Can Elections Motivate Responsiveness in a Single-Party Regime? Experimental Evidence from Vietnam

A growing body of evidence attests that legislators are sometimes responsive to the policy preferences of citizens in single-party regimes, yet debate surrounds the mechanisms driving this relationship. We experimentally test two potential responsiveness mechanisms—elections versus mandates from party leaders—by provisioning delegates to the Vietnamese National Assembly with information on the policy preferences of their constituents and reminding them of either (1) the competitiveness of the upcoming 2021 elections or (2) a central decree that legislative activities should reflect constituents’ preferences. Consistent with existing work, delegates informed of citizens’ preferences are more likely to speak on the parliamentary floor and in closed-session caucuses. Importantly, we find that such responsiveness is entirely driven by election reminders; upward incentive reminders have virtually no effect on behavior.

A growing body of evidence attests that legislators are sometimes responsive to the policy preferences of citizens in single-party regimes, yet debate surrounds the mechanisms driving this relationship. We experimentally test two potential responsiveness mechanisms-elections versus mandates from party leaders-by provisioning delegates to the Vietnamese National Assembly with information on the policy preferences of their constituents and reminding them of either (1) the competitiveness of the upcoming 2021 elections or (2) a central decree that legislative activities should reflect constituents' preferences. Consistent with existing work, delegates informed of citizens' preferences are more likely to speak on the parliamentary floor and in closed-session caucuses. Importantly, we find that such responsiveness is entirely driven by election reminders; upward incentive reminders have virtually no effect on behavior.
A burgeoning literature finds support for the argument that legislators can be responsive to the policy preferences of citizens in singleparty regimes (Chen, Pan, and Xu 2016;Cleary 2007;Distelhorst and Hou 2017;Manion 2016;Todd et al. 2021;Truex 2016). Nevertheless, there is limited understanding of the mechanisms behind this relationship when it is observed. Are legislators acting in response to fears of losing their seats in nondemocratic elections, despite limited competition and concerns about regime manipulation (Gorokhovskaia 2017;Krol 2017;Little 2012;Martinez-Bravo et al. 2017)? Or are legislators responding to top-down mandates from regime elites to consider citizens' preferences in their decision making, thereby limiting threats of collective action from below (Fan and Yang 2019;Meng, Pan, and Yang 2017;Wang and Liu 2020)? Here we directly test these mechanisms using a randomized controlled trial (RCT), which provided national legislators with infographics conveying the preferences of citizens and firms over the Labor Code.
Empirically establishing authoritarian responsiveness requires finding a causal connection between two factors: the articulation of constituency preferences regarding a policy debate, and the behavior of delegates seeking to enact those articulated preferences (Chen, Pan, and Xu 2016;Distelhorst and Hou 2017;Manin, Przeworski, and Stokes 1999). Therefore, we collaborated with the Vietnamese National Assembly (VNA) Library to field an RCT targeting the Labor Code (LC), discussed at the May 2019 session of the 14th VNA. To simulate a signal of constituency preferences, we provided each treated delegate with public opinion data on local preferences over amendments to the LC drawn from nationally representative surveys of citizens and firms. To measure delegate responsiveness, we obtained transcripts from three legislative fora to determine whether delegates spoke, what they said, and whether their speeches reflected the interests expressed on the treatment infographics. Consistent with prior work, we find that delegates were responsive to the citizen treatment but not the firm treatment (Todd et al. 2021).
Building on earlier studies, we presented treated delegates with cover letters designed to prime consideration of either electoral or upward motivations for responsiveness. Treated delegates were reminded of either (1) the competitiveness of the upcoming 2021 elections or (2) their responsibilities under an existing central decree to reflect citizens' preferences in their legislative activities. Delegates receiving a citizen infographic were 10.5 percentage points more likely to be responsive through speech on the VNA floor or in closed caucuses than were those receiving a firm infographic or no information. Importantly, responsiveness to citizens is entirely driven by the election reminder; the central mandate prime has almost no behavioral effects. Delegates receiving both a citizen infographic and election reminder were 15.2 percentage points more likely to use information to speak up in debates (a 37% marginal increase) and 12.7 percentage points more likely to criticize proposed antilabor amendments (56%). Results from hand-coding of speeches and a text-as-data approach reinforce each other in finding strong effects of the citizen-electoral treatment on pro-labor expressions. From a policy responsiveness perspective, delegates in the citizen-electoral treatment were 8.2 percentage points (69%) more likely to have their suggestions reflected in the LC upon passage. These results, together with the null results for other treatment groups, highlight the incentives generated by learning about citizen preferences in the context of an upcoming election.
Deeper analysis probes whether these findings stem from true fears of electoral loss or cadre advancement, in which delegates seek promotions through demonstrated popularity at the polls (Boix and Svolik 2013;Geddes 2006;Salomon 2007). We interact the citizenelectoral treatment with indicators for electoral competitiveness and central nomination status, as delegates could be nominated by central institutions or local electoral commissions. 1 Although we uncover tentative evidence for both interpretations, the cadre advancement story finds greater support: effect sizes on speaking are roughly four to five times larger for delegates from noncompetitive seats (22% versus 4%) and for central nominees (31% versus 8%). This suggests that responsive delegates are less concerned about being voted out of office than they are worried about failing to secure landslide reelections that signal their eligibility for appointments to higher office.
In what follows, we first discuss three prominent existing explanations for authoritarian responsiveness, introduce a new mechanism, and review previous findings that motivated this experiment. Second, we describe the VNA's electoral and reporting institutions, the applicability of the literature's mechanisms to the VNA, and the history of the 2019 amendments to the LC, which serves as our focal legislation. Third, we detail our 2 Â 2 factorial design with a pure control, which informed delegates about constituency demands while priming them to consider electoral or upward accountability channels. Fourth, we present treatment effects on legislative speech making and speech content. Finally, we probe why the citizen-electoral treatment produced such strong effects upon delegate behavior.

