threatening by the state. To test that assumption, Table 2 displays regression models on the daily number of such highly visible protests only. The models estimate that national focal events (threeweek window) and national disruptive events (two-week window) both reduce protests by up to 24%. These reductions are stronger than those for all protests and provide support for H3. However, protests in city centers during local focal events are predicted to increase by up to 35% (one-week window) even though the share of workers protests is controlled for. Thus, when local political elites congregate, noticeably more protests occurs in symbolically potent central areas.
Finally, two tests were conducted to test the robustness of findings. First, to confirm if the reductions of protests during national focal and disruptive events are not the result of increased media censorship during these periods, Appendix E displays the same models computed only with events recorded by non-domestic sources, which are not subject to censorship by the state. Although significance levels are lower (which is unsurprising given the much smaller number of cases), the models confirm the negative findings for national-level focal and disruptive events. The predicted reductions of protests by 17% for national focal events (two-week window) and 24% for national disruptive events (three-week window) are within the same magnitude as those found in data from all sources. 63 Second, to examine if the effects for the selected dates are not simply random findings, four However, the increased number of protests for local focal events could not be reproduced with this data. A potential explanation is that the protests during these periods tend to focus on local issues and therefore do not make it into nondomestic sources of information. 64 Random dates were generated on www.random.org. Dates with screenshots are available upon request.
'placebo' variables were generated by reproducing the data structure of the events comprising the four types of sensitive periods with random calendar dates. 64 Appendix F shows that the placebos are not significant. The placebos are also insignificant in models for highly visible protest and for protests from non-domestic sources (not displayed). Therefore, the above results are highly likely not random findings, but systematic protest intensity changes in response to sensitive periods. To summarize, the analysis has indicated that the number of protests is depressed before and during national focal events and after national disruptive events. Given the theoretical expectations and substantial case study evidence, this strongly suggests that preemptive repression is taking place during these periods. However, potential protesters' fear of responsive repression may also have contributed to this effect. The same could not be found for local focal and disruptive events. For local focal events, in particular, potentially occurring preemptive measures are not strong enough to avoid a robust acceleration of contention. How do these patterns translate into responsive repression? Table 3 displays the raw percentages of responsive repression for the four types of sensitive periods again with one, two and three-week windows. The last row of the table displays the distribution of the repression variable in the complete sample. For 47.86% of protests no police presence was reported, in 28.21% police was present only, and in 23.93% the police applied coercive measures or arrests. 65 This is a notable finding in its own right. It indicates that the state deploys its considerable capacity for responsive repression in just slightly over half of the protests recorded.

Sensitive Periods and Responsive Repression
When it comes to the impact of sensitive periods, Table 3 indicates that national focal events and local disruptive events leave no observable mark on patterns of responsive repression: the distributions of policing and coercion or arrests for all three temporal windows remain close to the sample means and are statistically insignificant.
Discernable differences in responsive repression exist for local focal events and national disruptive events. For local focal events, the intensity of repression decreases. The share of protests without police presence increases from 47.86% to 52.89% (two-week window). By contrast, for national  disruptive events the intensity of repression increases. The growth is, however, exclusively one of the police presence (from 28.21% to 38.24%, two-week window). Table 4 reports results from a multinomial logistic regression. The analysis controls for protest attributes that are known to affect the state's inclination to repress: protest issues, participant size, the occurrence of protester violence, whether a protest takes place in a highly visible city center or in a rural location, as well as the protest target. 66 Also controlled are the city, the day of the week, a month counter variable and the source type in which the protest was recorded. The displayed relative risk ratios indicate the effect of protest attributes on the risk of being policed or being met with arrests or coercion, relative to not being policed (a value above 1 indicating elevated risk and values between 0 and 1 indicating reduced risk).
Results confirm that repression is somewhat reduced before and during local focal events and slightly increased after national disruptive events. National focal events and local disruptive events do not leave a discernible trace on the state's inclination to repress. The effect of local focal events is most pronounced when the time window is increased to three weeks before the events. Holding all else equal, a protest in the three weeks before and during a local focal event is 3% less likely to be policed and 4% less likely to be met with arrests or coercion. These effects are strongest for LPC meetings, where the likelihood of policing and coercion or arrests are reduced by up to 7% and 8% respectively (three-week window; not displayed). Protest policing after national disruptive events is most pronounced two weeks after the events, while coercion or arrests are not affected. All else equal, a protest occurring in this period is 12% more likely to be policed. Interestingly, this combined effect is not driven by protests in Hong Kong and Taiwan, which on their own do not lead to significant changes in responsive repression. Yet when purges of national leaders were announcedthe second component of the national disruptive events indicator-protest policing is predicted to increase by up to 19% (two-week window; not displayed).
