2.1 NSL's heterogeneous effects

First and foremost, as we have mentioned above, the recent institutionalist turn in the comparative authoritarianism literature has put the inclusiveness of political institutions under dictatorships at center stage. Beyond their nominal functionalities, the key reason for authoritarian institutions to be the mainstay of regime stability is their ability to credibly include and coopt all the major actors whose support is essential to dictators' rule. As the theory goes, the people who are institutionally included and shared with resources will be incentivized to support not only the institutions that make the cooptation possible, but also the regime itself (Boix and Svolik, Reference Boix and Svolik2013). While the theoretical framework is initially formulated for analyzing the intra-elite relationship, the idea of inclusiveness can be readily extended to the public opinions of rank-and-file citizens. Neundorf et al. (Reference Neundorf, Gerschewski and Olar2020) adopt this approach by tapping into the cross-national variation in the pre-transition authoritarian inclusiveness – defined as ‘wider redistribution of socioeconomic and political benefits’ – for explaining citizens' democratic support in post-authoritarian countries. They find that the support tends to be lower among the citizens in countries with a more inclusive regime before the transition since their living standards might not necessarily be substantially enhanced under democracy.

Along the similar line, since our focus is within and not across countries, we argue that institutional trust can also vary between ordinary citizens who feel themselves a part of the regime and those who don't. As far as the effect of repression is concerned, this argument implies that repression will have heterogeneous effects between these two groups. For those who perceive themselves to be outcasts, repressions can further alienate them from the regime and even have a long-term dampening effect on their trust in political institutions and the government (Lawrence, Reference Lawrence2016; Dinas and Northmore-Ball, Reference Dinas and Northmore-Ball2020; Desposato et al., Reference Desposato, Wang and Wu2021; Wang, Reference Wang2021). For those self-perceived regime insiders, however, the ability of the regime to clamp down on protest mobilizations and restore social order should have an opposite effect of boosting their confidence in the regime and its institutions. We, therefore, hypothesize that:

2.2 Protest expectations

Second, while protest repression certainly presents a shock to everyone, it doesn't mean its effect will necessarily be uniform to all. This, therefore, creates another layer of heterogeneity in repression's effect on institutional trust. From an informational perspective, protest repression as an external shock brings people new information about their governments, which can also help them – assuming they are Bayesian – update their levels of institutional trust by forming expectations about the future of political institutions. Frye and Borisova (Reference Frye and Borisova2019), for example, provide a similar reasoning that unexpected protests in Russia gave its citizens who had a strong prior about the government's repression intention new information regarding the ruling elite's tolerance of public opposition, which updated their trust in political institutions upward. In the context of repression, we argue that this Bayesian theory gives rise to different predictions according to one's assessment of its effects. Among the regime insiders, those who believe repression to be effective at taming protesters will find institutions more trustworthy as this piece of new information reconfirms them the resolve and ability of the regime to protect the institutions that benefit them. By contrast, those who don't hold the same belief will lower their trust for expecting declining benefits from their institutional memberships.

Among the regime outsiders, despite for a different reason, the direction of the prediction remains unchanged. While repression is predicted to induce a negative effect on their institutional trust since they derive no utility from political institutions, the degree of the downward Bayesian update will be higher among those who perceive the effect of a repression shock to be low. The reason is that, if the dictator is unable to rein in protest mobilizations through repression, it also implies that there might be more protests and greater needs for repressions that make institutions even less inclusive and trustworthy in the future. By contrast, if repression works from the outset, then it is less likely for such actions to be taken, thus making political institutions ‘less’ untrustworthy. Combined, the Bayesian perspective on repression as a shock gives us a hypothesis:

This second hypothesis also echoes the recent findings by Rozenas and Zhukov (Reference Rozenas and Zhukov2019) that people's loyalty under repressive dictatorship is contingent on the credibility of the dictator's retribution threat to the opposition.

4.1 Data collection, sample, and variable construction

To test the hypotheses specified in Section 2, this study exploits the NSL as an external shock by conducting two surveys with almost the same set of questions in Hong Kong immediately before and after its passage. On 21 May 2020, the Chinese government announced that a new version of the NSL tailor-made for Hong Kong would be deliberated and voted on in the coming session of the National People's Congress held in Beijing between 22 and 28 May 2020. The news brought a shock not only to Hong Kong, but also to the rest of the world since the intensity of the Anti-ELAB Movement had gradually tapered off owing to the COVID-19 outbreak since January 2020. The draft was eventually passed on 28 May.

