Elections with candidate filtering and two mechanisms of demobilization effect: the prologue to Hong Kong’s authoritarian turn

ABSTRACT Previous studies of electoral authoritarianism identified that manipulations demobilize opposition supporters. Yet, less is known about whether radicals are more prone to abstention than moderates in manipulated elections. To answer this question, we disentangle two mechanisms of demobilization effect—the efficacy mechanism and the electoral supply mechanism—that have different expectations on the turnout rate of radicals and moderates. Our research leverages the disqualification controversy in Hong Kong in 2016, after which radical candidates who advocate self-determination or independence were filtered out from the electoral market. Using both aggregate-level and individual-level data, our analysis shows that a substantive demobilization effect exists. Crucially, we find that radicals and moderates are demobilized to a similar extent, and the decreases in perceived electoral fairness and importance of voting are similar between the two factions. These findings suggest that the efficacy mechanism is a more plausible explanation of the demobilization effect. Overall, this study extends our understanding of voting behavior and political attitude of opposition supporters in face of autocratization.


Introduction
Various studies of EA regimes found that manipulations demobilize opposition supporters (Frantz 2018;Hiskey and Bowler 2005;McCann and Domınguez 1998). At the same time, scholars also noted that the opposition camp in EA regimes is filled with divisions, as some factions have more radical demands than the others (Greene 2007;Magaloni 2006). These two phenomena-demobilization effects driven by manipulations and divisions within the opposition-naturally point to the question of whether radicals are more likely to abstain than moderates. Though this question has important implications for opposition electoral performance, regime cooptation strategies, and democratization, it has not caught much researchers' attention.
This lacuna is puzzling because the current literature implicitly identified two main mechanisms of demobilization effects, which have different expectations on the turnout rate of radicals and moderates. One mechanism is the efficacy mechanism that focuses on voters' perceived electoral fairness and their perceived importance of voting (Hiskey and Bowler 2005;Nikolayenko 2015;Simpser 2013). This mechanism views opposition as a bloc, and electoral manipulation affects moderates as much as it affects radicals. As such, when there is manipulation, both factions should similarly perceive elections as more unfair and voting as less crucial. Hence, the efficacy mechanism considers that radicals and moderates are demobilized to a similar extent.
Another mechanism is what we call the electoral supply mechanism. It suggests that voters are unwilling to trade off factional interests/ideology for their commitment to democracy, when their most favorable candidate is unavailable in the electoral market (Gandhi and Ong 2019). In case dictators exclude radical candidates from the electoral arena, radical voters will not set aside factional interests/ideology and vote for moderate candidates whom they disfavor. Accordingly, radicals should be more likely to disengage from voting than moderate voters. Although these two mechanisms have different expectations on the turnout rate of radicals and moderates, previous works have not explicitly disentangled them. This is what motivates our research: we test whether radicals are more likely to abstain than moderates when there is electoral manipulation and ascertain which mechanism is more plausible in explaining demobilization effects. To do so, we use the case of Hong Kong, where there was a clear authoritarian turn in the 2016 general election and the sovereignty issue has become a conflict line within the opposition camp since the early 2010s. The emergence of the center-periphery cleavage was mainly driven by different mainland-Hong Kong integration projects and political stagnation (Fong 2017b;Kwong 2016a;Ma 2015). Along this cleavage, radicals consider that only by demanding sovereignty can democracy be attained (Veg 2017), while moderates do not contest such issues. But since demanding sovereignty is unacceptable for the Chinese Communist Party (CCP), the opportunities of public contestation for radicals were significantly reduced during the 2016 Legislative Council (LegCo) general election. Specifically, several members-elect of the radical faction were disqualified after this election, which was unprecedented. In the subsequent by-elections, radical candidates that advocate self-determination or independence were clearly barred from entering the election. This autocratization episode provides us with a valuable opportunity to fill the above research gap. As a preview, our results show that opposition supporters were substantially demobilized in the post-disqualification by-election. Crucially, we find that radicals and moderates are demobilized to a similar extent, and the decreases in perceived electoral fairness and importance of voting are similar between the two factions. Hence, we suggest that the efficacy mechanism is more likely to explain the demobilization effect.