MECHANISMS OF RESPONSIVENESS
The primary outcome variable we wish to explain in our research is the responsiveness of elected politicians. Responsiveness must be distinguished from three other closely related concepts in political science: representation, receptiveness, and accountability. Scholars define policy responsiveness as efforts to enact policies "reflecting and giving expression to the will of the people" (Pennock 1952, 790). Later work refined this definition to delineate responsiveness from the broader idea of representation by limiting the concept's reach to political behavior where the legislator's policy choices are taken in response to clear signals of voter preferences expressed through direct contact, surveys, or support for well-articulated party platforms (Manin, Przeworski, and Stokes 1999). Receptiveness-defined as the openness of legislators to receiving signals from voters-is best viewed as a precursor to responsiveness in that it captures an awareness of citizen preferences and a willingness to engage them-both necessary for accurate policy design (Meng, Pan, and Yang 2017). Accountability, in contrast, is conceptually downstream from responsiveness, as it gauges the ability of voters to sanction legislators if they do not take policy actions that reflect their interests (Manin, Przeworski, and Stokes 1999;Tsai 2007). In democratic settings, sanctioning most often takes place through electoral losses. By definition, electoral accountability is restricted-if not impossible-in authoritarian settings, prompting the question of whether responsiveness is even possible without it.
In this paper, we build on a growing body of evidence attesting that bureaucrats and politicians in authoritarian countries may take actions that appear responsive to signals of voters' preferences, such as emails or phone calls (Chen, Pan, and Xu 2016;Cleary 2007;Distelhorst and Hou 2017;Manion 2016;Todd et al. 2021;Truex 2016). Although receptiveness is certainly a first step in the causal chain, we focus on responsiveness because we are interested in delegates' legislative behavior and whether the policies they advocate match the signaled preferences of their constituencies. Because a great deal of the activity analyzed occurs in closed-door sessions invisible to voters, we follow Manin, Przeworski, and Stokes (1999) in assuming that delegates expect responsiveness to be judged by voters' observations of policy outputs as well as their speeches. Although electoral motivations animate our theory and empirical analyses, we lack direct evidence of accountability. Rather than observing delegates' electoral sanctioning, we observe their behavioral responses to reminders about the competitiveness of upcoming elections.
Clear boundary conditions exist around previous findings on authoritarian responsiveness. Politicians and bureaucrats cannot be responsive to every voter preference, especially those that raise existential questions for the authoritarian regime or pose national security threats. Moreover, with a few exceptions (Miller 2015;Truex 2016), most of the research has been conducted on subnational officials, not national politicians. Although Miller (2015) examines crossnational patterns, most work considers only a small set of countries. In addition, the issues that prompt voters to "send" signals are neither sensitive nor embarrassing topics. This matters because citizens and regime leaders are less likely to hold diverging views and demands are less likely to fall outside of the regime's policy window (Kosterina 2017;Schuler 2020a). Despite these scope conditions, the published evidence suggests that providing information regarding citizens' preferences can sometimes motivate authoritarian responsiveness, especially in single-party regimes where opposition parties are prohibited. Todd et al. (2021) reported the first RCT on legislator responsiveness in an authoritarian national assembly, directly testing the causal link between a national legislator's knowledge of constituents' preferences and her consequent legislative behavior. The paper contributed to debates by answering two questions-are authoritarian legislators responsive and, if so, to whom? Delegates were responsive, but only to signals of citizens' preferences-not those of firms. Left unaddressed, however, was the motivation behind authoritarian responsiveness. Identifying the most important mechanism is the goal of this research project. The literature offers three potential motivations: (1) public spiritedness of individual politicians, which cannot be randomized and we therefore omit, (2) downward incentives generated by quasidemocratic elections, and (3) upward incentives from elite leaders seeking legitimacy or public support. We introduce and explore a fourth variant called cadre advancement theory, which we did not anticipate in our preanalysis plan. 2 Under this logic, delegates seek to maximize vote shares not because they fear losing the election but because the regime uses electoral returns to reward politicians with promotion (Boix and Svolik 2013;Geddes 2006). We now discuss the latter three mechanisms in turn.

Electoral Incentives
Today, over four-fifths of the world's countries-both democracies and nondemocracies-hold elections (Egorov and Sonin 2021;Hyde and Marinov 2012). Given the prominence of flawed elections in authoritarian regimes, do they actually shape the behavior of the politicians they select? If limited competition renders electoral defeat possible while also conveying the demands of citizens, authoritarian elections may generate a bottom-up motivation for legislators to take policy actions reflecting citizens' preferences. In this story, incumbents may be more responsive to citizens' policy concerns in order to boost their probability of renomination by party elites or reelection by citizens. Little (2017) demonstrates formally that this may lead to greater policy responsiveness.
However, the literature has not explored whether flawed elections may motivate responsiveness. Some authoritarian politicians claim to be both receptive and responsive to constituents (Manion 2016;Meng, Pan, and Yang 2017). Despite these findings, electoral incentives are presumed absent in authoritarian regimes. At first blush, this makes sense, as flawed elections rarely lead to regime turnover. But short of regime change, it is nevertheless possible that limited competition motivates incumbent parliamentarians to alter their behavior in ways that might appear responsive (Little 2017;Miller 2015).
Research on China demonstrates that elected village leaders are more likely to take actions in line with local demands and oppose central initiatives contradicting local interests, such as the former one-child policy (Martinez-Bravo et al. 2017). Distelhorst and Hou (2017) find that Chinese mayors are equally likely to respond to complaints regardless of collective action potential, suggesting that it is dissatisfaction rather than unrest that they seek to reduce. In Uganda, Grossman and Michelitch (2018) find that politicians in more competitive districts are more responsive to a performance scorecard. These findings lend some credence to the suggestion that, in authoritarian regimes with some intraparty competition, elections may still matter for responsiveness. Although regime survival and party authority is rarely at stake, voters' limited power to choose alternative regime candidates may nonetheless spur responsiveness through possible electoral sanctioning or performance-based promotion incentives.
To be clear, authoritarian leaders erect electoral rules and parameterize candidacy and participation for their own purposes. In line with existing theories of authoritarian elections, regimes create the conditions under which delegates can learn from their voters and apply such knowledge in legislative debates. Thus, a finding that authoritarian delegates are responsive to electoral concerns would not indicate that central authorities have abandoned control of the polity to voters. Rather, it would indicate that authoritarian elections may constitute a more efficient strategy for learning voter preferences and channeling them into policy than do central diktats.

Upward Incentives
In lieu of electoral sanctions or "good" types, upward accountability to central elites may also ensure that policy outcomes eventually conform to the expectations of constituents. In order to foster legitimacy and ward off social unrest, the single-party regime may direct delegates to represent constituents' perspectives (Dickson 2016;Tang 2016;Wang and Liu 2020). Proponents of the upward incentive mechanism argue that single-party regimes actually encourage responsiveness within a limited space on nonthreatening issues (Truex 2016) by creating structures, such as hotlines and mailboxes (Cleary 2007;Distelhorst and Hou 2017), for citizens to request improvements in services and by providing financial resources for delegates to learn about citizens' demands (Wang and Liu 2020).
In the Vietnamese context, where the party delegates executive tasks to a government cabinet led by a prime minister, Schuler (2020b) argues that demanding responsiveness when citizens' views align with regime interests is one way for the party to hold cabinet officials accountable. Incumbent politicians are motivated to be responsive to citizens' interests, not because of any concerns about elections but because regime leaders want them to be. And leaders want legislators to be responsive because they worry that unsatisfied citizens may engage in costly social unrest. Therefore, citizens' demands are pacified, contributing to regime resilience and stability (Guriev and Treisman 2019;Morgenbesser 2020). Chen, Pan, and Xu (2016) test this mechanism in China by showing that mayors are more likely to respond to citizens' complaints when reminded about the potential for collective action.
Importantly, the upward incentives mechanism does not depend in any way on electoral returns. Delegates are rewarded for taking legislative actions that reflect local interests and may therefore mollify citizens. Moreover, under upward incentives, delegates' only principal is the central leadership, which communicates to them an objective. Parliamentarians then target their behavior toward central elites, who directly observe their legislative activities and reward them based on elite interpretations of local needs.

