Policy Research Working Paper 9416 Perceptions, Contagion, and Civil Unrest Christophe Abi-Nassif Asif Mohammed Islam Daniel Lederman Middle East and North Africa Region Office of the Chief Economist September 2020 Policy Research Working Paper 9416 Abstract This paper investigates the empirical relationship between conditions. The heterogeneous effects of perceptions on citizens’ perceptions of economic and political conditions uprisings across geography and income groups, however, and the incidence of nonviolent uprisings. Perceptions are are not robust and susceptible to changes in estimators and measured by aggregating individual-level data from regional model specification. In particular, the international conta- barometer surveys. The main results show that negative gion of protests eliminates this international heterogeneity, perceptions of political conditions—proxied by the share implying that the incidence of uprisings in nearby countries of the population that is generally dissatisfied with the way tends to generate protests at home through its effect on democracy works—have a significant positive effect on the perceptions related to political conditions in high-income number of protests and strikes. Negative perceptions of countries. Overall, the effect of perceptions about political economic conditions do not seem to be significantly related conditions, along with protest contagion, is robust to the to the latter. This generally holds across a large sample of inclusion of numerous control variables that capture actual countries and is particularly the case for Western and Cen- economic conditions and the quality of governance across tral European countries as well as high-income countries. countries. The results are also robust to the use of seemingly In developing economies, however, social protests appear valid instrumental variables, alternative count-data estima- to be driven by dissatisfaction with economic and political tors, and sample composition. This paper is a product of the Office of the Chief Economist, Middle East and North Africa Region. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at dlederman@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Perceptions, Contagion, and Civil Unrest Christophe Abi-Nassif Asif Mohammed Islam Daniel Lederman1 World Bank Group Middle East and North Africa Region Office of the Chief Economist Keywords: Citizen Perceptions, Civil Resistance, Nonviolent Uprisings JEL Codes: D74, D60 1 Abi-Nassif was a consultant with the World Bank’s Office of the Chief Economist for the Middle East and North Africa (MNACE) Region. Islam is a Senior Economist with MNACE, and Lederman is the Deputy Chief Economist for MENA. The opinions expressed in this paper do not represent the views of the World Bank Group, its Board of Directors or the Governments they represent. All errors and omissions are the authors’ responsibility. We also gratefully acknowledge the invaluable support and comments from Rabah Arezki and Michael Robbins. I. Introduction Nonviolent protests have been making headlines. The last decade alone witnessed a significant global spike in peaceful uprisings and street mobilizations against established regimes and governments (Arezki et al., 2020a). The existing literature on civil resistance has already made the case that nonviolent methods yield better results than violent confrontations, most notably because of the former’s ability to garner more domestic and international legitimacy, support and sympathy, and the latter’s reduced ability to justify violent repressions of nonviolent mobilizations (Chenoweth and Stephan, 2008). In parallel, an emerging literature has explored the determinants of nonviolent protests (Chenoweth and Ulfelder, 2017; Witte et al., 2020; Arezki et al., 2020a). A pertinent question concerns the role of a populace’s perceptions in generating nonviolent conflict. Weak economic conditions and deficient institutions may fuel negative perceptions of the economy and democracy (lack of voice), which then translate into nonviolent protests. However, there may be situations where perceptions can lead to unrest despite stable or improving economic or institutional conditions. For instance, a government can engage in reforms that require time to have any material effects. Furthermore, economies suffering from weak transparency may have limited avenues through which governments can communicate reforms or economic progress to their citizens. Worse, non-transparent governments may have fundamentally lost the trust of their citizens, and thus may find it difficult to alter perceptions despite implementing reforms and improving economic conditions (Arezki et al., 2020b). The central point is that perceptions may matter, sometimes independent of the realities on the ground. Over the past two decades, research institutions and analytics companies have facilitated access to reliable data to add quantitative rigor to public opinion research and contribute to policy making. Such opinion surveys inherently capture citizens’ perceptions and concerns – including what makes them angry or dissatisfied. In this paper, we investigate the relationship between citizen perceptions and the incidence of nonviolent protests and strikes using regional public-opinion barometer surveys. Following existing literature, we hypothesize that negative perceptions of the state of the economy and the perceived lack of voice in the political system in a country translate into street protests and civil mobilization. Indeed, Figure 1 shows that higher terciles of negative political and economic condition perceptions are correlated with higher levels of protests. This unconditional relationship varies depending on whether it is within or outside the European economies covered in the sample. In addition, for the sample of countries outside Europe, the relationship between perceptions about the state of the economy appear to be uncorrelated with the incidence of civil unrest. The paper explores the robustness of these relationships as we add control variables and experiment with various econometric estimators. 2 Figure 1: Political and Economic Perceptions and Civil Unrest 4 4.5 Average Number of Strikes and Average Number of Strikes and 4 3.5 3.5 3 3 Demonstrations Demonstrations 2.5 2.5 2 2 1.5 1.5 1 1 0.5 0.5 0 0 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd 1st 2nd 3rd Political Perception (Terciles) Economic Perception (Terciles) All Economies Euro Non-Euro All Economies Euro Non-Euro Note: Perception data is obtained from the barometer surveys. Protest data is obtained from the National Domestic Conflict Database. The x-axis presents terciles of perceptions. The data covers 742 country-year observations. Appendix 2C provides details on the sample. We find that at the global level, after accounting for a host of factors, the perceived lack of voice in the political system, as measured by a general dissatisfaction with the way democracy works, materializes in anti-government demonstrations and strikes. Negative perceptions of the state of the economy on average do not seem to influence the incidence of such events. Some evidence suggests that this finding is particularly the case for Western and Central European and high-income countries. In contrast, nonviolent civil resistance in non-European countries is related to a general dissatisfaction with current economic conditions, and in developing economies it is driven by a dissatisfaction with both economic and political conditions. However, controlling for international contagion of protests appears to eliminate this international heterogeneity, implying that the incidence of uprisings in nearby countries tends to generate protests at home through its effect on perceptions of political conditions in high-income countries. Overall, these findings shed new light on the different perceptions that motivate citizens to take to the streets in different regions and income groups. 