The contributions of our research are twofold. First, previous research noted that one easily underestimates the extent of demobilization effects, for dictators can push up turnout figures by stuffing ballot boxes and forcing citizens to vote (Simpser 2012). But because manipulations that enhance turnout had been scarce in Hong Kong, our research circumscribes this concern. Second, our research contributes to the literature on autocratization, for electoral manipulations are common during autocratization (Lührmann and Lindberg 2019; Tomini 2021). By disentangling the two mechanisms of demobilization effects, our research helps us understand how citizens' voting behavior and attitude can change when EA regimes experience autocratization.

Electoral manipulation and two mechanisms of demobilization
Electoral manipulation can take different forms. In terms of the demand side, rulers or pro-government parties can intimidate voters by force, engage in electoral fraud, threaten voters by withholding pork-barrel projects, and gerrymandering to change partisans' distribution (Frye et al. 2019;Reuter and Szakonyi 2021;Wong 2019). On the other hand, EA regimes can manipulate the supply side by barring opposition parties and candidates from entering the electoral arena, intimidating them by judicial process, and restricting their access to mass media and campaign finance (Hyde and Marinov 2012;Szakonyi 2021).
Regarding the impact of electoral manipulation on turnout among opposition supporters, the literature has mainly found that the effect is negative in direction. Yet, there are two mechanisms to explain such demobilization effects, which have different expectations on the turnout rate of radicals and moderates. One mechanism is the efficacy mechanism that focuses on voters' perceptions of electoral fairness and the significance of voting. In this regard, opposition supporters who perceived elections as fraudulent and/or insignificant should be more likely to abstain (Hiskey and Bowler 2005;McCann and Domınguez 1998;Nikolayenko 2015). Simpser, who is a key advocate of this mechanism, argues that manipulation can reveal information about the manipulator's strength and can therefore shape public perceptions (2012). If opposition supporters perceive the EA regime as invincible such that electing any single candidate would not make any appreciable impact, they would believe election outcomes are already predetermined.
That is why some opposition supporters become disenchanted or apathetic and abstain from voting when there is manipulation (Simpser 2013).
The efficacy mechanism certainly echoes the rational choice theory of voter abstention, in the sense that low pivotality in the electoral arena curtails turnout. If opposition candidates in EA regimes are perceived as unable to respond to citizens' demands, opposition supporters abstain from voting because they doubt the benefit of voting in a rigged election (Bahry and Silver 1990). Also, the efficacy mechanism can be linked to the literature on expressive political behavior (Hillman 2010). That is to say, abstention in EA regimes is an "expressive choice" by some opposition voters, for they view the unfair electoral process as a farce and do not want to be part of it (Ansolabehere and Persily 2008; Davis and Coleman 1983).
One crucial idea of the efficacy mechanism is that electoral manipulation affects opposition supporters as a bloc. Suppose an EA regime manipulates elections by excluding radical candidates from the electoral arena, the efficacy mechanism would argue that opposition supporters as a bloc are less likely to vote, regardless they are moderates or radicals. It is because this mechanism considers that candidate filtering, as a kind of manipulation, is targeting the opposition camp as a whole. As such, the perceived fairness and importance of elections should drop similarly among radicals and moderates. Hence, according to the efficacy mechanism, the demobilization effect should not be heterogeneous among different factions. This leads us to the following hypothesis: H1a: If radical candidates are filtered out, moderates and radicals are similarly more likely to abstain.
H1b: If radical candidates are filtered out, the drops in perceived fairness and importance of elections are similar among radicals and moderates.
The other mechanism that explains demobilization effects-the electoral supply mechanism-is not well formulated in the literature, but it can be extrapolated from the research that investigates divisions within the opposition camp (Greene 2007;Magaloni 2006). From the EA regime literature, we know that the opposition is often a heterogeneous entity and is filled with ideological divisions (Howard and Roessler 2006). The ideological divisions naturally raise the issue of whether opposition supporters from different factions are willing to trade off their factional interests/ideology for commitment to democracy, when their most preferred candidate is not in the electoral market. The answer provided by the electoral supply mechanism is negative.