Cadre Advancement
A final potential motivation for responsiveness, known as the cadre advancement theory, combines features of both electoral and upward incentives. Adopting the informational logic for elections sketched above, it suggests that authoritarian elections furnish the regime with information on the performance of subordinate officials. Elections serve as "popularity contests" in which citizens call attention to venal cadres or rising stars (Salomon 2007). The cadre advancement logic acknowledges the existence of dual principals in the legislative calculus. Rewards and punishment are meted out, not directly by citizens, but by regime leaders who read the electoral returns and promote or co-opt the high performers (Boix and Svolik 2013;Gandhi and Przeworski 2007;Geddes 2006). Yet it is citizens at the ballot box who provide the raw material for these promotion decisions.
In line with the upward incentive logic, central officials want delegates to reflect constituents' preferences to enhance regime legitimacy and ward off local unrest. Unlike the upward incentives logic, however, cadre advancement theory sees delegates as incentivized to maximize vote shares rather than tailor policy behavior to earn promotion. As a result, delegates must be highly attuned to the nuances of voter preferences, especially when these deviate from central preferences without directly threatening the regime. At the same time, cadre advancement theory distinguishes itself from pure electoral incentives by emphasizing vote share maximization as a signal of widespread popularity above and beyond mere electoral victory (Martinez i Coma and Morgenbesser 2020). In contrast to the logic of single-member districts with plurality rules, where vote maximization is the norm, any candidate in Vietnam's multimember districts can be elected by earning the third-highest vote share, whatever that may be. Thus, it is cadre advancement theory, not standard electoral logic, that motivates vote maximization.

RESEARCH CONTEXT
The setting for our field experiment is the VNA. Although the role of the body has changed over time, the basic structure of two-to-three-week legislative sessions has not. Twice a year, delegates gather in Hanoi to receive the final legislative docket and then enter group caucuses-composed of combinations of provincial delegations-to prepare for debates on draft laws. Delegates grill cabinet officials on their performance and policy choices in televised query sessions. Floor debates constitute the final stage of VNA sessions. Although individual delegates are prohibited from introducing new legislation, they may propose and vote on amendments during legislative debates. Many floor comments cite citizens' concerns regarding pressing issues such as food safety, health care, and infrastructure (Schuler 2020b). Although floor debates and query sessions are time limited and delegates with relevant expertise are favored for speaking time, caucuses are more freewheeling and place no limits on delegates' contributions.
With each new session, the share of full-time delegates has increased, and research has charted steady increases in educational attainment and professional expertise within the body (Malesky and Schuler 2011;2013;Schuler 2020b). As a result, the quality of legislative debates and query sessions has improved. Increasing professionalism and expertise means new responsibilities for the VNA, including public committee hearings soliciting expert testimony and annual confidence votes (Bui 2018). In fact, a majority of Vietnamese voters views legislative debates as meaningful and influential (Schuler 2020b).
Four main types of delegates serve in the VNA. Ninety-six (19%) of the 496 delegates serve full time in the VNA Standing Committee (NASC) or as Chairs or Deputy Chairs of the VNA's nine committees and were nominated by the center (described in detail below). An additional 25 (5%) central nominees serve part time; these elites comprise the Politburo and Secretariat and were dropped from the experiment due to political sensitivity. A further 65 (13%) locally nominated, full-time delegates maintain provincial offices, provide constituency services, and receive voter petitions. The remaining 310 (63%) delegates are local nominees who serve part time, attending only the biannual plenary sessions while maintaining full-time employment in other fields. 3 Full-time delegates are more likely to have the time and capacity to be 3 In addition to the 25 sensitive elites, we dropped one full-time central nominee who for logistical reasons never received a treatment as planned. Our final sample contains 470 delegates: 95 (20%) fulltime central nominees, 65 (14%) full-time local nominees, and 310 (66%) part-time local nominees. responsive to citizens, and locally nominated, full-time delegates were the most responsive in previous studies (Malesky and Schuler 2010).

Elections
Vietnam's 2013 Constitution stipulates that the VNA must be democratically elected. After a rigorous nomination process to get on the ballot, voters directly elect vetted delegates through quasi-competitive elections to serve five-year terms in 184 multimember districts. Voters in each district choose multiple delegates from a slate with a roughly 2:1 candidate-to-seat ratio (Bui 2018;Schuler 2020b). The median 2016 vote share among all candidates was 60%, increasing to 72% among elected delegates. All candidates' vote shares are published in national and local newspapers (e.g., VnExpress 2016) as well as on the VNA website (Quoc Hoi Viet Nam 2021). This electoral connection raises the possibility of responsiveness, as voters may be able to leverage their ballots to select "good types" ex ante or "vote the bums out" ex post (Fearon 1999;Manion 2014). Although elections are not entirely free and fair, voters do enjoy genuine choice among candidates nominated and vetted by central institutions or local election commissions, and this may suffice to motivate delegate behavior. Scholars of Vietnam have also echoed Geddes (1999) and Boix and Svolik (2013) in claiming that central elites use electoral returns to assess the popularity of local officials for cooptation or promotion (Salomon 2007).
The Central Election Board (CEB) and the NASC delegate substantial authority to Vietnam's 63 provincial officials. First, provincial authorities decide who will compete against whom and where. With each province containing between two and nine districts, provincial authorities wield considerable discretion to place candidates in more or less competitive districts, making the province the appropriate unit for assessing population interests. Second, provincial party leaders nominate and vet most candidates running in the provinces. Although the CEB distributes an outline for the descriptive characteristics each provincial delegation should eventually match, the provinces nominate and vet the specific individuals who run to fill those roles. However, the central leadership does not relinquish all nomination and vetting duties to the provinces. During the run-up to the election, the CEB works closely with the NASC and the Vietnam Fatherland Front to produce a list of vetted candidates. This centralized process occurs in parallel with provincial vetting, wherein provincial election commissions draw up candidate lists (Malesky and Schuler 2009;Schuler 2020b).
Of 869 candidates in the 2016 election, 193 were centrally nominated-put forward by central party, government, or military organizations and vetted by the CEB. An additional 665 were locally nominatednamed by local agencies and vetted by a provincial election board. All local nominees reside in the provinces in which they compete; central nominees may reside anywhere, but they often work as high-ranking officials in Hanoi. Eleven self-nominated candidates survived the vetting process and ran as independents.
Once candidate slates have been determined, central nominees are sent to the provinces and assigned to specific electoral districts to compete with local nominees. After dropping Politburo elites, 95 (20%) of the remaining 470 delegates in our sample are central nominees who only nominally represent the interests of their "provinces"-many having never previously visited. These central nominees are fulltime politicians or high-ranking bureaucrats, often designated for leadership positions within the VNA itself. They are therefore placed in easier-to-win districts with lower candidate-to-seat ratios (Malesky and Schuler 2011). Former Deputy Chairman of the Office of the VNA Nguyen Si Dung (2017) suggests that increasing the share of central nominees would improve professionalism and expertise because local nominees, particularly part-timers, lack preparation and technical knowledge. The 375 (75%) local nominees in our experiment are primarily local officials or other notables residing in the province put forward by provincial election boards and expected to represent local interests.
There is substantial within-party competition. The average vote share for party members was 59%, and fully 48% of party members failed to win the seats they contested. Electoral turnover is also quite high, with a mere 30% incumbency rate among candidates as many part-time, local nominees either opted not to run or failed to gain nominations. When they did run, incumbents in the 2016 election performed exceedingly well: 94% were reelected with an average vote share of 74%. Although the majority of party losses occurred among local nominees, 15 out of 193 central nominees lost and an additional eight barely eked out victories with less than 60% of the vote. 4 Because of media attention, such losses are considered quite embarrassing for the candidates themselves and may explain why two-thirds of Vietnamese voters consider VNA elections competitive (Schuler 2020b). Even winning with a small vote share can prove embarrassing, as Vietnamese newspapers highlight the shares of all elite leaders (VOA Tieng Viet 2016). One official even demanded that the newspaper Tien Phong correct a typo erroneously announcing his winning vote share as 68.32%, placing him behind two other elites, to the more substantial share of 86.32% (VOA Tieng Viet 2016).
To summarize, despite constraints on the nomination and vetting of candidates, the structure of VNA elections may yet motivate delegates to alter their behavior during legislative sessions. We experimentally prime this motivation by reminding randomly assigned delegates about the upcoming 2021 election. However, debate remains about why exactly elections matter. Do delegates care about electoral returns because they value election to the VNA (electoral incentives)? Or do electoral returns signal quality to the central officials who might promote delegates (cadre advancement)? Speculative analysis below leverages variation in electoral competitiveness and nomination status to disentangle these two incentives.