The first survey began on 15 May and ended on 21 May, the day of the Chinese government's initial announcement about the NSL. We recruited 1, 424 Hong Kong respondents from the online panel maintained by the Rakuten Insight, a global survey company and the surveys were scripted in Qualtrics. The sampling strategy was quota sampling, taking into consideration Hong Kong's population distributions of age and gender.Footnote ¹²

To make sure zero suspicions about the NSL's adoption and wide awareness of it among Hong Kongers, we waited until 10 June, nearly 2 weeks after the law had been officially passed at the National People's Congress in Beijing, to administer the second round, which ended on 26 June. As for the sample attrition of the post-NSL round, given the short span between the two surveys, we were able to retain almost 90% of our pre-NSL respondents for our post-NSL round (1, 256).Footnote ¹³

We included in both surveys questions regarding respondents' basic demographics (e.g., gender, age, and education), attitudes and beliefs regarding Hong Kong's economic and political prospects, protest expectations (to be detailed in Section 5.2 and Appendix C), institutional trust (our main variables of interest), and political stances. See Table B.1 in Appendix B for how our variables were constructed from these questions. While our empirical design certainly enjoyed the advantage of having the NSL as a natural experiment, what should be noted here is that soliciting people's (including both protesters and non-protesters) political attitudes in a contentious context where the freedom of expression was potentially under threat could be difficult. To address the issue, we not only avoided asking our respondents overly sensitive questions such as their previous protest participation, but also employed several de-identification measures to protect them (see Appendix A for details).

4.2 Estimation

The paper's main objectives are to estimate (1) the heterogeneous effect of NSL on institutional trust according to one's sense of regime inclusiveness and (2) the correlation between one's post-NSL protest expectation and his or her institutional trust. To estimate such effects, we first tapped into the exogenous variation in the institutional trust induced by the NSL. For each institution j, individual i's institutional trust Y at time t can be empirically modeled as:

(1)$$Y_{ijt} = \beta NSL_t + \lambda C_{it} + \alpha _{it} + \gamma _{ij} + \delta _{\,jt} + \varepsilon _{ijt}$$

NSL is an indicator variable with 1 denoting the post-NSL period and its coefficient β captures the average effect of the NSL as a shock to i's trust in institution j. C is the variable that denotes i's post-NSL shock in his or belief of Hong Kong society's collective action potential, and its coefficient, λ, is the parameter of interest that captures C's effects on i's institutional trust. If significant, our theory predicts λ's sign to be positive. The model also includes several control variables. First, α _it is a set of individual-time fixed effects that captures other time-varying individual-level attitudes and beliefs that might also affect one's institutional trust. In contrast, δ _jt is a set of institution-time fixed effects that accounts for the influence arising from institutions themselves. For example, an institutional change such as the ongoing electoral reform in Hong Kong that drastically changes how the Chief Executive and LegCo members are elected will definitely have an impact on people's trust in them. In addition, γ _ij is a collection of time-invariant variables at both individual and institutional levels that might also affect institutional trust. For instance, depending on one's view toward the colonial period, he or she might place a higher or lower trust from the very beginning in the institutions that are part of the colonial legacies compared to those created after the handover in 1997. In other words, the institutional origin might play a role in determining one's institutional preferences. Alternatively, the demographic variables that don't change over time cannot be ignored either. The last variable ɛ_ijt is the error term.