This unwillingness to prioritize democracy over factional interests/ideology is clearly demonstrated in the experimental study conducted by Gandhi and Ong (2019). Using the case of Malaysia, they test whether supporters of different parties are willing to engage in cross-party voting, when their favored parties form alliances with other parties. They found that opposition supporters are less likely to vote for the alliance when they know the coalition partner whom they least preferred is going to lead the government. In other words, opposition supporters uphold their factional interests/ideology and do not engage in cross-party voting, even though such voting is necessary to win over the authoritarian incumbent. Following this line of thought, we can conceive that if one faction's most preferred candidate is not on the ballot, supporters from this faction would not vote for a candidate from another faction, although such voting can increase the opposition candidate's chance to win. It is because, according to the electoral supply mechanism, these supporters care about factional interests/ideology. Put differently, faction supporters are unwilling to make compromises on factional interests/ideology and prioritize democracy.
The electoral supply mechanism is certainly more likely to occur when rulers create a divided structure of contestation, in which moderate candidates are allowed to participate in elections while radical candidates are filtered out (Lust-Okar 2005, 1). Since the most favorable candidate for radicals is unavailable, the electoral supply mechanism would argue that radicals are less likely to vote for the moderate candidate than moderate supporters. It is because the factional differences between the moderate candidates and the radical supporters are huge enough to trump over radicals' commitment to democracy. In other words, radicals put factional interest/ideology first and commitment to democracy second. Hence, the electoral supply mechanism has the following expectation: H2: If radical candidates are filtered out, radicals are more likely to abstain than moderates.
Although the above two mechanisms have different expectations on the turnout rate of radicals and moderates, previous studies have not ascertained which mechanism is more plausible. This research gap is surprising because opposition camps in EA regimes are often filled with divisions. Considering this, we leverage the disqualification controversy in Hong Kong to disentangle these two mechanisms.

Hong Kong's candidate filtering
In this section, we describe the political landscape of Hong Kong and explain why its autocratization episode during 2016-2018 can disentangle the two mechanisms. As a subnational entity under CCP's rule, Hong Kong had neither been considered totally authoritarian nor democratic (Fong 2017a;Ma 2007). Just like other EA regimes, Hong Kong's opposition camp is hardly homogenous. One key ideological division is the center-periphery cleavage, which concerns Hong Kong's relationship with mainland China (Kwong 2016b;Ma 2011;Wong 2020;Wong and Wan 2018). After the Umbrella Movement, the opposition was clearly separated into the moderate and radical factions along the center-periphery cleavage. 1 On the one hand, the moderates are led by traditional democrats, who consider the democratization of Hong Kong as a building block of a "democratic China" (Kwong 2018). As such, the issue of sovereignty is not on the moderates' agenda. On the other hand, the radicals advocate self-determination or even independence (Fong 2020;Veg 2017). These demands are certainly beyond the acceptable range of CCP due to its threat to China's sovereignty.
Nevertheless, the localist candidate that belong to the radical faction could still enter the Feb 2016 by-election. In this by-election, the moderate candidate, Alvin Yeung, barely won the seat against the establishment candidate Holden Chow (37% against 34% of vote share). The localist candidate, Hong Kong Indigenous's Edward Leung, obtained more than 15% of vote share. Since this by-election was held for only one vacancy using the plurality system, he failed to get a seat. But after this by-election, it was highly expected that the localist camp could break into the LegCo in the September 2016 general election under a PR system. However, before the general election took place, the Electoral Affairs Commission (EAC) started to restrict the candidature. The EAC required all candidates to sign a "confirmation form" in the nomination process, declaring that they recognize Hong Kong as an inalienable part of China. Consequently, Edward Leung and four other candidates were barred from candidacy, but all other radicals could still enter the electoral competition. Youngspiration, which belonged to the radical faction, could get the blessing of Indigenous, and two of its candidates became members-elect. Yet, their oaths were deemed invalid because they changed the wordings and brought a flag of "Hong Kong is not China" during oath-taking. The National People's Congress Standing Committee subsequently declared that the members-elect who took the oath "in a manner in which is not sincere or not solemn" shall be disqualified. After this decree, the Hong Kong government launched a judicial review seeking to invalidate the oath of four other opposition members. The court decided in favour of the government. In total, six opposition LegCo members/members-elect were disqualified, and such disqualification was the first since the handover. Soon after, in March 2018, a by-election was held. In this election, four vacancies were to be elected (three Geographical Constituencies (GCs) and one Functional Constituency (FC)) 2 and all radical candidates were barred from candidacy. Full support from the radical faction was not in place, as the ideological positions of these moderate candidates are rather different from those of the disqualified radical candidates. In the end, the opposition camp won two GC seats but lost one GC seat and the FC one. As a remark, the opposition had never lost a GC seat in a LegCo by-election before the disqualification controversy.
