1. Introduction
This paper focuses on people's attitudes towards democracy and non-democracy in Myanmar and asks whether the extent to which citizens dislike the authoritarian regime is moderated by the severity of the coronavirus disease-2019 (COVID-19) pandemic. Earlier research has investigated the relationship between regime types and the government's response to the COVID-19 pandemic. Authoritarian regimes can take more forcible and prompt actions, but they typically lack incentives to respond to citizens' needs (Stasavage, Reference Stasavage2020). If the authoritarian regime's ability to take swift and decisive actions is recognized as an advantage, citizens' opposition to the regime may be lower. We explore this hypothesis by conducting a survey experiment in Myanmar.
The COVID-19 pandemic quickly spread throughout the world. However, the damage to different populations varied, and regime type may be an important explanatory factor. Some show that the damage – such as the number of cases and death rates – has been more severe in democratic countries (Yao et al., Reference Yao, Li, Wan, Howard, Bailey and Graff2022), whereas others suggest the possibility that authoritarian regimes do not provide data or underreport the extent of the harm (Annaka, Reference Annaka2021). Responses to the pandemic also varied. For example, Engler et al. (Reference Engler, Brunner, Loviat, Abou-Chadi, Leemann, Glaser and Kübler2021) find that within electoral democracies, countries with higher-quality democracies were less reluctant to impose restrictions on individual rights in the first wave of the pandemic.
In addition to the questions on cross-country variations in the responses and their effectiveness, scholars examine the link between the pandemic and political attitudes (e.g. Herrera et al., Reference Herrera, Ordoñez, Konradt and Trebesch2020; Hartman et al., Reference Hartman, Stocks, McKay, Gibson-Miller, Levita, Martinez, Mason, McBride, Murphy, Shevlin, Bennett, Hyland, Karatzias, Vallières and Bentall2021; Kritzinger et al., Reference Kritzinger, Foucault, Lachat, Partheymüller, Plescia and Brouard2021; Filsinger and Freitag, Reference Filsinger and Freitag2022). However, to the best of our knowledge, research is scarce on attitudes towards democracy and authoritarian regimes. We contribute to the body of research by conducting a survey experiment in Myanmar, a country that experienced a profound political transition in February 2021, approximately 1 year after the onset of the pandemic. Specifically, we examine attitudes towards democracy and authoritarian regimes – and how they are affected by the seriousness of the threat from COVID-19 – among the people of Myanmar, who experienced both the elected government and military rule during the pandemic.
To examine attitudes towards regime types, we conducted a vignette experiment in June 2022 as part of an online survey of 756 Myanmar people. Specifically, we presented a hypothetical situation of Myanmar in May 2023, approximately 1 year after the survey was conducted, and asked the respondents how favourable they felt about the hypothetical situation. We used a two-by-two design where we varied the seriousness of the threat from COVID-19 (many people are suffering; not so many people are suffering) and the regime (the current non-elected government continues to run the country; democratic elections have been restored). Thus, there were four vignettes, one of which was presented to the respondent. As will be explained in detail in the subsequent section, we test the hypothesis that when the COVID-19 situation is bad, the extent to which people support democracy is weaker than when the COVID-19 situation is good.
In addition, as a supplementary analysis, we report the results of a conjoint experiment that was part of the same survey. As some respondents might perceive the question about democracy as sensitive, they may hesitate to reveal their truthful opinions. In the conjoint experiment, we provided a pair of hypothetical situations of Myanmar society in 2023. One attribute was the political condition – whether the election had been restored, or the current government remained in power without holding an election. One of the other attributes was how bad the COVID-19 situation would be (bad, moderate, good). We examine whether the average marginal component effect (AMCE) of the attribute of democracy varies across the levels of the attribute of the COVID-19 situation.
As background for the experiment context, Myanmar held a general election in November 2020 that led to a landslide victory for the National League for Democracy (NLD). It was the third election following the reinstallation of elections in 2010, and the Union Solidarity and Development Party – the former ruling party that controlled the legislative majority and elected the president from 2010 to 2015 – experienced defeat. Subsequently, the military claimed that there were numerous cases of election irregularities; a request for investigation and recounting of votes was made by the military, but the NLD-led government declined it. In February 2021, the military declared a state of emergency, nullified the election result, and took power. Thus, an authoritarian regime returned in the middle of the COVID-19 pandemic, and people's memory of both the elected government and the military regime was presumably still new when the survey was conducted (June 2022). Consequently, their attitudes towards democracy and authoritarian regimes in the hypothetical situation asked in the experiment would be based on their recent experiences living under both regimes.
The rest of the paper is organized in the following way. Section 2 explains our hypothesis. Section 3 describes the experimental design, followed by the presentation of the findings in Section 4. Section 5 reports the results of the conjoint experiment as a supplementary analysis. Section 6 discusses the results and their implications.