The estimation of β is straightforward. We took advantage of the NSL as an exogenous shock to the vast majority of Hong Kongers between the two surveys by pooling their responses together to obtain individual-wise repeated observations over the two periods, i.e., a panel data structure. Since C was only measured in the post-NSL period, the panel specification is given by:

(2)$$Y_{ijt} = \beta NSL_t + \alpha _{it} + \gamma _{ij} + \delta _{\,jt} + \varepsilon _{ijt}$$

Moreover, to estimate λ in (1), we first notice that the individual-institution fixed effects, γ _ij, can be readily eliminated by first-differencing (1) to obtain:

(3)$$\Delta Y_{ij1} = \lambda C_{i1} + \alpha _i^{\prime} + \delta _j^{\prime} + \varepsilon _{ij}^{\prime} $$

where ΔY _ij1 = Y _ij1 − Y _ij0 indicates the difference between i's pre-NSL and post-NSL trust in institution j, and, given the absence of any collective action shock before the law (C _i0 = 0), ΔC _i1 = C _i1. Furthermore, $\alpha _i^{\prime} = \alpha _{i1}-\alpha _{i0}$ accounts for other NSL-induced effects on one's institutional trust, and $\delta _j^{\prime} = \delta _{j1}-\delta _{j0}$ captures post-NSL institutional shocks. Finally, the error term is also re-specified as $\varepsilon _{ij}^{\prime} = \varepsilon _{ij1}-\varepsilon _{ij0}$. What should be noted here is that, given the extremely short span between the two surveys (about 2 weeks), institutional shocks were essentially non-existent for this study and therefore $\delta _j^{\prime}$ should be set to zero to yield:

(4)$$\Delta Y_{ij1} = \lambda C_{i1} + \alpha _i^{\prime} + \varepsilon _{ij}^{\prime} $$

Finally, as our theories predict, the effects of the NSL should be heterogeneous between respondents with different perceptions of regime inclusiveness. To incorporate the effect heterogeneities, the estimation equation is further generalized to be as follows:

(5)$$\Delta Y_{ij1} = \left\{{\matrix{ {\lambda_{PD}C_{i1} + \alpha_i^{\prime} + \varepsilon_{ij}^{\prime} } & {{\rm Pro}\hbox{-}{\rm Democracy}} \cr {\lambda_{PE}C_{i1} + \alpha_i^{\prime} + \varepsilon_{ij}^{\prime} } & {{\rm Pro}\hbox{-}{\rm Establishment}} \cr } } \right.$$

5.1 National Security Law and institutional trust

To measure institutional trust, we asked respondents in both pre- and post-NSL surveys to rate their levels of trust in eight government institutions using a 100-point scale where 0 denotes ‘no trust at all’ and 100 denotes ‘full trust.’ Both central-level – the Central Government, the Liaison Office, the People's Liberation Army (PLA) – and local institutions – the Chief Executive, the Court, the LegCo, the Police, and the Registration and Electoral Office – were included. We took an average of one's levels of trust in the three Central Government institutions that were directly responsible for the NSL legislation and crackdown to construct our main dependent variable of interest, Average Central Government Trust (ACGT).

To formally estimate the NSL's effects on the average government trust, we took advantage of our repeated observations of the same set of respondents right before and after the Law was passed by pooling together the data from the two surveys for our regressional analyses. Moreover, to test our theoretical argument about the NSL's heterogeneous effects, we further divided our respondents further into three groups of political stances: (1) pro-democracy (PRO_DEMO, self-perceived regime outcast), (2) pro-establishment (PRO_ESTAB, self-perceived regime insider), and (3) no stance (NO_STANCE) (see Table B.1 in Appendix B for details).Footnote ¹⁴ Given our tripartite typology of political stances, we set the no-stance to be the baseline for the analyses.Footnote ¹⁵

Our findings not only confirm the NSL's effects on one's institutional trust (Inclusiveness hypothesis), but also paint a very polarized Hong Kong. As Table 2 (model (1) baseline) documents, the NSL induced a mild upward shift by roughly 6% for no-stancers in their average institutional trust in the Central Government institutions. Furthermore, Figure 1 visualizes more clearly the NSL's heterogeneous effects vis-á-vis is the baseline between pro-democracy and pro-establishment Hong Kongers. On the one hand, the NSL substantially reduced the former's trust by large margins (−26%), while induced diametrically opposed effects (17%) on institutional trust for those who leaned toward the establishment on the other.

Table 2.

NSL's effects on institutional trust: ACGT

Robust standard errors in parentheses.

***P<0.01, **P<0.05, *P<0.1.

Figure 1.

NSL's effects on institutional trust by political stances: ACGT.