In summary, Hypothesis H4a, which posited that responsive repression is reduced for focal events, receives some support for local focal events. The slightly reduced level of responsive repression here may be a spill-over of an increased inclination to pacify protesters and avoid repression backlash during periods of local political congregations. However, the assumption that responsive repression is reduced because preemptive repression filters out the most threatening protests is not empirically supported. The depressed protest frequency before and during national focal events suggested preemptive repression was increased during these events. However, responsive repression here is unmoved. By contrast, when protest frequencies decline after national disruptive events, suggesting ad-hoc preemptive repression may take place, responsive repression is slightly elevated. Hypothesis H4b assumed that responsive repression is increased before and during focal events. This hypothesis is disconfirmed. Committed activists who stage protests in spite of increased political sensitivity to not seem to pose a sufficient enough threat for the state to turn to more responsive repression. Lastly, Hypothesis H5 assumed that responsive repression is intensified after disruptive events. It is supported for national disruptive events where protest policing is elevated. Given that such events come without much preparation time, and thus preemptive repression can only commence after the threatening event has occurred, the security apparatus does not take any chances and also increases the policing of protests. This pattern further confirms the impression of a rather fine-tuned repressive apparatus. Even under these circumstances, security organs do not turn to significantly more coercion or arrests than usual.
To scrutinize the robustness of findings, two alternative strategies to measure responsive repression have been tested. Since state repression against protesters is sometimes outsourced to nonstate actors, a responsive repression measure which included coercion by private security or thugs was used. 67 Findings are highly similar (results are available upon request). Another variant of modelling repression is to first estimate policing of protests, and in a second step estimate coercion and arrests only for events that have been policed. 68 This variant does not change the gist of the findings either (results are available upon request).

Conclusion
This study has investigated how the Chinese state responds to protests during sensitive periods, when it perceives popular contention as more threatening than usual. The observed reductions in the frequency of protest provided indirect evidence that preemptive repression increases considerably when sensitive periods are within the purview of the central leadership. This finding highlights the fact that the security apparatus in China is tuned to protect the political center and defuse threats to the regime. Results further indicate that the threat of diffusion of pro-democracy protests into the Mainland generates a particularly strong preemptive response by the state. When it comes to responsive repression, the analysis has shown that this type of state coercion is unaffected by national focal and local disruptive events. It declines slightly before and during local focal events and increases slightly after national disruptive events. The magnitude of these differences, however, is limited. In terms of the link between preemptive and responsive repression, it was found that when disruptive events generate political sensitivity overnight, and there is no time to prepare with preemptive repression, the state responds with somewhat augmented protest policing.
A limitation of this study has been the inability to clearly separate the effect of preemptive repression and protester self-censorship. Protest event research in China and other authoritarian contexts should explore how to improve the measurement of preemptive repression of protests, in order to dig deeper into its interaction with observable protests and responsive repression. For China, pro-democracy protests in Taiwan and Hong Kong appear to have a particularly strong depressing effect on contention on the Mainland. This finding mirrors previous research on dissident detentions and warrants closer investigation. 69 Future research should also explore further how protesters use focal events to extract concessions, and how the state manages this strategically timed contention. This study has identified city centers before and during meetings of local elites as particularly fertile grounds for this type of mobilization. Here, both quantitative and qualitative research is needed to investigate these processes in greater detail. Beyond what this study did find, it is also notable what it did not find. The analysis allows to rule out that protests during sensitive periods come to a standstill. The discernible reductions of popular contention have been mostly moderate. The number of protests during periods when local political elites meet was even markedly increased. Likewise, the analysis did not find a dramatically increased level of responsive repression. Hence, stability maintenance during sensitive periods appears much less rigid and more managerial than is often presumed. This points to a rather fine-tuned coercive apparatus in China's urban centers. The state preempts some protests when it manages both previously known and sudden periods of sensitivity. Yet, it still tolerates a considerable amount of contention and avoids costly bouts of responsive repression even during periods when it perceives the political order as more vulnerable than usual. ***p < 0.001, **p < 0.05, *p < 0.1. Cell entries represent unstandardized regression coefficients and standard errors in brackets. Not displayed are the following control variables: City (Chongqing, Guangzhou, Shanghai), Day of week (Mo-Sun), lagged protest, and month counter (1-29).