3 The main findings are robust to a plethora of factors and estimation models. We address concerns of simultaneity bias using an Instrumental Variables (IV) approach. Lagged economic growth, economic perceptions, and political perceptions as instruments appear to be statistically valid IVs, thus suggesting that the build-up of negative perceptions as well as poor economic performance are likely to result in social protests. We also experiment with economic and political perceptions contagion from other countries. This paper complements the literature on the underlying causes of social unrest. The analytical and empirical approach to analyzing nonviolent civil resistance re-emerged in the second half of the 2000s (Chenoweth & Cunningham, 2013). Multiple factors, ranging from outrage over dictatorial rule, polarization, and restrictions on basic liberties to food and commodity price increases are some of the factors that have been studied as drivers of social unrest (Abu-Bader and Ianchovichina, 2019; Cammett & Diwan, 2013; Arezki & Bruckner, 2011). Other more subtle drivers include contagion and social diffusion effects as well as spillovers facilitated by social media penetration in both source and destination countries (Braha, 2012; Arezki et al., 2020a). Our study builds on Witte et al. (2020) and Arezki et al (2020a). The former explores the effects of subjective well-being on both violent and nonviolent uprisings using the Gallup World Poll; the latter studies contagious protests with data from both protest counts and news media articles. Witte et al. (2020) find that an increase in the percentage of self-reported suffering is positively related to nonviolent protests, but not their violent counterparts. Their measure of subjective well-being is based on individuals’ ratings of their current and anticipated future life satisfaction. We improve on this general measure by honing into specific perceptions regarding the state of the economy and political conditions. We assess individuals’ perceptions of economic and political conditions as indicators of subjective well-being associated with the health of the economy and of institutions. Differentiating perceptions by economic or political conditions yields different findings. In turn, this paper studies the robustness of the role of public perceptions as determinants of the frequency of protests to the inclusion of variables that capture the contagion of protests across international borders. In summary, our study contributes to the literature by exploring the relationship between economic and political perceptions and nonviolent protests using regional barometer surveys. Generally speaking, negative perceptions of political conditions – or perceived lack of voice – seem to increase the incidence of strikes and anti-demonstrations, but not negative perceptions of economic conditions. This may be flipped for non-European countries where perceptions of economic conditions matter, but not those of political conditions. In developing economies, including those from Eastern Europe, both political and economic perceptions seem to play a role in generating nonviolent conflict. The rest of the paper is organized as follows. Section II provides an overview of the data. Section III presents the estimation framework and main regression results. Section IV presents 4 robustness checks including count data models and instrumental variable regressions. Section V concludes. II. Data Data on nonviolent uprisings are available from the National Domestic Conflict Database through the Cross-National Time-Series (CNTS) Data Archive (Banks and Wilson, 2020). The data set has been used widely in the literature primarily because of the extent of both its geographic coverage of 200 plus countries and time coverage since 1815 (Witte et al., 2020; Chenoweth & Ulfelder, 2017; Davenport, 1995).2 We define nonviolent uprisings (also interchangeably referred to as civil resistance or civil unrest) as the sum of strikes and anti- government demonstrations in a country (Witte et al., 2020). Perceptions data on economic and political conditions are obtained from six regional barometer surveys. These include the Arab Barometer, the Afrobarometer, Latinobarómetro, the Eurobarometer, the Asian Barometer and the South Asian Barometer.3 We harmonized and aggregated individual-level data for two perceptions variables that had ample coverage across the barometers and were pertinent to the analysis. For perceptions of economic conditions, we use data from answers to the following question: “How would you describe (evaluate) the present economic conditions (situation) of your country?” For perceptions of political conditions, we use answers to the following question: “How satisfied are you with the way democracy works in your country?”4 This question captures the perceived lack of voice in the political system.5 The share of the population that thinks that current economic conditions are either bad or very bad constitutes our variable of interest for the perception of economic conditions. Similarly, the share of the population that is either quite dissatisfied or 2 We also considered other popular conflict databases such the Armed Conflict Location & Event Data (ACLED) Project but proceed with the CNTS database because of ACLED’s exclusion of Western European countries and its limited sample size prior to 2010 (Arezki et al., 2020a). 3 The Eurasian Barometer – a regional barometer covering countries in Eastern Europe, the Caucasus and Central Asia – is not included because of lack of access to the data. Only one wave of the Eurasian Barometer data is publicly available and comes from the Global Barometer Survey data. That wave ends up being dropped nevertheless because of singleton observations for these countries and the resulting impact of country fixed effects. 4 The Arab Barometer does not cover this exact question and is therefore excluded from the main results. As a robustness check in section IV, we include Arab Barometer data by leveraging a comparable question: “How would you evaluate the state of democracy and human rights in your country?” This approach was used in the Global Barometer Survey where the aforementioned question was used as a proxy to the satisfaction with democracy variable we use in our main regression. 5 An alternative interpretation could be that this question captures whether one considers a democratic system less favorable than an authoritarian one, but this seems unlikely. 5 totally dissatisfied with the way democracy works constitutes our variable of interest for the perception of political conditions. We harmonize the data by adjusting inverted answer scales (e.g., from 1-4 to 4-1) and apply weights when available. Some of the barometers allow for neutral answer choices (i.e., ‘neither good nor bad’, ‘neither satisfied nor dissatisfied’) for the two perceptions questions while others do not. In our main estimation, we include the sample with all responses (i.e., positive, neutral and negative answer choices). As robustness checks, we re-estimate the baseline models by including non-neutral answers only (i.e., positive and negative answer choices only) as well as unweighted responses. Other data on GDP per capita, GDP per capita growth, total population, inflation, unemployment, oil rents, infant mortality, cellular subscriptions, and urban population come from the World Bank’s World Development Indicators. We account for institutional variables from the Polity5 data set, the Political Terror Scale (PTS) data set, and the World Governance Indicators. From Polity5, we use the institutionalized autocracy score which is derived from an 11-point additively constructed autocracy scale measuring political characteristics of authoritarian regimes. From PTS, we use an average score of the three indicators of political terror as defined by Amnesty International, Human Rights Watch, and the US Department of State. From the World Governance Indicators, we use an average score of the six governance sub-indexes: voice and accountability, political stability and absence of violence and terrorism, government effectiveness, regulatory quality, rule of law, and control of corruption. It is worth noting that none of these indicators is based on popular perceptions. Our estimation sample consists of an unbalanced panel of 742 country-year observations from 1996 to 2017 and spanning 88 countries. Descriptions and summary statistics of all variables are provided in Appendix 1A and Appendix 1B respectively. A covariate pairwise correlation table is presented in Appendix 1C for the baseline regression sample. Appendix 2A, Appendix 2B, and Appendix 2C detail the number and distribution of sample observations per year, barometer, and country-year respectively. 