2. Hypothesis
2.1 Pandemic and regime types
The COVID-19 pandemic quickly spread worldwide, and countries confronted this public health challenge in a variety of ways. Considering the rapid spread of the virus, the pandemic provides an opportunity to investigate variations in government responses and health outcomes. Earlier studies point to better outcomes in East Asian countries, indicating certain cultural practices – such as wearing face masks and behaviours useful for protecting the population from viruses – as a source of success (Matuschek et al., Reference Matuschek, Moll, Fangerau, Fischer, Zänker, Van Griensven and Schneider2020; Yamamoto and Bauer, Reference Yamamoto and Bauer2020; Hyun et al., Reference Hyun, Setoguchi and Brazelton2022).Footnote 1
Others focus on institutional characteristics such as federalism and regime types in explaining variations in government responses and health-related outcomes (Cheibub et al., Reference Cheibub, Hong and Przeworski2020; Frey et al., Reference Frey, Chen and Presidente2020; Gordon et al., Reference Gordon, Huberfeld and Jones2020; Huberfeld et al., Reference Huberfeld, Gordon and Jones2020; Kettl, Reference Kettl, Cockerham and Cockerham2020; Stasavage, Reference Stasavage2020; Bennouna et al., Reference Bennouna, Giraudy, Moncada, Rios, Snyder and Testa2021; Choutagunta et al., Reference Choutagunta, Manish and Rajagopalan2021; Engler et al., Reference Engler, Brunner, Loviat, Abou-Chadi, Leemann, Glaser and Kübler2021; Hegele and Schnabel, Reference Hegele and Schnabel2021; Yao et al., Reference Yao, Li, Wan, Howard, Bailey and Graff2022). For example, centralized coordination and response would be difficult in the USA due to its federal and decentralized intergovernmental system, leading to a wide variation across states in the response to and damages from the pandemic (e.g. Gordon et al., Reference Gordon, Huberfeld and Jones2020).
Regime types could also matter. According to Stasavage (Reference Stasavage2020), authoritarian regimes can take swift and decisive actions, bypassing steps that would have been necessary in democracies. However, they lack incentives to respond to public needs because there is no electoral consequence for ignoring them; as a result, despite the possibility of robust actions, authoritarian governments do not always respond in ways that best protect the population.Footnote 2 Additionally, an authoritarian regime's tendency to hide information makes it difficult to accurately understand the conditions on the ground, which may hinder effective measures by government agencies or broader society.
Empirical research has been conducted with mixed results. Some find that the pandemic outcome – such as the cumulative numbers of infections and deaths – is worse in a democracy than in a non-democracy. Karabulut et al. (Reference Karabulut, Zimmermann, Bilgin and Doker2021) find a positive association between countries' level of democracy and infection rate. Yao et al. (Reference Yao, Li, Wan, Howard, Bailey and Graff2022) find that, on average, the difference in the outcomes between the two groups is insignificant; however, when analysing relatively wealthy countries alone, democracies perform worse than authoritarian regimes, with the former having a higher fatality rate than the latter. The authoritarian regime's capacity to take forcible actions is assumed to be a key factor contributing to its success relative to democracy (e.g. Yao et al., Reference Yao, Li, Wan, Howard, Bailey and Graff2022: 8702). On the contrary, Annaka (Reference Annaka2021) demonstrates that critical data are either missing or unreliable for some authoritarian regimes. This view – the ‘biasing autocracy’ hypothesis (Cassan and Van Steenvoort, Reference Cassan and Van Steenvoort2021: 2) – implies that the negative association between the level of democracy and health-related performance during the pandemic is driven by a lack of accurate data.
Studies have also been carried out on political attitudes during a public health crisis. Using survey data from six European countries, Filsinger and Freitag (Reference Filsinger and Freitag2022) find that those with a greater sense of fear of the pandemic tend to support authoritarian attitudes, indicating ‘a desire for collective security at the expense of individual liberty (4)’. Similarly, Hartman et al. (Reference Hartman, Stocks, McKay, Gibson-Miller, Levita, Martinez, Mason, McBride, Murphy, Shevlin, Bennett, Hyland, Karatzias, Vallières and Bentall2021) reveal that survey respondents in the UK and Ireland with a greater sense of anxiety express a higher level of support for authoritarianism, nationalism, and anti-immigrant sentiment. Some studies also show that the damage from COVID-19 is associated with a decline in support for the incumbent (Baccini et al., Reference Baccini, Brodeur and Weymouth2021; Mendoza Aviña and Sevi, Reference Mendoza Aviña and Sevi2021).
Others examine whether crises affect the popularity of leaders, a topic of interest even before the current pandemic. In international crises such as wars, a leader's popularity tends to increase in the short run, which is referred to as the rally around the flag effect (e.g. Mueller, Reference Mueller1970; Oneal and Bryan, Reference Oneal and Bryan1995). For example, US President George W. Bush's popularity increased by 40 percentage points after the 11th September attacks in 2001 (Lambert et al., Reference Lambert, Schott and Scherer2011).