Moreover, the estimates are also very stable and robust with the addition of different sets of controlsFootnote ¹⁶: (1) model (2): one's probability assessment of how likely the elected opposition LegCo members were going to be disqualified after the September election in 2020Footnote ¹⁷ (DQ_OPM), (2) model (3): one's expectations of Hog Kong's future economy (HKECON) and social welfare (WELEXP), (3) model (4): demographic variables (age, gender, education, residential district, class, and occupation), and (4) model (5): the full specification that includes all the variables above along with one's stance in the first survey (PRO_DEMO_PRE and PRO_ESTAB_PRE).Footnote ¹⁸

What is worth noting here is that the full specification also gives us several additional insights into Hong Kongers' institutional trust independent of the NSL's effect. According to Table 2, one's age is found consistently to be a significant positive predictor of his or her trust level in the Central Government institutions on average (models (2) and (5)). We also find additional correlational patterns that the respondents who trusted the Central Government more tended to be those who had lower expectations of Hong Kong's economic prospect, social welfare provision, and DQ incidents (models (3)–(5)).

The results above help further enrich the literature on how protests and their repression affect public opinions because they not only extend it to the case of Hong Kong, but also paint a more complete picture through the effect heterogeneities our analyses identify. On the one hand, our results about the NSL's effects on one's institutional trust partially echo Sangnier and Zylberberg's (Reference Sangnier and Zylberberg2017) study in Africa in finding the plunge in trust among pro-democracy Hong Kongers during the post-repression (NSL) period. On the other hand, despite through a different theoretical mechanism, the surge of institutional trust among the pro-establishment camp is consistent with Frye and Borisova's (Reference Frye and Borisova2019) finding that the information about unexpected protests can positively update Russian citizens' perceptions of the government trustworthiness.

5.2 Protest expectations

As two recent related studies on Egypt (El-Mallakh, Reference El-Mallakh2020) and Russia (Tertytchnaya and Lankina, Reference Tertytchnaya and Lankina2020) show, people became more inclined to value order and stability after their exposure to protests and repression. In the context of our current study, this implies that one's institutional trust might actually go up if he or she had a stronger belief in the NSL's ability to rein in protest mobilization and dampen the social support for the Movement. In other words, this introduces another layer of effect heterogeneity through which the NSL affected one's institutional trust. The following question in our second survey allowed us to investigate this question and further unpack the relationship between the NSL and one's institutional trust in Hong Kong:

Respondents were then presented with an 11-grade scale on which a larger number denoted a higher probability. A dummy variable of protest expectation, CA, was constructed with responses higher than its average (3.79) designated as low expectations (CA = 0) and otherwise as high ones (CA = 1).

Before reporting the estimation of CA's effects on institutional trust, we show below how CA as a summary measure of one's overall protest expectation does co-vary with his or her assessments for individual protest forms. During the second half of 2019, we had witnessed several of the Hong Kong protesters used to vent, either peacefully or violently, their anger and grievances. Table 3 provides a comprehensive summary (and the variable names we used in the regressions).

Table 3.

Protest repertoire

To see if one's protest expectation measured by CA was correlated with the full gamut of protest repertoire as predicted, we regressed the former on respondents' likelihood assessment (on an 11-grade scale) for each form of the latter adopted by ordinary citizens and protesters, respectively (see Appendix C for more details about how we constructed the variables), and the results illustrated in Figure 2 show exactly that.

Figure 2.

Protest expectations and repertoire.

That is, despite different extents, across all the protest forms and for both citizens and protesters, the less one expected the NSL to rein in the intensity of the Movement, the more likely he or she would expect any one of the protest forms to take place. These results not only make us reassured about the internal validity of CA as a measure of one's overall protest expectation, but also reveal that the protest optimists expected more peaceful protest forms to transpire – i.e., Chanting, Help, and, Yellow.