6 III. Estimation Framework and Results A. Estimation Model The regression model to investigate the relationship between nonviolent uprisings and citizens’ perceptions of economic and political conditions for country i in year t can be written as: = 0 + 1 + 2 + + + + (1) In the baseline OLS estimations of equation (1) nonviolent uprisings (or civil resistance) is measured as the sum of strikes and anti-government demonstrations transformed using the hyperbolic inverse sine as in Witte et al. (2020). PECit and PPCit are proxies for the perception of economic conditions and perception of political conditions respectively. PEC refers to the share of the population in country i at time t that thinks that the current economic conditions are either bad or very bad. PPC refers to the share of that same population that perceives that there is a lack of voice in the political system as proxied by whether they are either quite dissatisfied or very dissatisfied with the current state of democracy. Xit is a vector of control variables. The latter can be grouped into four main categories. The first category consists of major economic indicators that reflect economic conditions. It includes the log of GDP per capita, GDP per capita growth, inflation, and unemployment rates. The second category consists of demographic indicators such as the log of total population and infant mortality rates per 1,000 live births. The third category takes into account the structure and level of development of the domestic economy and includes oil rents, the urban population percentage as a share of total population as well as the number of mobile cellular subscriptions per 100 people. The fourth and final category focuses on institutional development and includes the institutionalized autocracy score, the average political terror scale indicators as well as the average world governance indicators. FEi and FEt are country and year fixed effects. For all baseline estimations, robust standard errors are presented clustered at the country-level. The OLS estimates could be biased because of the transformation of the underlying counts of protests before taking logs, with the resulting sample distribution violating the OLS assumption of a normal distribution. Consequently, we also estimate equation (1) using three alternative count-data estimators: a conditional-mean Poisson regression with fixed effects captured by the country mean of the dependent variable, a conditional-mean negative binomial regression, and a negative binomial regression with country dummy variables (instead of country means of the dependent variable to control for country fixed effects). These estimators are discussed in detail in the count data models section IV-A. We use the untransformed count data (i.e., incidence of strikes and anti-government demonstrations) for 7 estimations using these count data models. We also conduct a series of robustness checks to address concerns related to specification, sample composition, and survey harmonization (section IV-B). The estimation equation (1) is arguably subject to endogeneity concerns due to omitted variable bias and simultaneity bias. We account for the former by including a comprehensive array of control variables in the main specification. Since protests could influence perceptions, we account for the potential of reverse causality by implementing an Instrumental Variables (IV) approach. We use the lagged levels of economic growth, economic perceptions, and institutional perceptions as instruments, under the assumption that that the build-up of negative perceptions and poor economic conditions could result in worsening perception and thus in social protests. We also use economic and political perceptions contagion from other countries given existing evidence of contagious protests whereby dissatisfaction with economic and political conditions in nearby countries may also influence popular perceptions at home. These instruments are discussed in detail in section IV-C. B. Main Results Table 1 reports the OLS results with country and year fixed effects across the full sample with the stepwise addition of control variables as we move to the right of the table. The effect of the perception of political conditions on the incidence of nonviolent civil resistance is positive and statistically significant at the 1% level. Taking the full-model specification in column 13, a one percentage-point increase in the percentage of people who are either quite dissatisfied or very dissatisfied with how democracy works in their country leads to a 1.9% increase in the number of nonviolent civil resistance events. Regarding the economic magnitude of this coefficient estimate, a one standard deviation increase in the percentage of the population that is generally dissatisfied with democracy leads to a 0.38% increase in the number of nonviolent civil resistance events. The coefficient of the economic condition perception variable is not statistically significant and switches signs across specifications. Of the control variables, only the coefficient of the inflation rate is statistically significant (at the 5% or 1% significance level) with a positive sign across all specifications. Using the full specification in column 13, we find that a one-percentage point rise in inflation is associated with a 0.9% increase in the incidence of nonviolent civil resistance events. The magnitude of this coefficient implies that a one standard-deviation shock to inflation (6.1 p.p.) is associated with a 5.5% increase in the number of protests. Hence it is plausible that inflation is a powerful driver of protests, perhaps even to a greater extent than perceptions. This finding is consistent with the results reported by Arezki and Bruckner (2011) on the role of the food price inflation. However, it remains an open question whether the estimated effect of inflation on civil unrest is robust to changes in samples, the inclusion of the protest-contagion variable, and alternative count-data estimators. 8 Table 1 – OLS with Country and Year Fixed Effects Regression Results (Full Sample) Model OLS with Country and Year Fixed Effects (Full Sample) Civil Resistance (Transformed at the Hyperbolic Inverse Sine) Outcome Variable Consists of the Number of Strikes and Anti-Government Demonstrations (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) % of Citizens Thinking Current Economic Conditions Are Bad or Very Bad 0.147 0.138 0.097 0.133 0.079 -0.034 -0.034 0.017 -0.013 -0.009 -0.017 -0.021 -0.142 (0.342) (0.364) (0.378) (0.369) (0.388) (0.396) (0.407) (0.418) (0.439) (0.441) (0.446) (0.448) (0.467) % of Citizens Generally Not Satisfied with Democracy in Their Country 1.613*** 1.545*** 1.537*** 1.642*** 1.668*** 1.820*** 1.775*** 1.793*** 1.829*** 1.805*** 1.814*** 1.897*** 1.923*** (0.534) (0.535) (0.532) (0.526) (0.554) (0.497) (0.530) (0.533) (0.547) (0.547) (0.550) (0.565) (0.572) Log of GDP per Capita (Constant 2010 US$) -0.412 -0.414 -0.335 -0.325 -0.464 -0.522 -0.405 -0.435 -0.414 -0.358 -0.305 -0.340 (0.456) (0.460) (0.446) (0.508) (0.494) (0.548) (0.574) (0.610) (0.617) (0.629) (0.653) (0.771) GDP Per Capita Growth (Annual %) -0.006 -0.007 0.002 -0.001 0.005 0.004 0.004 0.004 0.003 0.004 0.001 (0.013) (0.013) (0.013) (0.013) (0.012) (0.012) (0.013) (0.013) (0.012) (0.012) (0.013) Log of Total Population -1.468* -1.558* -1.766** -1.906** -1.311 -1.036 -1.075 -1.016 -0.881 -0.672 (0.789) (0.806) (0.791) (0.794) (0.999) (0.982) (0.997) (1.027) (1.051) (1.233) Inflation (CPI) (Annual %) 0.013*** 0.013*** 0.012*** 0.011*** 0.013*** 0.013*** 0.013*** 0.013*** 0.009** (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) (0.004) Unemployment, Total (% of Total Labor Force) -0.001 0.003 0.001 0.003 0.003 0.002 0.004 0.002 (0.018) (0.019) (0.019) (0.019) (0.019) (0.019) (0.019) (0.021) Oil Rents (% of GDP) -0.013 -0.013 -0.006 -0.006 -0.006 -0.008 -0.006 (0.020) (0.021) (0.021) (0.020) (0.020) (0.021) (0.026) Infant Mortality Rate (Per 1,000 Live Births) 0.013 0.015 0.015 0.016 0.015 0.016 (0.011) (0.012) (0.012) (0.012) (0.012) (0.016) Polity5 - Institutionalized Autocracy Score 0.047 0.038 0.039 0.041 -0.003 (0.084) (0.085) (0.085) (0.084) (0.101) Political Terror Scale (Average) 0.050 0.052 0.066 0.046 (0.083) (0.083) (0.084) (0.096) Mobile Cellular Subscriptions (Per 100 People) -0.001 -0.000 -0.001 (0.002) (0.002) (0.002) Urban population (% of total population) -0.025 -0.032 (0.024) (0.024) World Governance Indicators (Average of Sub-Indexes) -0.237 (0.389) Constant 0.381 3.412 3.476 26.206** 27.309** 31.985** 35.216*** 24.147 19.948 20.264 18.811 17.579 14.759 (0.248) (4.002) (4.048) (12.241) (12.601) (12.174) (12.981) (17.155) (17.287) (17.340) (18.057) (18.304) (22.489) Country Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES YES Year Fixed Effects YES YES YES YES YES YES YES YES YES YES YES YES YES Number of observations 975 939 939 939 892 879 824 824 784 784 784 784 742 R2 0.526 0.450 0.450 0.455 0.463 0.477 0.397 0.399 0.401 0.401 0.401 0.403 0.405 Adjusted R2 0.513 0.434 0.434 0.438 0.445 0.458 0.374 0.375 0.375 0.375 0.374 0.375 0.378 note: *** p<0.01, ** p<0.05, * p<0.1. Robust Standard Errors Reported, clustered at the country level 9 It is also possible that the average effects estimated for the full sample of countries mask international heterogeneity. More specifically, we investigate differences in geography and levels of development. For geography, we find interesting differences between Eurobarometer (i.e., Western and Central Europe) and non-Eurobarometer countries. For levels of development, we find differences between high-income countries and countries with