Other types of crises may impact leaders' popularity, too. According to the theory of retrospective voting, if voters perceive that the government is performing well, they will continue to support incumbent leaders, whereas bad performance reduces support, possibly leading to electoral sanction (e.g. Fiorina, Reference Fiorina1981). More specifically, if voters perceive the government is not handling the crisis properly when a severe natural disaster occurs, they will likely blame the incumbent (Achen and Bartels, Reference Achen and Bartels2004). For example, following the earthquake, tsunami, and nuclear incident in 2011, the incumbent government and prime minister in Japan were widely criticized for mishandling the crisis, although some question whether the poor handling of the crisis is solely attributable to the incumbent (Kushida, Reference Kushida2014).Footnote 3
Economic crises can also result in a decline in the popularity of the leader. In assessing incumbent performance, many voters use the economic situation – either their own (pocketbook) or that of the society (sociotropic): they support the incumbent when the economic condition is good and stop supporting the incumbent in leaner years. Therefore, economic crises are likely to reduce support for the incumbent and cause their electoral defeat. For example, in Europe, ruling parties suffered electorally following the Great Recession, particularly in countries such as Ireland, where elections were held shortly after the financial crisis began (Magalhães, Reference Magalhães2014).
The COVID-19 pandemic is a public health crisis, and in its early phase, many observe the rally around the flag effect (Herrera et al., Reference Herrera, Ordoñez, Konradt and Trebesch2020; Yam et al., Reference Yam, Jackson, Barnes, Lau, Qin and Lee2020; Kritzinger et al., Reference Kritzinger, Foucault, Lachat, Partheymüller, Plescia and Brouard2021). Using data from 11 countries and regions, Yam et al. (Reference Yam, Jackson, Barnes, Lau, Qin and Lee2020) show that the approval ratings of incumbent leaders increased alongside an uptick in COVID-19 cases. Herrera et al. (Reference Herrera, Ordoñez, Konradt and Trebesch2020) point out that the rally around the flag effect has occurred in many countries, where the approval ratings of incumbent leaders increased in the short run after the onset of the pandemic. However, the high level of popularity lasted beyond several weeks only in countries where the government managed the pandemic well.Footnote 4 Similarly, Thies and Yanai (Reference Thies, Yanai, Pekkanen, Reed and Smith2022) find that the popularity of the prime minister in Japan remained low while COVID-19 cases were high. Thus, consistent with the theory of retrospective voting, if voters perceive that the government is not handling the crisis adequately, the incumbent seems to lose support.
Despite earlier studies, little research has been carried out on attitudes towards democracy and authoritarian regimes during the pandemic, particularly focusing on those who live under authoritarian regimes. Given the review of research from broader comparative perspectives, we now proceed to the country-specific discussion and the presentation of our hypothesis.
2.2 Hypothesis
Our hypothesis builds upon an assumption that people strongly prefer democracy, and we focus on the case of Myanmar. The lower house of Myanmar (Pyithu Hluttaw) has 440 seats, 330 of which are elected by the people – the other 110 are reserved for the military. The NLD, which assumed power after its victory in the 2015 election, experienced another landslide victory in 2020. It won 258 out of the 330 constituencies in the lower house. Even with the 110 seats reserved for the military, the party retained a legislative majority. The electoral college – comprised of popularly elected parliamentary members in the upper and lower houses and parliamentary members appointed by the military – elected the president from the NLD (Htin Kyaw) as well.
Although voting data were not accessible, the election results show that the level of support for the party has been high in the recent past. On 1 February 2021, the military seized power by declaring a state of emergency after the NLD government rejected its accusation that there were incidences of election fraud. A large number of people protested throughout the country following the coup. Given the most recent election results and anecdotal evidence following the transition in February 2021, it is improbable that the survey respondents substantially support the authoritarian regime.
Despite people's likely resentment of military rule, however, it is possible that the extraordinary public health crisis – when prompt and forcible actions may be useful – moderates it. In democratic societies, such decisions by the government may conflict with democratic principles such as freedom of association and individual liberty. In addition, leaders of more democratic countries would feel more concerned than those of less democratic countries about post-crisis criticisms of their forcible actions; thus, greater post-crisis accountability likely discourages them from taking more decisive actions (Engler et al., Reference Engler, Brunner, Loviat, Abou-Chadi, Leemann, Glaser and Kübler2021). If people perceive that forcible actions are needed but the democratic government has limitations in taking such actions, their animosity against dictatorships may decrease.Footnote 5 Thus, while acknowledging the likely presence of strong support for democracy, we test the following hypothesis: The extent to which people prefer democracy is expected to be lower when the COVID-19 situation is bad than when it is good.
We also stress that the rally around the flag effect is unlikely for the military government because it assumed power almost 1 year after the onset of the pandemic. The short-run surge in popularity documented in the literature usually lasts several weeks to a few months after a crisis hits a country. If the rally around the flag effect had been observed, it would have been for the NLD government. Thus, to the extent that people's memory of the NLD's handling of the pandemic is influenced by its incumbency in the early phase, it would increase people's support for democracy during the crisis. In other words, it would make it more difficult for our hypothesis to be supported.