Upward Incentives
A competing mechanism of legislator responsiveness holds that delegates abide by regime directives to reflect the views of their constituents. The VCP's 2011 Political Report specifically called for the VNA to enhance its representation of and connections to voters. Analyzing the motivation for this mandate, Thiem H. Bui, an officer in the VNA's Institute for Legislative Studies (ILS), argues that the VCP views the development of a "socialist, law-based" state as central to regime legitimacy (Bui 2018). As part of the process, the VCP wants the VNA to check the power of the prime minister and cabinet through query sessions and confidence votes (Schuler 2020a). Bui further argues that empowering delegates to speak up furnishes the regime with information about the needs of local constituents, thereby warding off potential disruptions. Related survey experiments have found that priming citizens about the role of the VNA increases their support for the VCP and reduces the likelihood of protests (Schuler 2020b).
To this end, there are organs within the VNA mandated by the party leadership to improve information provision and responsiveness (Nguyen 2017). Article 83 of the 2013 Vietnamese Constitution mandates that all delegates meet with constituents at least twice per year. The VNA Standing Committee, chaired by the Politburo's fourth-ranked member Nguyen Sinh Hung, detailed in Resolution No. 27/2012/QH13 how the VNA was to fulfill its mandate of improving debate and contact with local citizens. 5 Borrowing language from the Political Report and signed by Nguyen, the Resolution therefore bears the imprimatur of the VCP. 6 A related joint resolution of the NASC and the Vietnam Fatherland Front (525/2012) further mandates that delegates attend meet-the-voters sessions before and after the two annual VNA sessions. In practice, therefore, delegates are obligated to meet with constituents four times per year.
Resolution 27 provided each deputy a small budget for staff support and constituency services in all 63 provinces. The Office of the National Assembly was declared responsible for supporting delegates in their legislative activities through its oversight of the VNA Library, much like the Congressional Research Service in the United States. Operating under the purview of the VNA Standing Committee, the ILS was tasked with improving the quality of VNA proceedings through research endeavors. Although all of these institutions predated Resolution 27, the relationship between delegates and constituents is now more clearly delineated.

Labor Code Debate
The focus of legislative debate in our experiment is the revised Labor Code (LC), adopted in 1994 and last amended in 2012. Eyeing participation in the Trans-Pacific Partnership (TPP) and EU-Vietnam Free Trade Agreement, Vietnam adopted several highly pro-labor reforms. However, following the U.S. exit from the TPP, Vietnamese firms pushed back against the prolabor status quo. At the same time, women's organizations were pushing for greater protection of women in the work place. To address these concerns, several amendments were submitted to the seventh session of the 14th VNA for debate.
The proposed amendments were controversial, sparking disagreement between business and labor interests and providing a useful test of authoritarian responsiveness. The controversy is best exemplified by the LC's 50,000 page views on Du Thao Online, the VNA's portal for public comments; the second-most viewed bill recorded less than 10,000 page views. Major stakeholders such as the Vietnam Business Forum, Vietnamese Confederation of Labor, Vietnam Chamber of Commerce and Industry, and International Labour Organization issued position papers debating the costs and benefits of the proposed amendments. The final draft of the LC was endorsed the following legislative session, and its 17 chapters and 221 articles took effect in January 2021.
Our experiment centers on the May 2019 legislative session, where debate over the LC focused on (1) defining "worker" in a digital economy, (2) restrictions on overtime, (3) retirement age, (4) unilateral termination of labor contracts, (5) grassroots organizations for workers' representation, (6) public and New Year holidays, and (7) working hours. These LC revisions were presented in the form of two options for delegates (status quo versus amendment), allowing us to code whether the provision was more pro-labor or more probusiness and therefore capture the policy orientation of VNA speakers. Below, we provide more detail on the LC provisions up for debate and positions held by key actors as they relate to our research design.