upper- middle, lower-middle and low income. Our Western and Central European subsample includes some of the most advanced economies in our sample and is obtained from the Eurobarometer surveys. Results are reported in Table 2, which replicates the specification in column (13) of Table 1. Column (1) includes the full sample. Column (2) singles out Eurobarometer data, while column (3) looks at all barometer countries except those from the Eurobarometer (i.e., sample (1) = sample (2) + sample (3)). A closer look at Eurobarometer data only (Table 2, Column 2) reveals similar but notably larger coefficient estimates of the political perception variable. A one-percentage point increase in the percentage of people who are generally dissatisfied with democracy leads to a 3.35% increase in the incidence of nonviolent uprisings. Inflation, a key economic factor in the full sample, is no longer significant. For Western and Central European countries, this may signal a stronger drive for citizens to protest when they feel less satisfied with the state of democracy rather than with the state of the economy. This interpretation seems further supported by the negative and statistically significant (at the 10% level) coefficient of the World Governance Indicators control variable. Column (3) of Table 2 showcases the regression results for non-Eurobarometer countries. For this subset of the data, the coefficient of the perception of economic conditions becomes significant at the 5% level. A one-percentage point increase in the percentage of people thinking economic conditions are either bad or very bad corresponds to a comparable 1.028% increase in the number of strikes and protests. In contrast with European countries, the perception of political conditions is not significant. Inflation is significant at the 10% level indicating that a one-percentage point increase in inflation is associated with a 0.7% increase in civil resistance events. A potential interpretation of this statistically significant and positive coefficient is that reduced purchasing power could be a main driver of nonviolent uprisings in the sampled non-Eurobarometer countries. We find a similar pattern when splitting the sample into two subsamples for high-income countries (column 4) and middle- and lower-income countries (column 5). Negative perceptions of political conditions have a statistically significant effect on civil resistance for high-income countries, but negative perceptions of economic conditions do not. A one- percentage point increase in the percentage of people who are generally dissatisfied with democracy leads to a 2.9% increase in the incidence of nonviolent uprisings. However, for developing economies, negative perceptions of both economic and political conditions seem to increase the incidence of civil resistance events. A one-percentage point increase in the 10 percentage of people who are generally dissatisfied with democracy leads to a 1.06% increase in the incidence of nonviolent uprisings, while a one-percentage point increase in the percentage of people who are generally dissatisfied with economic conditions leads to a 1% increase in protests.6 The divergence in the statistical significance of the coefficients for each of the two perception variables could offer an explanation on what motivates citizens to take to the streets through strikes and anti-government demonstrations in different regions and income groups. In Western and Central European countries (from a geographic perspective) as well as in relatively more prosperous economies (from an income perspective), dissatisfaction with the state of democracy seems to be a main driver of street protests. On the other hand, perceptions of faltering economic conditions may be more immediate and urgent motives for protests in non-European economies. In the latter, perceptions about the health of the economy might take precedence over the lack of voice in the political system. The findings for developing economies indicate that both the perceived health of the economy and the perceived lack of voice in the political system are pressing drivers of civil resistance, along with inflation. 6 The Polity 5 variable drops for the Eurobarometer and high-income samples as there is no variation in the Polity 5 autocracy score for these subsamples. 11 Table 2 – OLS with Country and Year Fixed Effects Regression Results (Sample Splits) Model OLS with Country and Year Fixed Effects Civil Resistance (Transformed at the Hyperbolic Inverse Sine) Outcome Variable Consists of the Number of Strikes and Anti-Government Demonstrations Upper-Middle, Lower- Non-Eurobarometer Full Sample Eurobarometer Countries High Income Countries Middle, and Low-Income Countries Countries (1) (2) (3) (4) (5) % of Citizens Thinking Current Economic Conditions Are Bad or Very Bad -0.142 -0.837 1.028** -0.833 1.002** (0.467) (0.580) (0.484) (0.534) (0.484) % of Citizens Generally Not Satisfied with Democracy in Their Country 1.923*** 3.353*** 0.904 2.920*** 1.066** (0.572) (1.014) (0.543) (0.921) (0.512) Log of GDP per Capita (Constant 2010 US$) -0.340 0.779 -0.587 -0.092 0.623 (0.771) (1.589) (0.842) (1.255) (0.657) GDP Per Capita Growth (Annual %) 0.001 0.000 0.010 0.004 0.005 (0.013) (0.023) (0.018) (0.022) (0.018) Log of Total Population -0.672 -2.286 -0.667 -2.636 -0.409 (1.233) (2.718) (1.452) (1.712) (1.481) Inflation (CPI) (Annual %) 0.009** -0.010 0.007* -0.031 0.007* (0.004) (0.051) (0.004) (0.023) (0.004) Unemployment, Total (% of Total Labor Force) 0.002 -0.011 0.000 -0.018 0.010 (0.021) (0.033) (0.029) (0.027) (0.027) Oil Rents (% of GDP) -0.006 0.210 0.004 0.210 0.019 (0.026) (0.547) (0.031) (0.599) (0.030) Infant Mortality Rate (Per 1,000 Live Births) 0.016 0.062 0.019 0.259** 0.028* (0.016) (0.135) (0.016) (0.103) (0.016) Polity5 - Institutionalized Autocracy Score -0.003 (dropped) 0.029 (dropped) 0.039 (0.101) (0.107) (0.106) Political Terror Scale (Average) 0.046 -0.163 0.114 -0.191 0.170 (0.096) (0.258) (0.105) (0.210) (0.116) Mobile Cellular Subscriptions (Per 100 People) -0.001 -0.009 -0.000 0.002 -0.002 (0.002) (0.008) (0.002) (0.005) (0.003) Urban population (% of total population) -0.032 -0.046 -0.034 -0.067 -0.054* (0.024) (0.072) (0.026) (0.062) (0.028) World Governance Indicators (Average of Sub-Indexes) -0.237 -1.371* 0.339 -1.393** 0.322 (0.389) (0.759) (0.469) (0.649) (0.453) Constant 14.759 33.318 18.211 47.877* 4.814 -0.142 -0.837 1.028** -0.833 1.002** Country Fixed Effects YES YES YES YES YES Year Fixed Effects YES YES YES YES YES Number of observations 742 318 424 357 385 R2 0.405 0.487 0.382 0.484 0.431 Adjusted R2 0.378 0.445 0.331 0.435 0.379 note: *** p<0.01, ** p<0.05, * p<0.1. Robust Standard Errors Reported, clustered at the country level. The Polity 5 variable drops for the Eurobarometer and high-income economies as there is no variation in the polity 5 autocracy score for these subsamples. 12 IV. Robustness and the Role of Contagious Protests This section presents several robustness checks in relation to the estimation model. First, it presents results from count-data models, which are more appropriate for the dependent variable measured by the number of incidents. Second, the model specification is extended to include protest contagion across countries. Third, the results are tested to changes in sample composition and alternative survey-question harmonization. Fourth, IVs estimates are discussed at the end. A. Count Data Estimators Given the count nature of the civil unrest dependent variable, we build on the baseline OLS model by running a Poisson regression with fixed effects. To factor in potential overdispersion and the inherent limitation of a constant error variance over the mean that is assumed by the Poisson model, we also run a conditional-mean negative binomial regression (with fixed effects). We finally run a negative binomial regression with country dummy variables to control for the effects of country fixed effects directly on protests and indirectly through the overdispersion parameter (see Allison & Waterman, 2002). The results from the three count- data models reported in Table 3 below are qualitatively similar to the OLS estimates in terms of statistical significance, but the coefficients are notably larger than the OLS coefficients. Columns (1), (3) and (5) present regression results without any control variables and columns (2), (4) and (6) show results with the full set of controls from Table 1. A one percentage point increase in the percentage of people who are generally dissatisfied with democracy leads to a 4.3, 2.7, and 3.4 percent increase in the number of nonviolent civil resistance events for the conditional-mean Poisson model (with country fixed effects), the conditional mean negative binomial model (with country fixed effects), and the