3. Materials and methods
We test our hypothesis using a survey of 756 citizens in Myanmar conducted in June 2022.Footnote 6 The survey was run online through a survey firm in Yangon: the firm recruited participants by contacting people registered with the firm as potential respondents. We requested that the survey company recruit respondents to (1) have a balance between male and female respondents, (2) have respondents from all age groups, and (3) have respondents from both urban and rural areas. Approximately 4,500 individuals were contacted online (mainly by messaging applications), out of whom 756 responded. Therefore, the response rate was 16.8%. Given the sensitive nature of the topic, some declined to participate, leading to the relatively low response rate. In Table 1, we report the characteristics of our respondents, including age, gender, and other variables asked in the survey.Footnote 7
Characteristics of the respondents
Note: The following are the descriptions of the variables. Age: respondents were asked to report their age groups. Gender: respondents were asked to select male or female. Urban: respondents were asked to report whether they live in an urban area or not. Ethnicity: respondents were asked to select their ethnicity; those who are non-Burmese are categorized as ‘Other’. Education: respondents were asked to select their educational attainment.
Note that probability sampling was not feasible because there is no readily accessible sampling frame, e.g. a voter registry. Alternatively, one might think of a multistage sampling approach in which the country is divided into sampling units such as census tracts, and a stratified random sample of the units is selected; then, within each unit sampled, the list of households could be created prior to carrying out random sampling. However, this is extremely costly; furthermore, given the political instability and violence, it is not safe to visit households in various parts of the country.
The survey contained 40 closed-ended questions in the language of Myanmar (Burmese): the first section asked about respondents' health-protecting behaviours during the pandemic, such as washing hands and ventilation; the second section asked for their attitudes towards the government's responses to this public health crisis; the third section was about respondents' satisfaction with public services in general; the fourth section included a vignette experiment, and a conjoint experiment was presented in the fifth section. The last section asked several questions about respondents' characteristics, such as ethnicity and educational attainment.Footnote 8
We test the hypothesis using a vignette experiment. A vignette presented the hypothetical condition of Myanmar society in May 2023, approximately 1 year after the survey was conducted. We employed a two-by-two design in which we varied the severity of the COVID-19 situation (good, bad) and whether elections had been restored (restored, not). One of the four scenarios was randomly presented to each respondent; then, they were asked to rate the favourability of the scenario by responding to the following question: ‘How would you rate this hypothetical condition?’ Answers were presented in a five-point ordinal scale, including favourable, somewhat favourable, neither favourable nor unfavourable, somewhat unfavourable, and unfavourable. Table 2 summarizes the experimental design. Appendix A checks the balance across experimental groups.
Experimental design
Note: Respondents were randomly assigned to one of the four experimental groups. After the vignette was presented, the following question was asked: ‘How would you rate this hypothetical condition?’ The answer choices include: favourable (5), somewhat favourable (4), neither favourable nor unfavourable (3), somewhat unfavourable (2), and unfavourable (1). The ordinal variable was constructed with 5 indicating favourable and 1 unfavourable.
In terms of the COVID-19 situation in Myanmar, the country had difficulties handling the crisis even before the military took power in February 2021. During the first few months after the outbreak (May 2020), the number of infections and casualties reached 142,000 and 3,200, respectively, according to data from the World Health Organization (WHO), although these numbers likely underestimate the actual level of damage due to the limited testing capacity. An article in the Irrawaddy reports: ‘Yangon is finding it must rely not only on the strength of its medical professionals but also on volunteers in its fight against the coronavirus pandemic…There was a shortage of personnel – from medics to drivers to cleaners – to handle the volume of cases. Most of the existing staff complained of exhaustion (Htwe, Reference Htwe2020)’.
After the military coup in 2021, the virus continued to threaten public health. The lack of data transparency after the military coup, commonly observed in authoritarian regimes during a pandemic (Annaka, Reference Annaka2021), makes it difficult to gauge the degree of damage from the virus. However, anecdotal evidence suggests that the military government also had difficulties handling it. Wittekind (Reference Wittekind2021) argues that the situation surrounding the health sector remains challenging: ‘staff shortages, military violence against medical staff, and widespread distrust of authorities have weakened Myanmar's historically under-resourced healthcare sector, rendering it less able to manage care and vaccination (2)’.
Despite the government's lack of proper measures before and after the coup, at the time of the survey (June 2022), the country's COVID-19 situation was relatively stable. According to the WHO data, there were three major waves of confirmed cases and deaths: September–December 2020, June–October 2021, and January–April 2022. Although data are likely underreported, neighbouring countries such as Thailand experienced waves around the same time (June–October 2021 and January–April 2022 in particular). Thus, the time of the survey coincided with a period of relatively good conditions. This scenario might have led to less critical views of the military government with respect to its handling of the crisis; if the survey had been held during more severe times, the extent to which the respondents disliked non-democracy could have been stronger, and the extent their opposition to non-democracy was moderated by the severity of the pandemic in the vignette could have been weaker.