The disqualification controversy in 2016 and the subsequent by-election in Hong Kong provide us a valuable opportunity to disentangle the two mechanisms of the demobilization effect. As delineated above, a divided structure of contestation was created after the disqualification controversy (Yuen and Chung 2018), as radical candidates that advocate self-determination or independence were totally filtered out from the electoral market. In such a political context, the electoral supply mechanism would suggest that radical supporters are more likely to be demobilized than moderate supporters. It is because radical supporters care about their factional interests/ideology and their commitment to democracy is not strong enough to vote for a moderate candidate whom they disfavor. On the other hand, candidate filtering, as a kind of electoral manipulation, can also demoralize or disenchant the opposition supporters as a bloc. Thus, the efficacy mechanism expects both moderates and radicals would similarly perceive the electoral process as more unfair and voting as less crucial. Hence, opposition supporters, no matter moderates or radicals, would be demobilized to a similar extent in the post-disqualification by-election.
Also, from a methodological point of view, the case of Hong Kong is suitable to analyze the mobilization effect because manipulations that enhance turnout have been scarce. As said before, previous research is prone to the problem of underestimating the demobilization effect, for EA regimes can boost turnout figures by stuffing the ballot box and forcing citizens to vote (Simpser 2012). But according to the Clean Elections Index created by the V-Dem project (its scale ranges from 0 to 1), the index of Hong Kong hovered around 0.75-0.8 in 1991-2018, which is rather high compared to other EA regimes (Coppedge et al. 2019). And looking at the disaggregated indicators, electoral manipulations that enhance voter turnout, such as election violence, intimidation of the opposition, and registration fraud, had been largely absent in Hong Kong during this period. 3 Thus, the change in turnout 2 Before the electoral reform in 2021, half of the seats in the legislature were filled by the FCs that mostly represent major business and professional groups. The other half was chosen by universal suffrage based on the five GCs (Ma 2007). 3 The most likely turnout-enhancing manipulation in Hong Kong would be phantom voters, i.e. pro-government organizations use false addresses to register establishment camp supporters (Fong 2017a). We use a local news search engine to check the severity of this issue. Although there were sporadic reports on vote rigging from 2000 to 2018, these cases usually entailed legal consequences and among opposition supporters in the post-disqualification election can largely reflect the demobilization effect.

Data and research design
Our analyses use both aggregate-level and individual-level data and focus on the GC elections. In both levels of analysis, we compare the change in voting behavior and political attitude between the September 2016 LegCo general election and the March 2018 by-election. Using the terminology of natural experiments, the treatment is candidate filtering-the disqualification of members-elect after the 2016 LegCo election and the disqualification of radical candidates in the March 2018 by-election. Here, the treatment group is the opposition supporters while the control group is the establishment supporters, which are not expected to experience demobilization effects. Regarding the coding of moderate and radical candidates, our source includes candidates' social media platforms and newspapers. We coded opposition candidates as radicals if they publicly support self-determination or independence, and moderates otherwise. We did not find any mismatch between the candidate's advocacy and their party's advocacy along the center-periphery cleavage (see coding scheme in Appendix Table A1).