RESEARCH DESIGN
The research design for this project builds on existing work to identify the mechanism behind delegate responsiveness: downward electoral incentives or upward attentiveness to VCP elites. Through collaboration with the VNA Library, we determined in 2018 when the LC debate would transpire, which clauses would be debated, and what alternative wordings would be considered. From these, we selected seven clauses we believed best capture salient disagreements between citizens and firms. We then developed survey modules presenting the alternative language for these seven clauses and included them in two nationwide 5 The NASC is a permanent body that controls the activities of the VNA when the latter is not in session. Its chairperson presides over VNA sessions, authenticates VNA output, and liaises with other party-state organs. 6 Additional background on the Resolution is presented in Online Appendix I. public opinion surveys conducted later that year. 7 The Vietnam Provincial Competitiveness Index (PCI) surveys over 10,000 domestic, private firms each year; the Provincial Administration Performance Index (PAPI) surveys 14,000 Vietnamese citizens annually. Both surveys use probability samples from provincially representative sample frames and have earned the respect of Vietnamese policy makers. Additionally, VNA Library officials warned that delegates would place more confidence in well-known instruments than in tailor-made surveys. Each survey instrument obscured which amendment was the proposal and which the status quo, with proposal and clause ordering randomized.
Using the PAPI and PCI survey data, we calculated the average opinions of firms and citizens in each province and placed these on customized infographics, creating one for each constituency-citizenry or firms-in each province. Figure 1 translates a sample pair. Provincial variation in public opinion, as Figure 2 demonstrates, is minimal. The bars examine the relative strengths of the citizen and firm treatments across all 63 provinces, whereas the dashed lines indicate national averages. Even where individual provinces vary, the within-province preference ordering between citizens and firms remains the same.
To construct the treatments, we selected three amendments to use in both firm and citizen infographics and a fourth item concerning female employees unique to each constituency. Item 1 references LC Article 35 concerning contract termination. Although the amended Art. 35 would preserve the employee's right to terminate a contract unilaterally, it would also impose a 30-day (45-day) window for fixed-term (flexible) contracts, whereas the 2012 LC contained no constraints. According to Figure 2, a mere 16% of Vietnamese citizens supported the proposed change, compared with 40% of owners and managers. Item 2 concerns LC Art. 138, Clause 3, which proposed to strengthen an existing protection for female workers who are pregnant, on maternity leave, or raising young children by exempting them from labor discipline, whether by termination, docked wages, reduced bonuses, or favoritism in shift selection. The results indicated that 83% of citizens and 77% of Vietnamese firms supported this change.
Item 3 highlights LC Art. 169, a proposed increase to the retirement age. The amendment would gradually increase the retirement age from 60 to 62 and from 55 to 60 for men and women, respectively. This proposal garnered support from 35% of citizens and 27% of firms. The fourth item differs slightly between the citizen and firm treatments, although both concern women's issues. The citizen treatment references proposals in LC Art. 36-37 directed at strengthening provision of school and daycare facilities and enhancing occupational training for female employees; an overwhelming 97% of citizens nationwide favored this change. The firm treatment cites LC Art. 137, which proposed to enhance maternity protections by allowing pregnant women to avoid strenuous labor and potentially void their employment contracts; 83% of firms favored this change. Although the LC also considered changes to overtime restrictions, labor unions, and additional holidays, we chose not to use these questions in the infographics because of political sensitivity (i.e., collective bargaining in a single-party regime) and confusing language (i.e., overtime adjustments).
To ensure delivery of these infographics to delegates, we collaborated with the Institute of Public Policy Management (IPPM) at National Economics University. The IPPM provided its official seal on cover letters introducing the infographics, an imprimatur that increased the likelihood that delegates would actually read our treatment materials. Because both citizen-and firm-treated delegates received similar infographics from the same institute, any behavioral differences across groups are attributable to constituency itself rather than the uniqueness of the infographics, the identity of the sender, Hawthorne effects from the novel delivery, or potential assumptions on the part of delegates regarding central monitoring. Although IPPM is a prestigious body officially falling under the purview of the Ministry of Education, it is considered neither an official mouthpiece nor politically powerful relative to delegates' typical informational sources. 8 We then randomly assigned the 470 delegates into treatment groups, with approximately one-third receiving a citizen infographic (n = 152), a firm infographic (n = 160), or official VNA research on the revised LC (n = 158). Cover letters and infographics were mailed during the first week of May 2019, with group caucuses 8 In two caucus speeches, a delegate used information from a treatment but could not recall its source. Appendix C in the DVSI presents additional evidence that delegates used treatment infographics in speeches.
held on May 20th. A saturation design ensured continuous variation in the provincial shares of treated delegates, allowing us to estimate potential spillover and reinforcement effects (Malesky and Todd 2022). 9 Figure 3 sketches the four-stage randomization: (1) 63 provinces were assigned saturation dosages, (2) 470 delegates within provinces were assigned to treatment or control, and treated delegates were assigned to receive (3) the citizen or firm treatment and (4) the electoral or upward incentives treatment. The initial assignment of provincial dosages was a block randomization, whereas the latter three randomizations were delegate-level Bernoulli trials. The first set of Bernoulli trials was conducted for every delegate according to provincial dosages, whereas the latter two were equiprobable coin flips conducted only among the treated.

Ethics
Our RCT follows APSA's "Principles and Guidance for Human Subjects Research" and was approved by the principal investigator's institutional review board. Although the VNA Library sanctioned and facilitated our experiment, we did not receive informed consent from individual delegates prior to randomized assignment. Had delegates known they were being observed, they may have altered their behavior. To ensure natural behavior and minimize stress during the legislative session, we obtained a waiver of informed consent and agreed to postexperimentally debrief the VNA. The primary design features were also directed at minimizing potential effects on the policy process. First, all information on the treatment infographics delegates received was sourced from two high-profile and officially sanctioned surveys; we merely aggregate data from the surveys' Vietnamese-language, public-facing websites. As public figures, delegates frequently receive mailed information from governmental and quasi-governmental bodies, and neither they nor their staff likely viewed the treatments as unusual. Second, all treated provinces ultimately received both citizen and firm treatments prior to group caucuses, ensuring that we did not unduly influence the law through selective information revelation.

Outcome Variables
Following legislative adjournment, we collected transcripts from the group caucuses, floor speeches, and query sessions. 10 For each delegate in each forum, we

FIGURE 3. Randomization Scheme
Note: After provinces were assigned to treatment dosages of 25, 50, and 100%, delegates were assigned to treatment or control. Treated delegates were then assigned to either Citizen or Firm and either Electoral or Upward treatments. 9 Online Appendix E finds no such effects. 10 Because the Minister of Labor received only three inquiries, the analysis focuses on the caucuses and floor debate. recorded (1) whether she spoke, (2) whether she criticized a proposed amendment, (3) the pro-labor/profirm tone of her speech, and (4) whether the final law reflects those comments. 11 Because only aggregated vote totals are released, these speeches provide the best indicator of behavioral responsiveness in Vietnam (Malesky and Schuler 2010;Malesky, Schuler, and Tran 2012) and other countries (Noble 2020;van Dijk 2004). Speaking in authoritarian assemblies is intrinsically risky, as superiors and adversaries can use the information against the speaker. Delegates are therefore reluctant to speak without confidence in the information they present. Because treatments are randomly assigned, differences across treatment groups in their willingness to speak provide an excellent indicator of responsiveness.
We measured speech content in two ways. First, two native Vietnamese speakers, uninformed of the research question or treatment assignments, read all 205 speeches and coded each along four factors: the clause(s) cited, support or opposition to the amendments, a pro-labor or pro-firm tone, and whether the statement was reflected in the final Labor Code. 12 Nonspeaking delegates were coded as zero to avoid posttreatment bias caused by nonrandom selection into speaking.
Second, we used a supervised text-as-data approach called Wordscores to place speeches along a 100-point scale from pro-business to pro-labor (Laver, Benoit, and Garry 2003;Lowe 2008). Reference texts, based on postdebate commentaries from three organizations, anchor each pole. The pro-business position (0) was defined by a position paper put forward by the Vietnam Business Forum, an annual meeting of business associations and chambers of commerce (Human Resources Working Group 2020). Because the pro-business text is 14 pages long, we combined two pro-labor commentaries (Fair Labor Association 2020; International Labour Organization 2019), which collectively run to seven pages, to construct the prolabor pole (100).
Wordscores assigns each word a score representing a weighted average of the reference texts, where weights denote the conditional probability that the word appears in each reference text. 13 Pro-business words and phrases highlight labor agreements, labor contracts, discipline, and salary; pro-labor terms focus on issues like worker representation, collective bargaining, discrimination, and women's issues. New texts are each assigned a (pro-labor) score representing a weighted average of the dictionary's word scores, with weights corresponding to the relative frequencies of these words within the new texts.
Because most speeches (especially in caucuses) are not observed by voters, we assumed that delegates believe that voters assess responsiveness based on actual policy change rather than rhetorical actions. Thus, our final outcome variable indicates whether the views expressed in delegates' speeches were reflected in the LC as adopted, which voters should be able to observe directly. 14 Ding (2020) cautions that speaking may be more performative than responsive, yielding limited tangible improvements for citizens. Yet as excerpts (DVSI Appendices C and D) reveal, VNA speeches are technocratic and even boring-not performative. 15 Eighty-seven percent of LC-related speeches occurred in closed-door caucuses, which could not be performative because they are not televised and their transcripts are not publicly available. Responsiveness in this context, then, must be policybased rather than rhetorical. To measure this effect, we constructed a dichotomous variable that takes the value 1 when a delegate's suggestion was ultimately reflected in the LC as passed in November 2019. When multiple delegates made the same suggestion, all were coded the same way. Delegates whose suggestions did not appear in the final version or who did not speak were coded as zeroes.