negative binomial model with country dummy variables, respectively. These results are statistically significant at the 1% level. The magnitudes of these estimates are larger than that of the baseline OLS model with country fixed effects reported in Table 1, column 3 (1.9%). Similarly, the economic magnitudes of these estimates are larger than their OLS counterpart for the case of a one standard-deviation shock which is 0.38%. Across the count-data estimates, a one standard deviation increase in the percentage of the population that is dissatisfied with democracy leads to a 0.86 (conditional-mean Poisson), 0.54 (conditional-mean negative binomial) and 0.68 (negative binomial with country dummies) percent increase in the number of nonviolent civil resistance events. These results reveal a negative bias caused by overdispersion in the protest data. Indeed, the overdispersion parameter reported under columns 5 and 6 is statistically significant, thus indicating that overdispersion is an issue that probably biased the OLS estimates downwards. In principle, the negative binomial regression with country dummies probably yields the least biased results (see Allison & Waterman, 2002). Although not reported in Table 3 for the sake of brevity, the coefficient for inflation is statistically insignificant for all specifications for the conditional negative binominal estimates. 13 For the negative binomial with country fixed effects, the coefficient is not statistically significant with no controls. For the Poisson FE it is statistically significant in 2 specifications. Thus, we conclude that effect of inflation on civil unrest is sensitive to model specifications and choice of econometric estimator, in contrast to the political perceptions variable. To test for coefficient heterogeneity in the count data models, we also interact the perception variables with a dummy variable taking the value of 1 for Eurobarometer countries and 0 otherwise. The corresponding results are in Table 4 below. Likewise, interactions with a dummy variable for high-income economies are reported in Table 5. The coefficient of the interaction between perception of economic conditions and the Euro dummy variable is negative and statistically significant across all count data models, with or without control variables. The coefficient of the interaction between perceptions of political conditions and the Euro dummy is less robust across specifications. We find similar non- robust results when interacting the perception variables with the high-income dummy variable. The interaction between perceptions of economic conditions and the high-income dummy variable is negative and statistically significant for all but one specification. However, the interaction between perceptions of political conditions and the high-income dummy variable is barely statistically significant at the 10% level for a couple of specifications. Yet in both cases, in Tables 4 and 5, the interactions with political perceptions have large and positive coefficients, whereas the interactions with economic perceptions are negative, thus mimicking the OLS estimates even if the statistical significance is not robust. That is, it appears that the magnitudes of the estimated effects of economic conditions on protests are lower in the Eurobarometer data and among high-income economies than in other countries. The reverse is true for political perceptions. 14 Table 3 – Count-Data Estimates: Full Sample Civil Resistance (Count) Outcome Variable Consists of the Number of Strikes and Anti-Government Demonstrations Negative Binomial with Country Conditional-Mean Poisson Conditional-Mean Negative Binomial Model Dummies (Full Sample) (Full Sample) (Full Sample) (1) (2) (3) (4) (5) (6) Perceptions of Economic Conditions -0.510 -0.301 -0.097 0.588 -0.008 0.483 (0.604) (0.532) (0.293) (0.502) (0.403) (0.549) Perceptions of Political Conditions 2.957*** 4.268*** 2.127*** 2.680*** 2.929*** 3.411*** (0.788) (0.892) (0.419) (0.713) (0.566) (0.747) Overdispersion -0.925*** -1.003*** (0.123) (0.173) Control Variables NO YES NO YES NO YES Country Fixed Effects YES YES YES YES YES YES Year Fixed Effects YES YES YES YES YES YES Number of observations 924 675 924 675 975 742 note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors reported inside parentheses. Table 4 – Count-Data Estimates: Full Sample with Eurobarometer Dummy Interactions Civil Resistance (Count) Outcome Variable Consists of the Number of Strikes and Anti-Government Demonstrations Conditional-Mean Negative Binomial Negative Binomial with Country Model Conditional-Mean Poisson Dummies (1) (2) (3) (4) (5) (6) Perceptions of Economic Conditions 1.003* 1.623*** 0.924** 1.556** 1.226** 2.225*** (0.517) (0.621) (0.464) (0.673) (0.534) (0.789) Perceptions of Economic Conditions x Euro Dummy -3.414*** -3.576*** -1.862** -2.051** -2.700*** -2.980*** (1.019) (0.847) (0.761) (0.946) (0.774) (1.000) Perceptions of Political Conditions 2.006*** 2.471*** 1.432*** 1.988** 1.753** 1.652* (0.561) (0.847) (0.501) (0.812) (0.734) (0.998) Perceptions of Political Conditions x Euro Dummy 1.764 2.687* 1.294 1.763 2.528** 3.462** (1.180) (1.595) (0.921) (1.185) (1.109) (1.569) Overdispersion -0.979*** -1.070*** (0.126) (0.179) Control Variables NO YES NO YES NO YES Country Fixed Effects YES YES YES YES YES YES Year Fixed Effects YES YES YES YES YES YES Number of observations 924 675 924 675 975 742 note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors reported inside parentheses. 15 Table 5 – Count-Data Estimates: Full Sample with High-Income Countries Dummy Interactions Civil Resistance (Count) Outcome Variable Consists of the Number of Strikes and Anti-Government Demonstrations Negative Binomial with Country Model Conditional-Mean Poisson Conditional-Mean Negative Binomial Dummies (1) (2) (3) (4) (5) (6) Perceptions of Economic Conditions 1.017** 1.501** 0.668 1.181* 1.076** 1.848** (0.493) (0.600) (0.468) (0.690) (0.544) (0.798) Perceptions of Economic Conditions x High-Income -3.136*** -3.076*** -1.195* -1.094 -2.160*** -2.234** Dummy (0.940) (0.923) (0.682) (0.845) (0.731) (0.967) Perceptions of Political Conditions 1.923*** 2.450*** 1.618*** 2.069** 2.022*** 1.977* (0.531) (0.893) (0.506) (0.849) (0.739) (1.010) Perceptions of Political Conditions x High-Income 1.724 2.777* 0.694 1.277 1.697 2.671* Dummy (1.248) (1.566) (0.816) (1.060) (1.068) (1.490) Overdispersion -0.974*** -1.050*** (0.127) (0.178) Control Variables NO YES NO YES NO YES Country Fixed Effects YES YES YES YES YES YES Year Fixed Effects YES YES YES YES YES YES Number of observations 924 675 924 675 975 742 note: *** p<0.01, ** p<0.05, * p<0.1. 16 B. Protest Contagion across Countries Protests in a country may be influenced by protests in neighboring countries. Arezki et al., (2020a) find that distance-weighted and social-media-influenced protests spillover across countries and affect the incidence of protests per capita (Arezki et al., 2020a). The estimation of distance-weighted protests is based on weighting national protests by the inverse distance between two countries’ capital cities. Thus, protests in capital cities that are geographically closer to one another will carry more weight than protests in cities that are further apart. The methodology employed is similar to Bown et al. (2017) which studied the impact of growth of nearby countries on an economy’s growth path. This could potentially be an omitted variable in our estimations. We thus re-estimate our base specification (Table 1, column 13) using the measure of distance-weighted protest contagion and present our results in Table 6 for both OLS and count data models. The coefficient for perceptions of political conditions is positive and at the 5% level across all specifications. Thus, our main findings stand regardless of whether we control for distance-weighted protest contagion. However, the coefficient of the effect is lower across all models than in the specification without the protest contagion variable. Furthermore, for the count data models, the effect of economic perceptions is positive and statistically significant, which constitutes a deviation from our main findings. We also explore the robustness of the results for the Eurobarometer interactions (Table 7) and high-income countries interactions (Table 8) when protest contagion is accounted for. For each of these tables, we estimate the base OLS regression without the contagion variable (column 1), the base OLS regression with the contagion variable (column 2), and the three count data models with the protest contagion variable (columns 3-5). The inclusion of the protest contagion variable severely weakens the interaction effects with Eurobarometer and high-income dummy variables as the coefficients are not only statistically insignificant, but the sign of the interactions is unstable across specifications. The contagion variable thus appears to capture the differential effects by geography and income. 