Aggregate-level
The unit of the aggregate-level analysis is the counting station, and we retrieved election results from the Registration and Electoral Office. 4 We are aware that the boundaries of each counting station can be revised before each LegCo election. Fortunately, the issue of boundary shift is not a significant concern in our case, since the most substantial changes are often made before the District Council (DC) election (Wong 2019), and there was no DC election held between 2016 and 2018. We matched 284 counting stations out of 315 in the two elections (Details of the matching procedure and descriptive statistics are reported in Appendix B). Our analysis first verify whether a demobilization effect exists and estimate its magnitude. We use a difference-in-differences (DiD) approach to estimate the change in electorate share of the opposition camp from the September 2016 election to the March 2018 by-election, compared to that of the were mini-scale. Only one confirmed election fraud case (the 2011 King's Park case) involved more than a handful of votes, and a re-election was ordered in that case. 4 It is plausible that some opposition supporters give a void vote to express discontent. However, our analysis cannot explore this possibility because the Registration and Electoral Office does not provide such data. establishment camp. The regression model can be written as follows: where i indicates counting station, j indicates camp (establishment or opposition), and t indicates the period of either election. share ijt denotes the electorate share of each camp in either election. Electorate share is the number of votes each camp obtained divided by the number of eligible voters in that election. opposition j is a dummy variable with opposition camp coded as 1, and establishment camp coded as 0. byelection t is a dummy variable, with the March 2018 by-election coded as 1, and the September 2016 general election coded as 0. The coefficient of the interaction term opposition j × byelection t denotes the demobilization effect (i.e. the change in electorate share of the opposition camp across two elections, compared to that of the establishment camp). To account for possible correlation of the error term within the same counting station, we use cluster robust standard error at the countingstation level. Afterward, we disentangle the efficacy mechanism and the electoral supply mechanism. Here, our key concern is whether the demobilization effect is mainly driven by radicals, or similarly by both moderates and radicals. Since the aggregate-level analysis uses electorate share data, we measure the level of radical support by the proportion of votes radical candidates received in each counting station in the September 2016 general election. We make inference on the assumption that the electorate share in counting stations with high radical support are more representative of the voting behavior of radicals than that of moderates, and vice versa. 5 We calculate the level of radical support in all counting stations and classify them into terciles within each GC. Then, we decompose the DiD coefficient by categories: Since counting-station level data is used, we can only identify the number of votes obtained by radical and moderate candidates. Although one cannot be completely certain that voters who voted radical (moderate) candidates identify with the radical (moderate) faction, we consider the counting-station level data is a good proxy. It is because, from the HKES September 2016 General Election dataset, 58% of voters feeling closest to a moderate party voted for a moderate candidate, while 77% of those feeling closest to a radical party voted for a radical candidate.
where radical ik indicates the level of radical support tercile category k of counting station i. 6 The coefficients of our concern are b 72 and b 73 , which shows whether the candidate filtering has heterogeneous effects on counting stations with different levels of radical support. If the efficacy mechanism is more plausible, these coefficients should be indistinguishable from zero, as both moderates and radicals are demobilized to a very similar extent. Alternatively, if the electoral supply mechanism is correct, these coefficients should be negative and statistically significant since radicals are demobilized to a greater extent than moderates. To make valid inferences from a DiD design, one key assumption is the parallel trend. In our case, the parallel trend assumption is that the change in electorate share of the opposition camp and that of the establishment camp would be the same in the absence of treatment. To address this concern, we leverage the September 2012 LegCo general election and the February 2016 by-election as a placebo, since there had been no candidate filtering in these two elections. 7 Also, we use the placebo sample to check whether the parallel trend assumption holds for all three levels of radical support. This test is important because there is a shift in the electoral system from a general election (PR system) to a by-election (plurality system). Due to this shift in the electoral system in a by-election, some might argue radicals can be more likely to abstain than moderates. Thus, we ascertain the turnout rate of radicals and moderates in a by-election that had no candidates filtering.