Priming Mechanisms of Responsiveness
Our main contribution is to test experimentally two possible mechanisms behind responsiveness by priming delegates in the cover letters that accompanied the infographics. Two versions of the cover letter were sent, each varying slightly in its expressed motivation for informing delegates and designed to prime delegates to think about either the electoral or upward incentive mechanism as it applies in Vietnam. Table 1 presents the cover letter prefaces, with the electoral (upward) incentive treatment in the top (bottom) row and the citizen (firm) treatment in the left (right) column. 16 The electoral prime reminded delegates about the upcoming 2021 election and voters' attention to legislative debates. It did not cite particular vote shares, as these are common knowledge among incumbent delegates. The upward prime reminded delegates about their VCP-mandated responsibility to consider the viewpoints of constituents in their deliberations. Its reference to Resolution 27 implicitly reminded delegates of the guidelines specified in the VCP's 2011 Political Report. Cover letter language was adapted to match the citizen or firm treatment to which each treated delegate was assigned.
As Figure 3 illustrates, we employed a 2 Â 2 factorial design with pure control. Ultimately, 158 delegates were assigned to the control group, 80 to receive the citizen infographic and electoral prime (CE), 72 to receive the citizen infographic and upward prime 11 Online Appendix G presents our multiple outcomes adjustment accounting for the correlation structure across outcomes. 12 Appendix D in the DVSI offers examples of subjective coding decisions for 30 speeches over a wide range of debate topics. 13 Appendix E in the DVSI offers further explanation and presents a "page" from the dictionary derived from the reference texts. 14 We did not prespecify this outcome in our preanalysis plan. 15 Full transcripts of all floor speeches are available on the VNA website: http://quochoi.vn/Pages/sitemap.aspx. 16 See Appendix B in the DVSI for Vietnamese-language cover letters.
(CU), 81 to receive the firm infographic and electoral prime (FE), and 79 to receive the firm infographic and upward prime (FU). 17 Because each delegate was assigned based on an independent coin flip, treatment shares vary continuously across provinces. 18 Although previous work argues that business interests are important for political behavior in single-party regimes (Lü and Landry 2014;Meng and Pan 2015;Tsai 2007), Todd et al. (2021) found no significant effect of the firm treatment in their previous RCT, noting that firms' influence may work through channels outside of VNA speeches. On this basis, we do not expect the firm treatment to spur responsiveness, anticipating instead that the theoretical mechanisms will be best adjudicated through the CE and CU treatment groups. If the electoral (upward) mechanism is the primary driver, responsive behavior should be greater in the CE (CU) group than in the CU (CE) group. The FE and FU treatments guard against potential Hawthorne effects resulting from the infographics' novelty.

Empirical Strategy
Our primary estimation strategy is a delegate-level linear probability model, where we regress our outcome variables (spoke, critical, pro-labor, probusiness) on four treatment group dummies (CE i , CU i , FE i , and FU i ), provincial treatment shares (R p ) to account for potential intraprovince spillover effects (from treated to control) and reinforcement effects (from treated to treated), and three pretreatment controls (X i ) used by Todd et al. (2021). The control variables consist of dummies indicating whether a delegate (1) serves full time in the VNA, (2) was nominated by the center, and (3) experienced a competitive election. Consistent with prior work, this last variable takes the value 1 when a delegate's vote share falls below the national median for winning vote shares (72.4%). We run the following specification, where i indexes delegates and p provinces: For hypothesis testing, we follow Todd et al. (2021) in adopting a randomization inference (RI) approach coupled with tabular presentation of regression coefficients, standard errors, and RI-derived p-values. To compute these p-values, we reassigned delegates to As you know, voters are increasingly paying attention to whether their opinions and aspirations are included in the law-making discussions of the National Assembly. Many voters regularly follow these discussions and provide opinions in meetings with the Delegates, as well as in increasingly competitive elections. In order to support the Labor Law (amended) in the next National Assembly session, the Research Group of the Institute of Public Policy and Management of the National Economics University would like to send to the delegates voters' opinions in Province X about the rights of local workers As you know, the business community is increasingly paying attention to whether their opinions and aspirations are included in the law-making discussions of the National Assembly. Many investors and business associations regularly follow these discussions and provide opinions in meetings with the Delegates, as well as in increasingly competitive elections. In order to support the Labor Law (amended) in the next National Assembly session, the Research Group of the Institute of Public Policy and Management of the National Economics University would like to send to the delegates opinions of the business community in Province X about the rights of local workers.
Citizen and Upward: Firm and Upward: treatment and control 1,000 times in precise accordance with the four-stage randomization procedure (see Figure 3), with covariates and outcomes undisturbed. We conducted all analyses on each rerandomized dataset and compared our experimental estimates with these distributions of rerandomized estimates. We thus obtained an answer to the following question: under the sharp null hypothesis of no effects for all delegates, just how unusual are our experimental results? 19 Note that it is possible to reject the sharp null hypothesis without rejecting the null hypothesis for an average effect.

RESULTS
We present our findings in four subsections. First, we demonstrate that our experimental results are consistent with previous work on the VNA, providing a foundation of plausibility for novel findings. Second, we examine speaking rates across treatment groups. Third, and based on subjectively coded speeches, we analyze the influence of the constituency and mechanism treatments on delegates' tendency to criticize proposed amendments and favor labor or business interests. Fourth, we employ Wordscores to evaluate the content of delegates' speeches.