4. Results
4.1 Main results
The results are reported in this section. As the dependent variable is ordinal, we run ordered logit regressions in which the dependent variable is the favourability of the hypothetical scenario presented in the vignette. The key independent variables are a binary variable indicating the assignment to the hypothetical scenario with the election restored, another binary variable indicating the assignment to the hypothetical scenario with a good COVID-19 situation, and the interaction between the two. Table 3 reports four models: the first with the two binary variables indicating experimental assignments, the second with the two treatments and the interaction term, and the third and fourth with individual-level control variables added to the first two models. Despite randomization, control variables are added due to an imbalance in the gender variable and to check whether the results hold after taking into account respondents' characteristics. The variables included are age, educational level, gender, ethnicity, satisfaction with the government's COVID-19 measures before and after the coup, and respondents' health-protecting measures.Footnote 9 In Appendix B, we report raw results – a contingency table showing the relationship between the experimental assignment and the dependent variable.
Main results: impact of democracy on the favourability of the hypothetical situation under good and bad COVID-19 conditions
Note: The table reports the results of ordered logit regressions in which the dependent variable is the favourability of the hypothetical scenario presented in the vignette experiment (five-point scale). Model 1 is the baseline model, which has the experimental treatments as the independent variables. Model 2 adds to model 1 the interaction term between the two treatments. Models 3 and 4 add to models 1 and 2 the respondents' characteristics available from the survey responses. Standard errors are reported in parentheses.
***P < 0.01, **P < 0.05, *P < 0.1.
Table 3 reveals that respondents prefer democracy (restoration of election) to an authoritarian regime (status quo) and a good COVID-19 condition to a bad one (model 1); the pattern holds after taking into account control variables (model 3). Because the coefficients are difficult to interpret directly, we calculate the change in the predicted probability of selecting ‘favourable’ when the key independent variables take different values (not reported in the table). We find that the probability of selecting ‘favourable’ increases by 47.3 percentage points when the democracy variable changes from ‘election not restored’ to ‘election restored’. The probability increases by 17.8 percentage points for the COVID-19 treatment (from ‘bad’ to ‘good’). Thus, the respondents prefer democracy to non-democracy by a wide margin, while also feeling favourable about the good COVID-19 situation.
As our hypothesis centres on whether attitudes towards democracy differ depending on the pandemic situation, the interaction term should be interpreted. The significant coefficient suggests that the effect of democracy indeed varies (models 2 and 4). In Figure 1, we plot the marginal effect of democracy on the probability of selecting ‘favourable’ for the two levels of the COVID-19 situation. The figure shows the following: (1) the marginal effect of democracy is positive for both levels of the COVID-19 situation (good, bad); and (2) the magnitude of the effect is greater when the COVID-19 situation is good than when it is bad (59.8 and 34.6 percentage points, respectively). This finding suggests that although the respondents prefer democracy to non-democracy, the bad COVID-19 condition reduces their support, which is consistent with the hypothesis.
Marginal effect of democracy under good and bad COVID-19 conditions.
Note: The figure plots the marginal effect of democracy on the probability of selecting ‘favourable’ based on model 4 in Table 3. Among those who were assigned to the ‘COVID-19 situation is good’ condition, the random assignment to ‘democracy’ is expected to increase the probability of saying that the hypothetical situation is ‘favourable’ by 0.598 compared to the random assignment to ‘non-democracy’. For those assigned to the ‘COVID-19 situation is bad’ vignette, the hypothetical situation with democracy is expected to increase the probability by 0.346 compared to when the hypothetical situation is non-democratic.
We stress that respondents strongly prefer democracy. Even when the pandemic situation is bad, the level of support for democracy is substantially higher than that for non-democracy. This result is consistent with anecdotal evidence suggesting distrust in the military government, which has led to demonstrations, armed opposition, and civil disobedience movements in various parts of the country. What we demonstrate is that the extent to which democracy is preferred is lower when the pandemic condition in the vignette is classified as bad.Footnote 10
Finally, it is possible that people prefer a peaceful condition regardless of the regime, and the restoration of elections can be interpreted as a future scenario with a decline in unrest and increased stability. Unfortunately, our design does not allow us to test whether and to what extent people equate democracy and stability. Ideally, we would have included questions to confirm how the respondents perceived the treatment. Also, our vignette should have had one more variable – stability – that can take two values: status quo and substantial improvement. With this design, we could test whether people prefer democracy to non-democracy, holding the level of stability constant, and clarify whether the impact of democracy is moderated by the level of stability.
4.2 Additional analyses
Related to the issue of stability or democracy, we acknowledge that the causal mechanism remains untested. Ideally, the experiment should have been designed to allow for a direct test of the causal mechanism. Here, we refer to other questions asked in the survey to test an observable implication of our argument. Specifically, the hypothesized mechanism is that people perceive strong measures as useful and that the non-democratic government may be better at managing a pandemic. Questions asking these items were not included in the survey. However, we did ask respondents to report their satisfaction with the government's health-protecting measures before and after the military coup, using a five-point ordinal scale from dissatisfied (1) to satisfied (5). We calculated the change in the level of satisfaction from the pre- to post-transition period (average of six items, including travel restriction, stay-at-home policy, facilities quarantine, public healthcare services, government financial help for households and businesses, and the expansion of government health expenditure).