Individual-level
To corroborate our result and tap into opposition supporters' subjective perception, we use the Hong Kong Election Study (HKES) dataset, which conducted online surveys with representative samples of the voting population (S. H. W. Wong 2020a, 2020b). We use the cross-sectional data from the September 2016 LegCo post-election survey (n = 1805) and the March 2018 post-election survey (n = 591). We classify the respondents by the party they reported they feel closest to: in the general election, there were 716 establishment supporters, 705 moderate opposition supporters, and 384 radical opposition supporters. In the by-election, there were 236 establishment supporters, 273 moderate opposition supporters, and 82 radical opposition supporters. Respondents who do not identify with the two camps are excluded from our analysis.
Three dependent variables are used, namely voting behavior, the perceived electoral fairness, and the perceived importance of voting. First, for voting behavior, we create a variable "voted for own camp" that is coded as 1 if an establishment supporter voted for an establishment candidate, or an opposition supporter voted for an opposition candidate. It is coded as 0 if they abstain or voted for a candidate that does not belong to the respondent's own camp. Second, for the perceived electoral fairness variable, respondents were asked "Generally speaking, how fair is the Legislative Council election in your opinion?" The answer is an 11-point scale, where 0 means "extremely unfair" and 10 means "extremely fair". Third, regarding the importance of voting variable, respondents were asked to what extent they think voting makes a difference. Because the two cross-sections use different scales 8 , we normalized them to a 0-1 range, where 0 means "voting doesn't make any difference" and 1 means "voting makes a big difference".
For the independent variables, we include a dummy variable for election (September 2016 election is coded as 0 and March 2018 election as 1), two dummy variables for the moderate and radical faction (with establishment camp as the reference group), and their interaction terms. The coefficients of the interaction terms are of our main concern, as they show whether the impact of candidate filtering is homogeneous on moderates and radicals. We also add standard demographic controls in the model, including gender, age and its squared term, education level, income, and property ownership (see Appendix Table C1 for descriptive statistics).

Aggregate-level
To analyze whether the demobilization effect exists, we start with Figure 1. 9 The left panel is a map that shows the changes in electorate share from the 2016 general election to the March 2018 by-election for the opposition camp. Although it is normal that a by-election has a lower turnout, the drop in electorate share is rather severe for the opposition camp. It is not uncommon for the decrease to be in the 15-20 percentage points (pp) range. Moving on to the right panel, we see a scatterplot that shows the distribution of electorate shares of both the establishment camp and opposition camp in these two elections. Each dot represents a matched counting station between the two elections. The scatterplot clearly illustrates that the decline in electorate share is more severe for the opposition camp, for all dots fall below the diagonal line. On the other hand, for the establishment camp, most dots are close to the diagonal line, and some even stay above it, meaning the electorate share of the establishment camp increases in these stations. Thus, these descriptive statistics indicate that there is a demobilization effect among opposition supporters.
To estimate the extent of the demobilization effect, we turn to the DiD regression result reported in Model 1 of Table 1. The opposition camp has, on average, a 7.8 pp lead over the establishment camp in terms of electorate share in the September 2016 general election (p < 0.001). This confirms the traditional wisdom that the opposition camp could obtain majority support in GC elections before the disqualification controversy. Yet, in the March 2018 by-election, while the electorate share of the establishment camp decreases by 2.0 pp on average (p < 0.001), the opposition camp suffers a further 9.3 pp drop compared to the establishment camp (p < 0.001). To visualize the demobilization effect, we created the left panel of Figure 2 based on Model 1. It illustrates that the opposition camp obtains a higher electorate share than the establishment camp in the general election. Yet, in the post-disqualification election, there is a substantive drop for the opposition camp. Crucially, the electorate share of the opposition camp is on average lower than that of the establishment camp. In short, our finding echoes previous studies of demobilization effects, as the candidate filtering decreases the electorate share of the opposition camp by an enormous magnitude.