17 Table 6 – OLS and Count-Data Estimates with Distance Weighted-Protest Contagion Variable Civil Resistance (Hyperbolic Inverse Sine of Civil Resistance (Count) Outcome Variable the Number off Strikes and (Consists of the Number of Strikes and Anti-Government Demonstrations) Anti-Government Demonstrations) OLS with Country and Conditional-Mean Negative Binomial with Model Conditional-Mean Poisson Year Fixed Effects Negative Binomial Country Dummies (1) (2) (3) (4) Perceptions of Economic Conditions 0.236 1.101** 0.863* 1.370*** (0.334) (0.522) (0.494) (0.493) Perceptions of Political Conditions 0.879*** 1.363** 1.695** 1.357** (0.340) (0.653) (0.725) (0.680) Inverse Distance-Weighted Protest Contagion 0.951*** 0.775*** 0.791*** 0.928*** (0.116) (0.105) (0.072) (0.084) Control Variables YES YES YES YES Country Fixed Effects YES YES YES YES Year Fixed Effects YES YES YES YES Number of observations 742 675 675 742 note: *** p<0.01, ** p<0.05, * p<0.1 18 Table 7 – OLS and Count-Data Estimates with Distance Weighted-Protest Contagion Variable and Eurobarometer Interactions Civil Resistance Civil Resistance (Count) Outcome Variable (Hyperbolic Inverse Sine of the Number off (Consists of the Number of Strikes and Anti-Government Strikes and Anti-Government Demonstrations) Demonstrations) Negative Binomial Conditional-Mean Conditional-Mean OLS with Country and Year Fixed Effects with Country Poisson Negative Binomial Dummies (1) (2) (3) (4) (5) Perceptions of Economic Conditions 0.926** 0.602* 1.157** 1.132* 1.568** (0.439) (0.361) (0.523) (0.673) (0.693) Perceptions of Economic Conditions x Euro Dummy -1.761*** -0.650 -0.704 -0.405 -0.612 (0.607) (0.553) (0.905) (0.958) (0.908) Perceptions of Political Conditions 0.953* 0.741** 2.027** 2.003** 1.454 (0.497) (0.317) (0.808) (0.886) (0.922) Perceptions of Political Conditions x Euro Dummy 2.332** 0.193 -1.504 -1.136 -0.406 (1.086) (0.711) (1.264) (1.210) (1.429) Distance-Weighted Protest Contagion Variable 0.943*** 0.772*** 0.803*** 0.914*** (0.117) (0.108) (0.074) (0.086) Ln of Alpha -1.905*** (0.277) Control Variables YES YES YES YES YES Country Fixed Effects YES YES YES YES YES Year Fixed Effects YES YES YES YES YES Number of observations 742 742 675 675 742 Adjusted R2 0.390 0.591 - - 0.324 note: *** p<0.01, ** p<0.05, * p<0.1. Robust Standard Errors Reported, clustered at the country level 19 Table 8 – OLS and Count Data Estimates with Distance Weighted-Protest Contagion Variable and High-Income Interactions Civil Resistance Civil Resistance (Count) Outcome Variable (Hyperbolic Inverse Sine of the Number off Strikes (Consists of the Number of Strikes and Anti-Government and Anti-Government Demonstrations) Demonstrations) Negative Binomial Conditional-Mean Conditional-Mean OLS with Country and Year Fixed Effects with Country Poisson Negative Binomial Dummies (1) (2) (3) (4) (5) Perceptions of Economic Conditions 0.618 0.460 1.241*** 0.862 1.447** (0.423) (0.330) (0.479) (0.695) (0.691) Perceptions of Economic Conditions x High-Income Dummy -1.212* -0.368 -0.388 0.062 -0.235 (0.632) (0.488) (0.881) (0.869) (0.871) Perceptions of Political Conditions 1.198** 0.838** 1.758** 2.038** 1.498 (0.492) (0.327) (0.819) (0.908) (0.922) Perceptions of Political Conditions x High-Income Dummy 1.541 -0.010 -0.968 -0.837 -0.415 (1.091) (0.638) (1.248) (1.077) (1.346) Distance-Weighted Protest Contagion Variable 0.949*** 0.763*** 0.803*** 0.924*** (0.117) (0.108) (0.073) (0.086) Overdispersion (log) -1.880*** (0.269) Control Variables YES YES YES YES YES Country Fixed Effects YES YES YES YES YES Year Fixed Effects YES YES YES YES YES Number of observations 742 742 675 675 742 Adjusted R2 0.383 0.590 - - 0.323 note: *** p<0.01, ** p<0.05, * p<0.1. Standard errors reported inside parentheses. Errors are clustered at the country level for OLS estimates. 20 C. Robustness Checks: Sample Coverage and Survey Harmonization Regarding the cross-country coverage of the estimation sample, the main estimations exclude data from the Arab Barometer. The survey instrument in the Arab Barometer does not include the political condition perception question we utilize from the other barometer surveys. As a robustness check, we use a comparable question “How would you evaluate the state of democracy and human rights in your country?” which is covered in waves 2 and 3 of the Arab Barometer.7 This same question has been used before as a proxy for the satisfaction in the way democracy works in the Global Barometer Survey (a collection of regional barometers). The share of “bad” and “very bad” responses to the question are included in our measure of political perception. Our main results are unchanged.8 Furthermore, there are two survey harmonization issues across the barometers that may affect our findings. The first is that some of the response options for the perception questions were changed across barometers. Some versions of the surveys allowed for neutral responses (neither bad nor good; neither dissatisfied nor satisfied) while others did not. As a robustness check, we re-estimated the baseline regressions by excluding neutral responses to the perception variables and only include responses with clear positive or negative perceptions. Our main findings are unchanged.9 The second harmonization challenge is with regards to survey weights. For some of the surveys (South Asian barometer and Eurobarometer), weights were not available. It is also unclear if weights were estimated consistently across surveys. In our baseline regressions, the perceptions variables were computed by utilizing weights whenever available. As a robustness check, we re-estimate our baseline regressions without weights. The main results are unaffected.10 D. Instrumental Variables Estimations Civil unrest can in of itself alter perceptions, giving rise to simultaneity bias. To tackle potential reverse-causality bias, we employ an instrumental variables (IV) approach. We first use the 1- year lagged economic and political perception variables as well as the 1-year lagged GDP per capita growth variable as instruments. We expect that lagged levels of GDP per capita growth and economic and political perceptions are correlated with contemporaneous perceptions as previous negative perceptions and economic conditions are likely to affect contemporaneous 7 The additional country-year observations resulting from the Arab Barometer are Algeria 2011, the Arab Republic of Egypt 2011, Iraq 2011, Iraq 2013, Jordan 2010, Jordan 2013, Kuwait 2014, Lebanon 2010, Lebanon 2013, Saudi Arabi 2011, Sudan 2010, the Republic of Yemen 2011, and the Republic of Yemen 2013. Overlapping country-year observations between the Arab Barometer and the Afro Barometer (e.g., North African countries) were accounted for as part of the latter. 8 Results are not reported but available upon request. 9 Results are not reported but available upon request. 10 Results are not reported but available upon request. 21 perceptions. However, protests are likely to be fueled by immediate perceptions and conditions that are a consequence of pent up frustration of previous years. Thus, the instruments are unlikely to violate the exclusion restriction criteria. Standard tests of the validity of the IVs are presented below. In another iteration of IV regressions, we use an expanded set of instruments by also introducing economic and political perception contagion variables. These variables are inverse-distance-weighted measures of economic and political perceptions in all countries included in our baseline sample. We expect that such perceptions in both neighboring and distant countries are correlated with domestic perceptions in the home country as negative perceptions could spread across-countries (e.g., the Arab Spring). It is noteworthy that in our sample we do not find that domestic protests are directly influenced and solely fueled by perception contagion from other countries. Here again, the two additional instruments are unlikely to violate the exclusion restriction criteria. Table 9 presents TSLS-IV estimates using both a set of three instruments and five instruments (including economic and political perception contagion). We also account for the protest contagion variable as an additional control. The findings are consistent with our main results – a perceived lack of voice in the political system leads to more protests whereas negative perceptions of economic conditions have no statistically significant effect. In the first stage, the results reported in columns 3 and 4 show that the lagged economic perceptions variable is significantly and positively correlated with the contemporaneous economic perceptions. The same is true for the political perceptions variable. But the estimated effect of the lagged economic perceptions on political perceptions is positive with a small coefficient and is only significant at the 10% level. The estimated impact of lagged political perceptions on economic perceptions is actually negative but significant. The first stage results reported in columns 5 and 6 of Table 9 include the estimates of perceptions contagion across borders, in addition to the GDP per capita growth and the lagged own perceptions. The evidence suggests that there is significant international contagion of perceptions, with economic perceptions in nearby countries significantly affecting economic perceptions. The same applies to political perceptions. In both cases, the estimated elasticities of international contagion of perceptions are greater than one, indicating that there might a magnification effect. It is also noteworthy that GDP per capita growth rates are not robust predictors of perceptions. In specification 3 and 4, growth appears with a negative sign, as expected, but it is statistically significant at only the 10% level. However, in specifications 5 and 6, growth has no discernable impact on either economic or political perceptions, after controlling for international contagion of perceptions. 