Data are reported in Appendix Table D1. Overall, the level of satisfaction is lower after the transition than before. If our argument is correct, individuals who believe strict measures effectively combat the pandemic are likely to show weaker support for democracy when the COVID-19 condition is bad. Therefore, assuming those less critical of the military government's handling of the pandemic (those with relatively higher levels of satisfaction) prefer stricter measures, we would expect to observe a smaller marginal effect of the democracy treatment for this group than for those who are more critical.
To test this possibility, we divide the sample into two groups: those whose change in the level of satisfaction from the pre- to post-coup period is above the median (more critical of the military government's handling of the pandemic) and below the median (those who are less critical). In other words, the level of satisfaction declined more substantially for the former than the latter. Although this is not a direct test, if the marginal effect of democracy under the bad COVID-19 condition is smaller for those who are less critical of the military government, the finding would favour our argument.
Appendix Figure D1 reports the marginal effects of democracy for these two groups. It shows that the marginal effect of democracy under the bad COVID-19 condition is indeed smaller among those who are less critical of the military government: the marginal effect is 0.244 among those who are less critical of the military and 0.352 among those who are more critical of the military. In addition, in an ordered logit regression focusing on subjects assigned to the bad COVID-19 treatment, the coefficient on the interaction term between the democracy treatment and the binary variable indicating those who are less critical of the military is negative and statistically significant (P = 0.008). Although the experiment was not designed to test for causal mechanisms, we have a finding inclined towards our hypothesized mechanism.
Relatedly, we provide a brief description of the security situation in Myanmar. Specifically, the condition was relatively stable around the time of the survey (June 2022). Street protests were taking place in villages in the central and northern parts of the country, particularly in Magway and Sagain regions. Additionally, people dissatisfied with the military coup were expressing their opinions on social media. It is also important to point out that the regime was continuously arresting individuals. However, no protests were taking place in major towns or cities, including Yangon. Concurrently, since the coup, both Yangon and Mandalay have been under curfew from 8 p.m. to 4 a.m.
In other words, although certain restrictions were imposed by the regime, the situation was relatively stable when the survey was conducted. Compared to bad security conditions when confrontations and battles frequently occur (such as after October 2023), the situation in May 2022 was rather favourable for testing our hypothesis because people's level of concern about security was presumably not very high. Therefore, while acknowledging the need for an experimental design that varies security conditions, we argue that the concern about confounding is moderate. If the survey was conducted during periods of frequent physical confrontations in major cities, people's concern about security and stability would have been substantially higher.
5. Supplementary analysis: conjoint experiment
This section provides the results of a conjoint experiment carried out in the same survey to deal with the possible desirability bias and to check whether the results of the vignette experiment are robust to the inclusion of other hypothetical conditions.Footnote 11 The survey included a question asking whether the respondent would like to see the country develop under democracy, and 92.2% selected ‘Yes’. In contrast, when asked whether they would like to see the country develop under non-democracy, only 10.5% responded positively. This result suggests that the respondents were relatively open to stating their opinion on democracy. However, in case some hesitated to report their attitudes, the results of the conjoint experiment are reported.Footnote 12
The following design was used. We presented a pair of hypothetical conditions of Myanmar society in May 2023, approximately 1 year after the current survey. The question was: ‘Imagine that Myanmar in 2023 looks as follows. Which of the following two situations is more favourable to you?’ The profile (hypothetical Myanmar society in May 2023) consisted of four attributes, including the seriousness of the pandemic condition, availability of vaccines, availability of medical services, and whether democracy is restored. The aim was to examine the relative importance of several attributes related to the COVID-19 condition and whether the impact of the attributes varies across the pandemic conditions. The respondents were asked to select which of the two hypothetical conditions they prefer. The task was repeated five times, resulting in the evaluation of 10 profiles from each respondent. Table 4 summarizes the attributes and levels.
Conjoint design
Note: The question asked is: ‘Please consider the following two different hypothetical conditions of Myanmar in May 2023. Which of the following two hypothetical situations do you prefer?’ Two profiles were presented for each task, and the respondent was asked to select the one that they preferred. The task was repeated five times. In the attribute of government, ‘Election is already restored’ is the level that corresponds to democracy; ‘Election is not restored yet. The government is the same as now’ corresponds to non-democracy. For each respondent, the order of the attributes is randomly decided by software and fixed throughout the five tasks.
Our task here is to examine whether the impact of democracy varies across the seriousness of the pandemic condition. The attribute of democracy has two levels: ‘election has been restored’ and ‘election has not been restored’; the attribute of the pandemic condition has three levels: ‘very bad’, ‘not so good’, and ‘good’. If the results of the conjoint experiment are consistent with the findings reported in Section 4, we should observe that the impact of democracy on the respondent's chance of selecting the profile as preferable is smaller for those who were randomly assigned to ‘very bad’ in the attribute of the pandemic condition than those assigned to ‘good’.