However, are radical voters more likely to be demobilized, as the electoral supply mechanism expects, or are moderate voters and radical voters similarly likely to abstain, as the efficacy mechanism suggests? We turn to Model 2 of Table 1 to ascertain which mechanism is more plausible. The estimation shows that both moderates and radicals are demobilized to a very similar extent, for the coefficient of "Opposition x By-election x Radical proportion medium" and that of "Opposition x By-election x Radical proportion high" are indistinguishable from zero. To facilitate interpretation, we plotted the right penal of Figure 2 that illustrates the electorate share of the two camps by the level of radical support. First, it indicates that the decline in electorate shares of the establishment camp between the two elections is unaffected by the level of radical support. More importantly, when we compare the general election and the post-disqualification election, the drops in electorate share of the opposition camp are almost parallel across all three levels. This finding is certainly more in line with the efficacy mechanism and contradicts the electoral supply mechanism. It is because the  . Cluster robust standard errors at counting-station level are in parentheses. *p < 0.05, **p < 0.01, ***p < 0.001 electoral supply mechanism expects radical voters are more likely to be demobilized than moderate voters, as radical candidates were filtered out in the March 2018 by-election. Taken together, the efficacy mechanism is a more plausible explanation for the demobilization effect, as both radicals and moderates are equally likely to be demobilized.
We conducted four robustness checks to strengthen our findings. First, to check the parallel trend assumption, we repeat the regression analysis with the September 2012 LegCo general election and the February 2016 by-election as a placebo test, for there had been no candidate filtering in these two elections. Since only one GC was involved in the February 2016 by-election and there were more substantial changes in counting station boundaries, we could only match 33 out of 133 counting stations between these two elections. Model 3 in Table 1 shows that the parallel trend assumption holds: unlike the post-disqualification election, the decline in electorate share of the opposition camp is indistinguishable from that of the establishment camp. That means the decline of electorate share in the pre-disqualification by-election is similar for both camps. Second, we check whether the parallel trend assumption holds for all three levels of radical support in the placebo sample, since there is a change in the electoral system in a by-election. As such, the results found in Model 2 can possibly mask a heterogenous effect among radicals and moderates if the impact of the by-election is not homogenous to begin with. Yet, Model 4 in Table 1 shows this is not the case: the declines in electorate shares are similar among all three categories of radical support in the Feb 2016 by-election. This signifies moderates and radicals in the pre-disqualification period are similarly likely to abstain in a byelection (see also Figure D1).
Third, we exclude the possibility that the decline in the electorate share of the opposition camp in the post-disqualification period is caused by a real change in popular support rather than candidate filtering. We find that the support towards the two main parties from each faction was rather stable in this period (see Appendix Figure D2). So, the decline in opposition electorate share cannot be attributed to the changes in support. In other words, the stable support rate of the opposition camp is less likely to translate into votes in the post-disqualification by-election. Lastly, we recode a fuzzy case that belongs to the radical faction (i.e. League of Social Democrats (LSD)) as moderate and rerun the regression (see Appendix Table E1). The results remain substantively the same and support the efficacy mechanism.

Individual-level
We now turn to the individual-level analysis to further analyze the two mechanisms. Recall that the efficacy mechanism argues electoral manipulations should affect opposition supporters as a bloc. Thus, when EA regimes manipulate elections by excluding radical candidates from the electoral arena, opposition supporters, no matter moderate or radicals, are more likely to abstain. Also, the perceived fairness and importance of elections should be dropped to a similar extent among radicals and moderates. Contrarily, the electoral supply mechanism argues the demobilization effect should be heterogeneous among different factions. As candidates with radical ideology are not in the electoral market, radical supporters are more likely to abstain than moderate supporters, for factional interest trumps their commitment to democracy. Table 2 and Figure 3 reaffirms that the efficacy mechanism is more plausible. 10 First, concurring with our aggregate-level analysis, moderates and radicals were both less likely to vote in the post-disqualification election (p < 0.001 and p < 0.05 respectively), and the decreases in turnout are statistically indistinguishable between the two factions (p = 0.239). In contrast, establishment supporters are about as likely to vote in both elections (p = 0.274). Next, moving on to the perception of voters, we find that establishment supporters' perceived electoral fairness and perceived importance of elections remained almost unchanged across the two elections (p = 0.809 and p = 0.369 respectively). However, that is not the case for opposition supporters. Compared to the general election, moderate supporters perceived the by-election as less fair (p < 0.01) and considered voting as less important (p < 0.05). Though the estimates are rather noisy, we observe similar negative trends among radical supporters as well. Also, the decreases in perceived electoral fairness and importance of voting are statistically indistinguishable between the two factions (p = 0.420 and p = 0.841 respectively).