22 Perhaps more importantly, the instruments pass the under-identification test, indicating they are correlated with the perception variables and thus relevant. The instruments also pass the weak identification test using the Stock and Yogo thresholds, indicating they are strong instruments in the sense that they explain a large share of the sample variance of the perception variables. And the relevant F-stat is larger in the first-stage specification that includes the variables that capture international contagion of perceptions. The instruments are also valid in the sense that they are sequentially not correlated with the regression errors, as reflected in the high p- values of the Sargan test for overidentification. Finally, Table 10 reports IV estimates with the conditional-mean Negative Binomial estimator.11 As discussed above, count-data estimates of the coefficient of interest tend to be larger than their OLS counterparts. This is also the case for the IV estimates. The Negative Binomial IV estimates of the elasticity of protests with respect to political perceptions is about 4.5, which is more than double the TSLS estimates reported in Table 9. The TSLS elasticity of protests with respect to protest contagion, however, is slightly larger than the count-data IV estimate (0.92 versus 0.78). Overall, political perceptions along with contagious protests tend to be robust predictors of the number of protests across countries and over time, regardless of estimator, control variables, and sample composition. In contrast, objective economic indicators appear not to play a significant role in predicting the incidence of social unrest, neither as direct determinants nor as indirect determinants via their lagged effect on perceptions. 11 The Negative Binomial estimator with country dummies did not converge with the reduced sample of the IV specification. However, as discussed above, the conditional-mean Negative Binomial estimator tends to yield coefficient estimates that are close to the alternative estimator and both tend to be larger than OLS estimates with the transformed dependent variable of civil unrest. Also, Table 11 does not report IV Negative Binomial results with the expanded set of instruments. 23 Table 9 – OLS and TSLS Estimates with Protest Contagion Variable Civil Civil Perception of Perception of Civil Perception of Perception of Outcome Variable Resistance Resistance Economic Political Resistance Economic Political (Transformed) (Transformed) Conditions Conditions (Transformed) Conditions Conditions OLS Regression Second Stage First Stage First Stage Second Stage First Stage First Stage (1) (2) (3) (4) (5) (6) (7) Perceptions of Economic Conditions 0.303 0.164 -0.051 (0.473) (0.469) (0.364) Perceptions of Political Conditions 1.288** 2.377** 1.890*** (0.517) (1.015) (0.552) Perceptions of Economic Conditions (1-Year Lag) 0.661*** 0.068* 0.406*** 0.059* (0.040) (0.035) (0.038) (0.030) Perceptions of Political Conditions (1-Year Lag) -0.151** 0.370*** -0.112** 0.186*** (0.060) (0.052) (0.051) (0.041) GDP Per Capita Growth (1-Year Lag) -0.003* -0.001 -0.000 0.000 (0.002) (0.001) (0.001) (0.001) Economic Condition Perception Contagion 1.357*** -0.204** (0.111) (0.088) Political Condition Perception Contagion -0.129 1.574*** (0.126) (0.100) Distance-Weighted Protest Contagion 0.934*** 0.905*** 0.002 0.022*** 0.915*** -0.010 0.009 (0.150) (0.065) (0.008) (0.007) (0.060) (0.007) (0.005) Controls YES YES YES YES YES YES YES Number of observations 468 468 468 468 468 468 468 Under-Identification Test (p-value) 0.0000 0.0000 Weak Identification Test (Cragg-Donald Wald F Statistic) 26.225 91.820 Stock-Yogo weak ID test critical values: 10% maximal IV size 13.43 19.45 15% maximal IV size 8.18 11.22 20% maximal IV size 6.40 8.38 25% maximal IV size 5.45 6.89 Over-Identification Test (Chi-Square p-value for Sargan Statistic) 0.348 0.292 R2 0.645 0.762 0.765 Adjusted R2 0.619 0.715 0.718 note: *** p<0.01, ** p<0.05, * p<0.1. Robust standard errors reported. 23 Table 10 -- IV Conditional-Mean Negative Binomial Estimates with Protest Contagion Variable Civil Resistance Perceptions of Perceptions of Political Outcome Variable (counts) Economic Conditions Conditions Second Stage First Stage First Stage (1) (2) (3) Predicted Perceptions of Economic Conditions 0.795 (1.013) Predicted Perceptions of Political Conditions 4.547** (2.176) Distance-Weighted Protest Contagion 0.781*** -0.003 0.022*** (0.102) (0.010) (0.006) Perceptions of Economic Conditions (1-Year Lag) 0.619*** 0.068* (0.043) (0.038) Perceptions of Political Conditions (1-Year Lag) -0.129** 0.370*** (0.060) (0.058) GDP Per Capita Growth (1-Year Lag) -0.003** -0.001 (0.002) (0.002) Year Fixed Effects YES YES YES Country Fixed Effects YES YES YES Number of observations 450 499 468 F Stat (First Stage) 155 211 note: *** p<0.01, ** p<0.05, * p<0.1 24 V. Conclusions This paper established a robust partial correlation between citizens’ perceptions of political conditions and the incidence of strikes and anti-government demonstrations. More specifically, the evidence indicates that a perceived lack of voice in the political system in a country translates into nonviolent uprisings on the ground. That is not necessarily the case for negative perceptions of economic conditions. Furthermore, the international contagion of protests is a strong predictor of civil unrest. These findings hold across several estimation models and robustness checks. A regional perspective shows that although these results hold holistically, they are flipped for countries outside Western and Central Europe: adverse economic circumstances are more pressing than the perceived lack of voice in the political system and are a driver of nonviolent uprisings. On the other hand, a look at subsamples by levels of income showed that civil unrest in high-income countries is driven by adverse perceptions of political conditions only. In developing economies, both negative economic and political condition perceptions seem to be drivers of such unrest. However, these findings of heterogeneous coefficients are not robust to changes in estimation models and specifications. That is, after controlling for international contagion of civil unrest, the statistical significance of this international heterogeneity disappears whereas the impact of dissatisfaction with lack of voice in the political system appears to be a robust central tendency in the data. One could speculate that in high- income economies and Europe, where information is transmitted freely, the effects of international contagion of protests may dominate. However, even for the whole sample, accounting for the international contagion of protests reduces the size of the estimated elasticity of civil unrest with respect to political perceptions. These results provide insights on citizen motivations to take to the streets in times when protests and uprisings are commonplace. Indeed, the introduction of opinion surveys and subjective perception metrics to the analysis of civil uprisings allows for a better understanding of the perceptions that may shape them. This has important implications for governments. Reforming policies that can improve economic conditions might be insufficient to quell civil unrest. Communicating such reforms to the public and enhancing transparency may be critical in shaping perceptions, which in turn may influence the incidence and likelihood of protests. An area for future research would therefore consist of exploring which factors influence the gaps between citizen perceptions and reality, particularly with respect to the extent of democratic governance. 