There are two major findings from our regressional analyses – based on the estimation frameworks (4) and (5) – that highlight our key contributions to the literature. First of all, Figure 3 illustrates the linear predictions of first-differenced institutional trust based on the stance-specific estimates for average Central Government institutions from Table 4 (model (1) baseline). The findings show that, regardless of one's political stance, a greater protest expectation, or a weaker belief in the NSL's stabilizing effect, was associated with a greater decrease in institutional trust. Counterintuitively, even for pro-democracy respondents, while the NSL had substantially reduced their institutional trust, the degree of the reduction however was smaller among the democrats who had stronger beliefs in the NSL's ability to restore social order and lower protest expectations. While this conclusion doesn't carry as much casual significance as the previous analyses on the NSL, it confirms our informational hypothesis that the lower the protest expectation one has, the smaller the decrease in his or her institutional trust will be. It also echoes El-Mallakh's (Reference El-Mallakh2020) and Tertytchnaya and Lankina's (Reference Tertytchnaya and Lankina2020) arguments about the post-protest crave for stability and we show it here that this desire is also positively associated with institutional trust.Footnote ²⁰

Figure 3.

Stance-wise predicted margins by protest expectations: ACGT.

Table 4.

Protest expectations and first-differenced institutional trust: ACGT

Robust standard errors in parentheses.

***P<0.01, **P<0.05, *P<0.1.

Second, while a lower protest expectation is associated with a greater increase in institutional trust for all political stances, the effect is however heterogeneous across them. First of all, using the no-stancers as the baseline, Figure 4 visualizes CA's marginal effects (the coefficients of the interaction terms in Table 4) on the difference between one's pre- and post-NSL average trust for Central Government institutions for pro-democracy and pro-establishment camps, respectively. As it shows, the differences in the increase in institutional trust between no-stancers and pro-establishmentarians are statistically ignorable. In contrast, among pro-democracy respondents, as their protest expectations get lower, the corresponding upward shifts in their institutional trusts are, however, substantially smaller in magnitude than those of other stances. From the informational perspective proposed above, this effect heterogeneity implies that democrats as self-perceived outcasts of authoritarian institutions didn't update their institutional beliefs with the NSL's shock on protesters' mobilization as much as self-perceived regime insiders.Footnote ²¹

Figure 4.

Protest expectations and first-differenced institutional trust: ACGT.

Finally, we also conducted a series of robustness tests for CA (models (2)–(4) in Table 4) – except that no demographic variables were added since they were exactly the same given a very short span between the two surveys. The results above were proven very stable and robust to more controls of respondents' other political as well as economic attitudes, pre-NSL political stances, and CA's original 11-grade scale.Footnote ²²

5.3 Hong Kong's local institutions

As a supplement to the main results above, we investigate in this section whether the NSL and Hong Kongers' protest expectations also affected their trust in Hong Kong's local institutions included in this study – i.e., the Chief Executive (CE), the Court, the LegCo, the Police (POL), the Registration, and Electoral Office (REO). First and foremost, according to Table 5 and Figure 5, although to somewhat different degrees, the NSL's heterogeneous effects among political stances were also present in Hong Kongers' trust in all of them and the polarizing trend between democrats and pro-establishmentarians was especially salient among the local executive institutions. While the former's institutional trust in the Chief Executive and the Police was slashed by the NSL by roughly 20%, it nonetheless went up by a similar magnitude among the latter.

Table 5.

NSL's effects on institutional trust: Hong Kong's local institutions

Robust standard errors in parentheses.

***P<0.01, **P<0.05, *P<0.1.

Figure 5.

NSL's effects on institutional trust by political stances: Hong Kong's local institutions.

In contrast, the trend was less obvious for the local monitoring institutions à la Sangnier and Zylberberg (Reference Sangnier and Zylberberg2017). First of all, while the direction of each effect remained unchanged for both stances, the magnitude of the difference between them shrank by half to 10% for the Registration and Electoral Office. Second, as far as the Court and the LegCo are concerned, the effect sizes were further reduced to digit numbers. In other words, the NSL's effect varied not only among political stances, but also among institutions. As far as the Court is concerned, this contrast is not surprising. Despite the aforementioned controversy about NPCSC's power of final interpretation, Hong Kong's Court is clearly the quintessential element of the common law tradition left from the colonial period and a crucial pillar of the rule of law, which has also become an integral part of Hong Kongers' political identity (Chan and Chan, Reference Chan and Chan2006). Finally, the results presented above are also very robust to different model specifications when we added more controls including respondents' demographics, other political and economic attitudes, and pre-NSL political stances (see Tables D.11–D.14 in Appendix D.3.1 for more details).