As a robustness check, we repeat the regression using the logit model or ordinal logit model (see Appendix Table F1). Next, we again recode LSD as moderate and rerun the regression (see Appendix Table F2). We also add Hong Kong-Chinese identity of respondents, together with their rating of  leaders from the establishment, democratic and localist camps as controls (see Appendix Table F3). The results remain substantively robust. Overall, we find that for moderate supporters, there are significant decreases in perceived electoral fairness and importance of voting, and such decreases are similar to those of radical supporters. Yet, these effects are not found among establishment supporters. In addition, the decrease in the propensity to vote is similar between the two factions. So, these findings are consistent with the expectations of the efficacy mechanism.

Conclusion
In this article, we focus on Hong Kong's autocratization episode in 2016-2018 and study how its candidate filtering affects voting behavior and political attitude. By using this case where turnout-enhancing manipulation is negligible, we first show that candidate filtering leads to a substantive demobilization effect: compared to the establishment camp, the opposition camp loses around 10 percentage points of their electorate in the post-disqualification election. More importantly, we show that both moderates and radicals are demobilized to a similar extent. In addition, supporters of both factions perceive the electoral process as more unfair and consider voting as less important. Taken together, our findings suggest the efficacy mechanism is a more plausible explanation than the electoral supply mechanism.
These findings certainly shed light on the recent political developments in Hong Kong. The 2019 DC election, which was held during the Anti-Extradition Bill movement, once again vindicated that candidate filtering affects opposition supporters' incentives to vote. In this election, EAC disqualified only one prominent radical candidate, Joshua Wong, but allowed all other radical candidates to enter the electoral market. In the end, the opposition camp won a landslide victory as it benefited from the high turnout driven by the movement. This speaks to the idea that demobilization effects can be rather sensitive to the political landscape. Also, our research has implications for the voting behavior in the recent autocratization episodes. In March 2021, the National People's Congress adopted a much more restrictive electoral system, and one key measure was to institutionalize a vetting system for the LegCo candidacy. Considering this more restrictive candidate filtering institution, we can foresee that more opposition supporters will abstain or cast invalid votes in LegCo elections in the National Security Law era, for the result is largely predetermined.
Theoretically, as electoral manipulations are prevalent during autocratization (Lührmann and Lindberg 2019;Tomini 2021), this research enriches our understanding of how opposition supporters of different factions react in face of autocratization. From previous studies, we know that manipulations demobilize opposition supporters (Frantz 2018;Hiskey and Bowler 2005) and the opposition camp is riddled with divisions (Gandhi and Ong 2019;Greene 2007;Magaloni 2006). Yet, scholars have not investigated whether radicals are more likely to abstain than moderates, even though the two mechanisms being studied in our research have different expectations on the turnout rate of radicals and moderates. Hence, our research contributes to the field by analyzing this question and ascertaining which mechanism is more plausible in explaining demobilization effects.
Nonetheless, our research has some limitations that can only be addressed by future studies. First, the post-disqualification election that we analyze is a by-election. Although the salience of this by-election was like that of a general election, this by-election is at most a quasigeneral election setting rather than a general election proper. Hence, our research is unable to ascertain the extent of the demobilization effect in a general election with candidate filtering. Second, the two mechanisms examined in this research can be confounded with candidates' valence factors. Given that Hong Kong elections are elite-centered, it is possible that opposition supporters abstain because the valencecompetence and integrity-of the opposition candidates were rather different in the September 2016 and the March 2018 elections. However, our data is not fine-grained enough to isolate and measure this valence effect on voter turnout. Third, just like other case studies, the scope conditions of our research are specific: the distinction between moderates and radicals is linked to the issue of sovereignty and the political context is a divided structure of contestation. Therefore, future studies can compare the intention to vote among moderates and radicals in EA regimes that have different cleavage configurations and a unified structure of contestation.

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