Consistency with Previous Studies
As we examine responsiveness in a particularly highprofile debate, it is important to assuage doubts about the generalizability of our results. Therefore, we replicate the basic findings of previous work on the VNA, both observational and experimental. Malesky and Schuler (2010) found that full-time, local nominees-who staff the VNA provincial offices and serve as the primary contact for local citizenswere roughly three times more likely to speak, on average. Despite our focus on a single debate, we observe a qualitatively similar difference: the probability that a full-time, local nominee spoke on the LC was 0.75, whereas the probability for other delegates was 0.39 (p = 0:0001). Similarly, several papers found that delegates who survived more competitive elections tended to speak more (Malesky and Schuler 2010;Malesky, Schuler, and Tran 2012). We, too, found competitively elected delegates to be more vocal than delegates from safe seats (p = 0:026). We return to the importance of these subgroups in our analysis of heterogeneous treatment effects below. Todd et al. (2021) found that citizen-treated delegates were 11 percentage points more likely to speak than others, a 27% marginal increase over the baseline of 41% (p = 0:07). By contrast, firm-treated delegates were not statistically distinguishable from their peers. As Table 2 shows, aggregating over the mechanism primes to assess the effects of the citizen and firm treatments, we estimate nearly identical effect sizes for the LC. Citizen-treated delegates were 10.5 percentage points more likely to speak (p = 0:092), whereas firm-treated delegates were marginally but statistically insignificantly more likely to speak. By successfully replicating previous findings from the VNA, we have increased confidence in the results that follow, despite inefficiency in the estimation. Figure 4 presents the unadjusted results for any speech made in any legislative forum (caucuses, floor, or query) by treatment groups. The effect of the CE treatment stands substantially above the that of other treatment conditions, with nearly 53% of CE-treated delegates speaking during the LC debates. For comparison, only 41% of control delegates-and no more than 45% of other treated delegates-spoke in any forum. Table 3 tests the robustness of this finding through four specifications. Model 1 presents the unadjusted regression results for our four treatment groups, whereas Model 2 adds the same pretreatment controls used in Todd et al. (2021). As we explore in Online Appendix E, the fact that delegates caucus together prior to floor debates prompts three concerns about violations of the stable unit treatment value assumption: treated delegates may (1) influence the behavior of untreated delegates through spillover, (2) raise the   100%), implying that treatment effects are not nationally representative. Therefore, Model 4 reweights the estimation by the inverse probability of selection, as defined by the dosage bins, to generate a nationally representative Average Treatment Effect (ATE). In our primary specification of Model 3, we estimate an ATE of 15.2 percentage points, which is both substantively meaningful (a 37% marginal increase above control) and statistically significant (p = 0:048). The nationally representative estimate of Model 4 is slightly smaller (13.7 percentage points) but more strongly significant (p = 0:038). No other experimental condition generates effects of this size. Importantly, the ATE for citizen-treated delegates primed about their responsibilities to party elites (CU) is one-third the size at 5.1 percentage points and statistically indistinguishable from zero (p = 0:273). In addition, neither firm-treated group exhibits a statistically significant treatment effect. These results indicate that reception of the citizen infographic is strongly enhanced by an election reminder.

Analysis of Hand-Coded Speeches
In addition to coding whether each delegate spoke, two native Vietnamese speakers also coded the tone of each delegate speech. First, because the amendments garnered varying levels of support from workers and firms, we asked each coder to identify whether speeches criticized any of the selected amendments. Second, each speech was coded as either pro-labor or pro-business. Third, we recorded whether a speech's suggestions were reflected in the LC as later adopted. Figure 5 presents the unadjusted averages across treatment groups for these subjectively coded outcomes. Although treated delegates were, on average, more critical of amendments than control delegates, the effect of the CE treatment appears substantially larger, as more than one in three signaled criticism. In addition, the CE group expressed slightly more pro-labor views and found considerably more of their suggestions (over 19%) reflected in the final law. Table 4 provides two models for each outcome. The first is covariate-adjusted and the second adds realized treatment saturation, finding that CE-treated delegates were 12.7 percentage points more likely to criticize the amendments than were control delegates, a 55% marginal increase (p = 0:064 ). Although not statistically significant in the fully specified model, they were also moderately more pro-labor than were control delegates. These results are consistent with the idea that delegates are primed by the CE treatment to reflect citizens' interests in their speeches. Column 6 also demonstrates that CE-treated delegates were

Reflected in Law
Note: Examination of unadjusted outcomes from subjective coding reveals tentative effects of the CE treatment.
8.2 percentage points more likely to have their comments reflected in changes to the law, a 69% marginal increase over control delegates (p = 0:060). This finding is critical because it is a clear indication of policy responsiveness in that citizens' preferences had direct effects on the resulting legislation. Although most of the decision making and drafting took place behind closed doors, there is little doubt that Vietnamese citizens will notice the results, particularly the greater protections and support for women.

Speech Analysis with Wordscores
Wordscores presents another way of gauging the tone of delegates' speeches by placing each on a scale from 0 (pro-business) to 100 (pro-labor). Table 5 regresses the pro-labor score for each speech upon treatment dummies, the usual controls, and provincial treatment shares. However, because coding occurs at the speech rather than the delegate level, three methodological adjustments are necessary. First, when a delegate spoke in both a caucus and on the floor, we averaged the prolabor sentiment across speeches. Second, the length and verbosity of delegates' speeches varied greatly.
To account for these differences, we weighted the regressions by the share of words scored for each delegate. Finally, Wordscores estimates are only available for delegates who spoke, yet subsetting the data to exclude nonspeakers only invites posttreatment bias. Therefore, we present in Columns 3-4 a second set of regressions in which all nonspeakers have imputed scores of 50, substantively implying neutrality between labor and firms. When we impute nonspeakers, the effect of the CE treatment is both substantively meaningful and statistically significant: CE-treated delegates made more pro-labor speeches. This is consistent with the more subjective analyses of hand-coded speeches discussed above and indicates that delegates were motivated to act when informed about citizen preferences and reminded of the upcoming election. Also consistent with prior results, the effects of the firm treatment are statistically insignificant, regardless of prime. However, one difference is that CU-treated delegates exhibited effects on par with those of their CE-treated peers. This is puzzling, for it indicates that CU-treated delegates used terminology similar to those in the CE treatment, yet they did not express pro-labor sentiments in a manner obvious to our coders.

Multiple Outcomes Adjustment
In this section, we address the heightened possibility of Type I errors given repeated hypothesis testing. To account for the correlation structure in our outcomes and in recognition that our p-values were constructed using randomization inference, we use our rerandomized treatment assignments to calculate the probability of obtaining a given number of significant effects under the null hypothesis of no effects for all delegates and all outcomes. 21 Specifically, we analyzed five outcomes-delegate spoke, delegate's speech was critical, delegate's speech was pro-labor, delegate's speech was reflected in law, and delegate's speech's pro-labor Wordscore-across two specifications-with and without the realized saturation shares (e.g., Table 3, Models 2 and 3). Setting alpha to 0.05 and examining our RI-based p-values, six of the 10 results are significant: spoke, critical, pro-labor, reflected, spoke-saturation, and Wordscores-saturation. We then set out to determine how likely we were to find six significant effects. For each of our 1,000 rerandomized treatment assignments, we then counted how many of these 10 ATEs were significant at the 0.05 level. From this, we calculated the probability of finding z or more significant effects and identified the smallest z such that Pr(# significant effects ≥ z|z, α = 0:05Þ < 0:05 . In our data, z = 4. In other words, at that alpha level, we were statistically unlikely (p < 0.05) to encounter four or more significant effects under the null hypothesis. At the same time, that means it was plausible (p ≥ 0.05) to find a maximum of three significant effects, all of which were false positives by virtue of the rerandomization. Subtracting the three false positives from the six significant ATEs yields a minimum bound on true positives; the three effects with the lowest RI p-values-whether delegates spoke, spoke critically, and saw speech reflected in law-are indeed significant at the 0.05 level.
Relaxing our alpha level to 0.10 and reexamining our RI p-values, another two results attain significance, for a total of eight out of 10. Taking the same approach, we find that z = 4, four or more significant results would be implausible, and a maximum of three significant effects would be plausible. Subtracting these three false positives from the eight significant ATEs yields five as the minimum bound on true positives. Sorting the results by p-value, we identify critical, reflected, spoke, Wordscores saturation, and pro-labor as true positives that survive our multiple outcomes adjustment.
By repeating this procedure under a fine grid of alphas from 0.005 to 0.99, we are able to revise the p-values upward for each of the 10 results in a way that accounts for the correlation structure across outcomes. These adjusted p-values are presented in Table 6.