As for the effects of one's protest expectation (CA), Figure 6 again visualizes its stance-wise linear predictions of first-differenced institutional trust for the same set of local institutions in Hong Kong based on the regression estimates in Table 6. Just as what we have found for the Central Government trust above, the results are also consistent with our informational hypothesis that, for all political stances, a higher expectation of future protests is associated with a lower level of institutional trust for all the local institutions across the board.Footnote ²³

Figure 6.

Stance-wise predicted margins by protest expectations: Hong Kong's local institutions.

Table 6.

Protest expectations and first-differenced institutional trust: Hong Kong's local institutions

Robust standard errors in parentheses.

***P<0.01, **P<0.05, *P<0.1.

Moreover, we also gained new insights from delving into CA's stance-specific marginal effects. First of all, as Figure 7 demonstrates, we did find the same kind of effect heterogeneity among political stances for Hong Kong's local institutions. Compared to no-stancers and pro-establishmentarians, CA's negative effect on one's institutional trust was dampened among democrats. Moreover, among pro-democracy respondents, there existed another effect heterogeneity between the monitoring – i.e., the Court, the LegCo, and the Registration and Electoral Office – and the executive institutions – i.e., the Chief Executive and the Police. As Figure 7 shows, compared to no-stancers, the (dampening) effect a democrat's protest expectation induces is smaller for the former vis-á-vis is the latter. Based on the same informational perspective, this implies that the CA-related belief for a particular institution to remain relatively unaffected in the post-NSL era was substantially stronger regarding the monitoring ones as opposed to the others, and therefore allowed the former to retain trust when people expected fewer protests to transpire. Finally, these results are also very robust to different model specifications (see Tables D.15–D.17 in Appendix D.3.2 for more details).

Figure 7.

Protest expectations and first-differenced institutional trust: Hong Kong's local institutions.

Combined, the two effect heterogeneities presented above help shed light on Hong Kong's post-NSL political landscape. On the one hand, it echoes Rozenas and Zhukov (Reference Rozenas and Zhukov2019) that the regime support – either institutional trust or loyalty – under dictatorships is closely related to the regime's credibility in suppressing political dissent. Moreover, this effect is dampened when the regime inclusiveness is low. On the other hand, while there was clearly an NSL-induced bifurcation in the trust in the executive institutions between pro-democracy and pro-establishment Hong Kongers, both camps, however, still shared in common similar levels of trust in monitoring institutions.

5.4 Robustness checks

We report more tests in this sectionFootnote ²⁴ to make sure our results are robust not only to different sets of controls, but also to other potential issues that might bias our empirical estimations.

5.4.1 Attrition and data censoring

We first address two sample-related issues. To begin with, despite our low attrition rate of roughly 10%, the fact that the NSL could affect the first-round respondents' decisions to accept or reject our second survey invitation might introduce a selection bias and make the sample nonrandom with respect to the NSL as an external shock. To address this issue, we adopted Lee (Reference Lee2009) bounds estimator that estimates an interval for the true value of the treatment effect in the presence of nonrandom sample selection. As Figure 8 shows, we obtain fairly narrow estimated treatment-effect bounds for the ACGT – −9 to − 24 for pro-democracy and 17–27 for pro-establishment Hong Kongers. While the former implies that our baseline model (model (1) in Table 2) slightly overestimates (in the absolute sense) the magnitude of the NSL effect among democrats, the Lee-bounds estimator still confirms the robustness of the directions and statistical significance of the hypothesized NSL effects across all political stances.Footnote ²⁵

Figure 8.

Treatment-effect bounds: ACGT.

Moreover, Figure 9 presents the estimated Lee bounds of the NSL effects for Hong Kong's local institutions respectively. On the one hand, the majority of the baseline estimates in Table 5 fall within the estimated bounds except those for democrats' institutional trust in the Chief Executive and the Police that slightly exceed the lower bounds.

Figure 9.

Treatment-effect bounds: Hong Kong's local institutions.