We first estimate the AMCE (Hainmueller et al., Reference Hainmueller, Hopkins and Yamamoto2014), which reports the change in the probability of selecting a profile when an attribute changes from the baseline level to a specific level. For example, we measure how much the probability of selecting a profile changes when the attribute ‘government’ changes from non-democracy (‘Election is not restored yet’) to democracy (‘Election is already restored’). We run ordinary least square (OLS) regressions in which the unit of analysis is profiles evaluated by the respondents, the dependent variable is a binary variable that takes the value of one if the profile is selected as a preferred alternative of the two, and the independent variables are attribute levels.
In addition, we calculate a conditional AMCE, which informs us of how AMCE changes as another attribute takes different values. For example, we can examine whether the impact of democracy is different between good and bad pandemic conditions by adding interaction terms between an attribute level (the key independent variable) and another attribute level (the one that can condition the effect of the independent variable). The results are reported in Table 5. Model 1 reports AMCEs of all the attributes, whereas models 2–4 report conditional AMCEs, showing the marginal effects of the attribute levels for the profiles with good, not so good, and bad COVID-19 conditions, respectively. In addition, to test whether the difference in AMCEs for the attribute of government across levels of the pandemic condition is statistically significant, we present two models. Model 5 includes interaction terms between the COVID-19 conditions and the election dummy. Model 6 parallels model 5 but only includes those profiles that have the good and bad COVID-19 conditions to ensure comparability with the vignette. Figures visualizing the findings in Table 5 are provided in Appendix E.
Results of conjoint experiment
Note: The dependent variable is a binary variable equal to one if the profile is selected by the respondents as a preferable alternative. The independent variables are binary variables indicating attribute levels. For each attribute, one level is used as the baseline value. Model 1 is the baseline model with all the attribute levels. Models 2–4 are subgroup analyses for conditional AMCEs: profiles with the COVID-19 attribute taking each level are analysed separately. Model 5 adds to model 1 interaction terms between the COVID-19 and election attributes. Model 6 is the same as model 5, except that the profiles with ‘not good’ COVID-19 conditions are excluded, so we have two attribute levels similar to the vignette. OLS is used. Each respondent evaluated five pairs (10 profiles).
***P < 0.01, **P < 0.05, *P < 0.1.
We first interpret AMCEs when all attribute levels are included (model 1). The results reveal that the profile has a lower chance of selection when the attribute of COVID-19 takes the value of ‘bad’ compared to the baseline level (‘not so good’). It also shows that restoration of election is very popular; the probability that the respondents select a profile increases substantially (by approximately 60 percentage points) when the attribute of government takes the value of ‘election is restored’ from the baseline value (‘election has not been restored’). These are consistent with the results of the vignette experiment reported in model 1 of Table 3. Thus, respondents still prefer democracy to non-democracy.
Second, we interpret conditional AMCEs. Models 2–4 reveal that the size of the coefficient on the ‘election restored’ is the largest when the COVID-19 attribute takes the value of ‘good’; the probability of selecting the profile increases by 64.6 percentage points. This result is followed by when the COVID-19 condition is ‘not so good’ and ‘bad’: the probability of selecting the profile increases by 60.3 and 56.2 percentage points, respectively. These findings are also consistent with our expectations.
Third, the coefficient on the interaction term between the good COVID-19 condition and the election dummy is statistically significant (models 5 and 6).Footnote 13 However, if multiple testing correction is incorporated as recommended by Liu and Shiraito (Reference Liu and Shiraito2023) and carried out by scholars analysing survey experimental data for hypotheses testing (e.g. Liu et al., Reference Liu, McElwain and Shiraito2023; Kuzushima et al., Reference Kuzushima, McElwain and Shiraito2024), the P-value increases; when Bonferroni correction is used in model 5, for example, the P-value is eight times larger because the number of independent variables is eight (six attribute levels and two interaction terms). As a result, the coefficient on the interaction term is not statistically significant at the 10% level. Therefore, multiple testing correction weakens evidence in favour of our argument; on the contrary, the direction of the effects is as expected.
Given the different experimental designs, the effect size in the vignette is not directly comparable to the conjoint experiment. However, we demonstrate that the direction of the results is the same in the two experiments.
6. Discussion
We hypothesized that because of the authoritarian regime's capacity to take swift and forcible actions to combat the pandemic, people's antipathy towards the regime may be weakened when the pandemic condition is bad than when it is good. To test this hypothesis, we conducted a vignette experiment that varied the severity of the COVID-19 situation (good or bad) and the regime (whether elections had been restored or not). We found that although the respondents preferred democracy (restoration of election) to non-democracy (continuation of the military rule) by a wide margin, the extent to which democracy was preferred was lower when the hypothetical pandemic situation was bad. A conjoint experiment was conducted for a supplementary analysis, presenting a pair of hypothetical conditions of Myanmar society consisting of several attributes. The findings were consistent with the vignette experiment.