DISENTANGLING THE EFFECTS OF THE ELECTORAL PRIME
The clear conclusion from the preceding analyses is that delegates were highly responsive to citizens' preferences and electoral priming but less responsive to the other treatments. Yet why would elections be consequential in a nondemocratic regime? The authoritarian institutions literature points toward two explanations. First, quasi-democratic elections matter and delegates do worry about defeat, however manipulated the elections may be (electoral motivation). Second, delegates are concerned about promotion and seek landslide victories to demonstrate their popularity and to angle for rewards or co-optation (cadre advancement).
To disentangle these explanations, we analyze heterogeneous treatment effects on the basis of two preregistered controls: competitive elections and central nomination status. 22 Recall that previous work has shown that VNA delegates surviving close elections are more likely to speak and be critical (Malesky and Schuler 2010) and that central nominees benefit from preelectoral engineering that guarantees them lower candidate-to-seat ratios and limited competition. Although this increases their probability of election, it also makes electoral loss highly embarrassing (Malesky and Schuler 2011;. Central nominees expect to inherit leadership positions and tend to demonstrate loyalty by speaking less often and less critically (Malesky and Schuler 2013).
To explore whether the CE effect is conditioned by the competitiveness of a delegate's election, we revise Equation 1 by interacting the CE treatment with the competitive election dummy that previously entered additively. 23 The resulting specification, Equation 2, is similar to the approach used by Grossman and Michelitch (2018) to evaluate scorecards in Uganda. To maximize statistical power, we only interact competitiveness with the CE treatment and simply control for the other treatment groups, treating the combined set of other treatment groups as the counterfactual comparison group. If the CE effect is driven by concerns about electoral motivation, we expect the marginal effect for competitive seats to be larger than that for safe seats, indicating that competitiveness strengthens the salience of the CE treatment for delegates. If the cadre advancement hypothesis is correct, however, we expect the marginal effect for safe seats to be larger than that for competitive seats, indicating that the CE treatment was most effective among safe-seat delegates seeking landslides to signal their suitability for promotion.
(2) Table 7 displays the predicted and marginal effects of the CE treatment on all six of the outcome variables analyzed above. The left panel displays the predicted effects for each of the four groups implied by the 2 Â 2 multiplicative interaction. Group A delegates, for example, easily won election and did not receive the CE treatment; their predicted probability of speaking in any forum is approximately 35%. The RI p-value (p = 0:990) indicates that among 1,000 permutations, only five generated a lower predicted probability of speaking. Group B delegates are similar by virtue of their safe seats but differ from Group A delegates due to receipt of the CE treatment. Their predicted probability of speaking, 57%, is highly significant (p = 0:012). In the right panel, we subtract A from B to calculate the marginal effect of the CE treatment in safe seats: 21.8 percentage points (p = 0:014). For delegates from competitive seats, in contrast, the marginal effect of the CE treatment is 4.0 percentage points (p = 0:323). These results tentatively suggest that cadre advancement is the more likely motivation.
Looking across the other outcome variables, we observe similar patterns, with the ATEs for safe-seat delegates frequently sizable and substantially larger than zero. Specifically, these delegates are 16.5 percentage points more critical, 14.2 percentage points more prolabor, 14.6 percentage points more likely to have their comments reflected in law, and 23.3 points more prolabor according to Wordscores. By contrast, the effects of the CE treatment among competitive-seat delegates are never statistically distinguishable from zero. 24 However, we should be cautious about jumping too quickly to a conclusion regarding cadre advancement Online Appendix D analyzes subgroups recommended by readers and reviewers but finds no significant heterogeneous treatment effects. 23 We prespecified the interactive specification but not the cadre advancement hypothesis. 24 Online Appendix H verifies that this result does not hinge on the choice of a safe-seat threshold at 72.4%. We observe substantively similar results for cutoffs ranging from 50% to 80%, and statistically significant results for thresholds above 64%. because a competitive seat has a strongly positive effect of its own. Consider first that the predicted probability of speaking for non-CE-treated delegates is 47.9% for those in competitive seats yet only 34.8% for those in noncompetitive seats-a 13.1-percentage-point marginal difference that is statistically significant at the 0.01 level. This implies that delegates from competitive seats were already keenly aware of the upcoming election and therefore the CE treatment may have been superfluous. Therefore, the smaller effect of the CE treatment in competitive seats could also be interpreted as an indication of diminishing marginal returns to electoral priming. This pattern is also generally replicated across outcomes.

CONCLUSION
This paper advances the literature on authoritarian regimes by experimentally inducing legislator responsiveness and simultaneously testing two competing theoretical mechanisms behind that responsiveness: electoral motivations and upward attentiveness to regime elites. Examining recent debates on the Labor Code amendments in the Vietnamese National Assembly, we find that delegates are responsive to the preferences of their citizens, and all the more so when they are reminded of the upcoming 2021 election. This responsiveness takes the form of speeches made in closed-door caucuses and televised floor debates, in their adoption of language from our treatment infographics, and in the pro-labor tone of many comments. Additional investigation reveals that the citizenelectoral treatment is more effective in safe seats and among central nominees, suggesting that delegate responsiveness is rooted not in a fear of electoral sanctions but in a desire for promotion and a need to demonstrate popularity at the polls in order to secure it. However, we also caution that local nominees and delegates in competitive seats were already more active than were their parliamentary peers, even in the absence of the electoral prime. Thus, the experiment may also suggest diminishing returns to reminders about electoral competitiveness. Further work will be necessary to disentangle these nuanced interpretations. These findings contribute to the literature on authoritarian responsiveness and fill a major gap in the authoritarian institutions literature by showing how electoral incentives can shape behavior in authoritarian parliaments. But what do our results say about electoral motivations in single-party settings? Does our finding that, under the right conditions, VNA delegates may be responsive to the demands of citizens before an election indicate that the Vietnamese regime is accurately characterized as possessing democratic features? We think such a conclusion remains premature.
For those attempting to build from our discoveries, it is important to understand that our experiment was gold plated. Just as a vacuum chamber generates conditions unobtainable in the natural world so as to enable physicists to test their hypotheses, so our experiment was precision engineered to engender responsiveness in ways that may seem artificial. Specifically, cooperation with the VNA Library was essential to overcoming the opacity and instability that typically plague the legislative docket and stymie organized efforts to influence policy prior to debate. In addition, we expended considerable time and money on treatment delivery for the simple reason that no comprehensive public registry of delegate contact information exists. Finally, even if ordinary citizens managed to contact the members of their provincial VNA delegation, they would likely face a far colder reception than did polished research from reputable surveys and a quasi-governmental body such as the IPPM.
To combat these obstacles, we put substantial effort into our experimental design. We raise them here because transparency about these obstacles is important for understanding the scope conditions of our findings. More optimistically, our research indicates that removing barriers to information on citizen preferences and electoral competition would enhance policy responsiveness in the VNA. Thus our experiment points toward clear policy solutions for central authorities interested in stimulating responsiveness.

SUPPLEMENTARY MATERIALS
To view supplementary material for this article, please visit http://doi.org/10.1017/S0003055422000879.