In addition, estimating the effects of the NSL on institutional trust critically depends on comparing one's choices over trust levels between the two surveys. Naturally, while they were independently administered, the size of the difference for some respondents might, however, be artificially limited by the choice range provided to them in the second round. For example, for one who already rated a given institution very low (high) during the first round, the fact that the range of values to be chosen was bounded by 0 (100) from below (above) had artificially forced the difference between his or her pre- and post-NSL trust levels to be smaller than it could have been if the lower (upper) bound was absent. In other words, the way in which our questionnaire was designed censored our dependent variable from both above and below, and this could bias our results. To address this second issue in our sample, we re-estimated them by running Tobit models with the cutoffs set at 0 from below and 100 from above.

Figures 10a and 10b show that the effect sizes are substantially magnified for the NSL's negative impacts on pro-democracy camp's ACGT across all model specifications and trust in Hong Kong's local executive institutions – the Chief Executive and the Police – , but the patterns remain the same. In other words, our previous findings about the bifurcated effects the NSL had on pro-democracy and pro-establishment camps' institutional trusts are robust to the data censoring in our design. We also estimated several augmented specifications with additional controls for Hong Kong's local institutions, and the results still hold. See Appendix E.1 for details.

Figure 10.

Coefficients plots for the NSL's effects on institutional trusts by political stances (Tobit estimates).

5.4.2 Endogeneity

Furthermore, since one's political stance choice could be endogenous to his or her institutional trust, we took a machine-learning approach by using the KNN (K-nearest neighbors) algorithm to find a totally exogenous proxy for one's political stance. For each respondent's choice of stance, the algorithm determined which other ones were ‘nearest’ to it in the parameter space composed of the variables of economic and political attitudes in the surveys that were essential to identifying its ‘neighbors,’ i.e., those who held very similar attitudes with the respondent in question along these dimensions. The algorithm then assigned a predicted political stance based on the most frequent choice among the identified nearest neighbors. This approach exploits the fact that individual survey responses in our surveys were independent of each other, and therefore one's newly assigned political stance based on the KNN algorithm would be entirely exogenous to his or her choice of institutional trust.Footnote ²⁶

As Figure 11a shows, while the magnitude of the NSL effects on the ACGT across all specifications among pro-democracy camp is slightly larger, their signs and statistical significance remain the same as the baseline estimations above. For the same set of institutions, the results in Figure 11b also show that our baseline estimates for the effects of protest expectations are robust. Regarding Hong Kong's local institutions, we can find from the estimates documented in Figures 12a and 12b that, despite some differences in magnitude, for both the effects of the NSL and protest expectations, they are very consistent with the baseline results. Combined, all the KNN-based estimations above confirm the robustness of our previous empirical findings to potential endogeneity.Footnote ²⁷

Figure 11.

KNN-based estimations: ACGT.

Figure 12.

KNN-based estimations: Hong Kong's local institutions.

5.4.3 Different cutoffs for CA (protest expectations)

In our third set of robustness tests, we tried to see if our main results for the first-differenced institutional trust in the Central Government (ACGT) remained stable under different cutoff choices for dichotomizing CA.Footnote ²⁸ We adopted two approaches. First of all, since our initial cutoff was 4 (>3.79), we compared its t-statistic with the counterparts under the cutoffs of 3 and 5, respectively, for the baseline model (1) in Table 4 – i.e., the coefficient of the interaction term between protest expectations and the stance dummy of pro-democracy camp. As Figure 13a – the red bar represents the t-statistics based on our initial choice – shows, changing the cutoff to either direction doesn't affect the statistical significance for the other coefficients. This implies that our baseline result is robust to different CA cutoffs.

Figure 13.

Different cutoffs of CA.

Alternatively, we also adopted Deaton and Cartwright's (Reference Deaton and Cartwright2018) and Cantoni et al.'s (Reference Cantoni, Yang, Yuchtman and Zhang2019) approach to randomly assign (fictional) CA groupings to our samples and estimated their effects for 10,000 times. Next, the t-statistics from the estimated effects for all the fictional trials were derived and compared to that under the initial CA cutoff for the same coefficient. This method helps us confirm that the effect we found is statistically very different from those based on random cutoffs and therefore is robust. Figure 13b – the vertical red line indicates the t-statistics from the initial CA assignment – clearly shows that our result is far from random.