Considering the broader implications of our research, our interpretation is that the severe pandemic condition weakened opposition to non-democracy even in Myanmar. Given the unpopularity of the military regime, Myanmar could be a less likely (if not the least likely) case to detect the negative effect of the pandemic on support for democracy. Beyond the context of Myanmar, the backsliding of democracy and the violation of democratic principles have been observed in many countries during the pandemic (e.g. Edgell et al., Reference Edgell, Lachapelle, Lührmann and Maerz2021). In 2020 and 2021, 73 and 60 countries experienced declines in the level of freedom (Repucci and Slipowitz, Reference Repucci and Slipowitz2022). Although we have not tested our argument elsewhere, democratic backsliding implies that similar results should be obtained had the experiment been replicated in other countries.
Finally, we conclude by discussing two extensions of the current research for future work. First, related to the point discussed above, whether our results can be generalizable beyond Myanmar should be tested. For example, what if the democratic government was perceived as corrupt and people supported non-democratic rule, or if the authoritarian government was perceived as competent? As mentioned, our setting – where the degree of support for the elected government was presumably very strong, and the military government is unpopular – seems to make it more challenging to find the negative effect of the pandemic on support for democracy. If conducted in a capable or not-so-unpopular autocracy, our results could be even stronger. Furthermore, in such societies, assuming some people have a positive evaluation of the non-democratic regime and a strong belief in the positive impact of forcible actions during a crisis, the impact of the democracy treatment could be negative under the hypothetical scenario of a bad COVID-19 situation.
Second, some respondents in our survey may have equated democracy with the NLD-led government. We avoided partisan labels in our experiments, such as ‘Election is restored and the NLD controls the government’, to understand people's attitudes towards the regime itself; however, it is not clear whether the true level of support for democracy (regardless of who is in power) would be as high in reality as it was in our experiment. Relatedly, Myanmar was not entirely democratic prior to the 2021 transition, according to some measures. For example, Myanmar was categorized as ‘not free’ as of 2020 by Freedom House, scoring low both in political rights and civil liberties (Freedom_House, n.d.). Thus, it is essential to verify in future work how people define terms such as ‘elected government’ or ‘democracy’ when they appear in surveys.
Acknowledgements
We truly appreciate helpful suggestions and comments from four anonymous reviewers. An earlier version of this paper was presented at the Lien Development Conference 2022 (November 2022) and the 2022 Japan Public Choice Society meeting (December 2022). We are grateful for comments and suggestions from Susumu Annaka, Kenneth Mori McElwain, and other participants. We would also like to thank Seunghoo Lim and Yusuke Jinnai for their comments and suggestions.
Competing interests
None.
Supplementary material
The supplementary material for this article can be found at https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/YUZSDO&version=DRAFT&faces-redirect=true.
Appendix A: Balance across experimental groups
Democracy treatment
COVID-19 treatment
Appendix B: Experimental results – contingency table
Appendix C: Additional analyses
Ordinary least squares
Logit
Regressions with the gender variable as control
Marginal effect.
Note: The figure plots the marginal effect of democracy on the probability of selecting ‘somewhat favourable’ or ‘favourable’ based on the logit regression reported in Appendix Table C2, model 2. Among those who were assigned to the ‘COVID-19 situation is bad’ condition, the random assignment to ‘democracy’ is expected to increase the predicted probability of favouring the hypothetical situation by approximately 0.44 compared to the random assignment to ‘non-democracy’. For those assigned to the ‘COVID-19 situation is good’ condition, the hypothetical situation with democracy is expected to increase the probability of being favoured by 0.75 compared to when the hypothetical situation is non-democratic.
Appendix D: Additional analyses
Summary table
Marginal effect of democracy.
Note: We ran ordered logit regressions in model 4 in Table 3 with subgroups: (1) those who are more critical of the military regime (those whose satisfaction levels declined more) and (2) those who are less critical. We find that the marginal effect of democracy is lower for those who are less critical of the military regime when the COVID-19 condition is bad than among those who are more critical, which is consistent with the expectation. The marginal effect of democracy is also smaller among those who are less critical of the military regime when the COVID-19 condition is good.
Appendix E: Visualization of the results of the conjoint experiment
Conditional AMCEs of the attribute levels.
Note: The figure reports the marginal effects of attribute levels at different values of the attribute of COVID-19 conditions based on model 2 (COVID good), model 3 (COVID not so good), and model 4 (COVID bad) of Table 5. Attributes and levels are described in Table 4. Dots show coefficients, and horizontal lines indicate 95% confidence intervals.
Conditional AMCE of democracy.
Note: The figure reports the marginal effects of the attribute level of ‘Election is already restored’ at different levels of COVID-19 conditions, based on model 2 (COVID good), model 3 (COVID not so good), and model 4 (COVID bad) of Table 5. The baseline level is ‘Election is not restored yet. The government is the same as now’. Dots show coefficients, and horizontal lines indicate 95% confidence intervals.