25 VI. References Abu Bader, Suleiman and Elena Ianchovichina (2019). “Polarization, Foreign Military Intervention, and Civil Conflict.” Journal of Development Economics 141, 102248 https://doi.org/10.1016/j.jdeveco.2018.06.006 Afrobarometer Data, All Countries, Rounds 1-7 (1999-2008). http://www.afrobarometer.org Allison, Paul and Richard Waterman (2002). “Fixed-Effects Negative Binomial Regression Models.” Sociological Methodology, 32(1), 247–265. https://doi.org/10.1111/1467-9531.00117 Arab Barometer Data, All Countries, Waves 1-5 (2006-2019). http://www.arabbarometer.org Arezki, Rabah and Markus Bruckner (2011). “Food Prices and Political Instability”. IMF Working Papers 2011/062, International Monetary Fund. Arezki, Rabah, Alou Adesse Dama, Simeon Djankov, and Ha Nguyen (2020a). Contagious Protests. Policy Research Working Paper; No. 9321. World Bank, Washington, DC Arezki, Rabah, Daniel Lederman, Amani Abou Harb, Nelly El-Mallakh, Rachel Yuting Fan, Asif Islam, Asif; Ha Nguyen, and Marwane Zouaidi (2020b). Middle East and North Africa Economic Update, April 2020: How Transparency Can Help the Middle East and North Africa. Washington, DC: World Bank. Banks, Arthur and Kenneth Wilson (2020). Cross-National Time-Series Data Archive. Databanks International. Jerusalem, Israel. Accessed through https://www.cntsdata.com Braha, Dan (2012). “Global Civil Unrest: Contagion, Self-Organization, and Prediction.” PLoS ONE 7(10): e48596. https://doi.org/10.1371/journal.pone.0048596 Bown, Chad P., Daniel Lederman, Samuel Pienknagura, Raymond Robertson (2017). Better Neighbors: Toward a Renewal of Economic Integration in Latin America. Latin America and Caribbean Studies; Washington, DC: World Bank. Cammett, Melani and Ishac Diwan (2013). The Political Economy of the Arab Uprisings. Westview Press. Chenoweth, Erica and Kathleen Gallagher Cunningham (2013). “Understanding Nonviolent Resistance: An Introduction.” Journal of Peace Research 50(3): 271–276. Chenoweth, Eric and Maria Stephan (2008). “Why Civil Resistance Works: The Strategic Logic of Nonviolent Conflict.” International Security 33(1): 7–44. Chenoweth, Erica and Jay Ulfelder (2017). “Can Structural Conditions Explain the Onset of Nonviolent Uprisings?” Journal of Conflict Resolution 61(2), 298–324. Davenport, Christian (1995). “Multi-Dimensional Threat Perception and State Repression: An Inquiry into Why States Apply Negative Sanctions.” American Journal of Political Science. 39(3):683–13 26 Eurobarometer Data, All Countries (2005-2019). https://www.gesis.org/en/eurobarometer-data- service/search-data-access/data-access Gibney, Mark, Linda Cornet, Reed Wood, Peter Haschke, Daniel Arnon, Attilio Piano, Gray Barrett; (2019). The Political Terror Scale 1976-2018. Accessed through the Political Terror Scale website: http://www.politicalterrorscale.org/ Hu, Fu and Yun-han Chu, (2001-2021). Asian Barometer Survey Project. Center for East Asia Democratic Studies. http://www.asianbarometer.org LatinoBarómetro Data, All Countries (1995-2018). https://www.latinobarometro.org/ Polity5 Annual Time-Series (1946-2018). Center for Systemic Peace. http://www.systemicpeace.org/inscrdata.html Witte, Caroline T., Martijn J. Burger and Elena Ianchovichina (2020). “Subjective Well-Being and Peaceful Uprisings.” KYKLOS 73(1): 120-158. WDI. (2019). World Development Indicators. http://data.worldbank.org/ WGI. (2019). Worldwide Governance Indicators. http://govindicators.org 27 Appendix Appendix 1A – Variables Definitions and Sources Name Definition Source Civil Resistance Measured as the Sum of Strikes and Anti- Computed Using Cross-National Time- Civil Resistance Government Demonstrations Transformed Using the Series Data Archive on National Domestic (Transformed) Hyperbolic Inverse Sine Conflict (2020) Perception of Economic Conditions Defined as the Perception of Computed Using Perception Data from Percentage of Non-Neutral Citizens Who Think Current Economic Conditions Regional Barometers Economic Conditions Are Bad or Very Bad Perception of Political Conditions Defined as the Perception of Political Computed Using Perception Data from Percentage of Non-Neutral Citizens Who Are Dissatisfied Conditions Regional Barometers or Very Dissatisfied with Democracy in Their Countries Log of GDP per Computed Using World Development Ln of GDP per Capita (Constant 2010 US$) Capita Indicators GDP per Capita GDP Per Capita Growth (Annual %) World Development Indicators Growth Log of Total Ln of Total Population World Development Indicators Population Inflation Inflation (CPI) (Annual %) World Development Indicators Total Unemployment (% of Total Labor Force) (Modeled Unemployment Rate World Development Indicators ILO Estimate) Oil Rents Oil Rents (% of GDP) World Development Indicators Infant Mortality Rate Infant Mortality Rate (Per 1,000 Live Births) World Development Indicators Cellular Subscription Mobile Cellular Subscriptions (Per 100 People) World Development Indicators Urban Population Urban Population (% of Total Population) World Development Indicators Center for Systemic Peace Data (Polity5 Autocracy Metric Institutionalized Autocracy Score Project) Average of the Three Indicators of Political Terror Political Terror Scale Computed Using Data from the Political (Amnesty International, Human Rights Watch, US Metric Terror Scale Department of State) Average of the Six World Governance Indicators Sub- Indexes (Voice and Accountability, Political Stability and World Governance Computed Using Data from the World Absence of Violence and Terrorism, Government Indicators Metric Governance Indicators Effectiveness, Regulatory Quality, Rule of Law, Control of Corruption) 28 Appendix 1B – Descriptive Statistics Variable Observations Mean St. Dev. Min Max Civil Resistance (Transformed) 742 0.773165 1.061435 0 4.219724 Perception of Economic Conditions 742 0.523706 0.24732 0.018132 0.995003 Perception of Political Conditions 742 0.521703 0.2016 0.058116 0.931941 Ln of GDP per Capita (Constant 2010 US$) 742 8.991888 1.386575 5.475694 11.62597 GDP Per Capita Growth (Annual %) 742 2.192606 3.439248 -22.3123 23.98551 Ln of Total Population 742 16.32672 1.382145 13.05013 21.03897 Inflation (CPI) (Annual %) 742 4.781616 6.144443 -4.4781 96.09411 Total Unemployment (% of Total Labor Force) 742 8.010958 5.525565 0.393 35.268 Oil Rents (% of GDP) 742 1.034778 2.972693 0 23.81619 Infant Mortality Rate (Per 1,000 Live Births) 742 19.1779 20.74064 1.5 110.9 Institutionalized Autocracy Score 742 0.292453 1.029614 0 9 Average of Political Terror Scale Indicator 742 2.128257 0.996237 1 5 Mobile Cellular Subscriptions (Per 100 People) 742 88.80605 46.09214 0.024533 182.4573 Urban Population (% of Total Population) 742 64.0512 18.8361 11.194 97.961 Average World Governance Indicators 742 0.332027 0.836041 -1.41226 1.893856 29 Appendix 1C – Correlation Matrix for Baseline Sample Variable (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Civil Resistance (Transformed) 1 Perception of Economic Conditions 0.08** 1 Perception of Political Conditions 0.21*** 0.50*** 1 Ln of GDP per Capita 0.07*** -0.01 -0.37*** 1 GDP Per Capita Growth (Annual %) -0.05*** -0.28*** -0.11*** 0.03*** 1 Ln of Total Population 0.37*** -0.04 0.12*** -0.22*** 0.01 1 Inflation (CPI) (Annual %) 0.02* 0.04 0.13*** -0.04*** -0.05*** 0.03** 1 Total Unemployment (% of Labor Force) -0.01 0.39*** 0.16*** 0.10*** -0.01 -0.17*** -0.01 1 Oil Rents (% of GDP) -0.08*** -0.16*** 0.12*** 0.11*** 0.03** 0.01 -0.01 0.01 1 Infant Mortality Rate (/1000) -0.14*** -0.04 0.15*** -0.75*** -0.05*** 0.08*** 0.04*** -0.13*** 0.03*** 1 Institutionalized Autocracy Score -0.19*** -0.09*** -0.02 -0.33*** -0.02* -0.14*** -0.01 -0.09*** 0.38*** 0.47*** 1 Average of Political Terror Scale Indicator 0.19*** -0.04 0.40*** -0.51*** -0.05*** 0.46*** 0.06*** -0.03** 0.12*** 0.39*** 0.29*** 1 Mobile Cellular Subscriptions (/100) 0.21*** 0.00 -0.02 0.39*** 0.00 0.04*** -0.05*** 0.00 0.00 -0.45*** -0.28*** -0.16*** 1 Urban Population (% of Population) 0.15*** -0.16*** -0.15*** 0.83*** 0.00 -0.12*** -0.01 0.08*** 0.15*** -0.67*** -0.29*** -0.29*** 0.32*** 1 Average World Governance Indicators -0.05*** -0.04 -0.52*** 0.81*** -0.04*** -0.31*** -0.22*** 0.02 0.27*** -0.67*** -0.39*** -0.74*** 0.43*** 0.52*** 1 note: *** p<0.01, ** p<0.05, * p<0.1. 30 Appendix 2A – Tabulation of Sample Observations by Year Year Number of Obs. 1996 14 1998 14 2000 18 2002 23 2003 25 2004 15 2005 61 2006 48 2007 45 2008 33 2009 45 2010 47 2011 48 2012 41 2013 59 2014 44 2015 60 2016 44 2017 58 Total 742 Appendix 2B – Tabulation of Sample Observations by Barometer Barometer Number of Obs. 13 (Inexact match; robustness check) Arab Barometer 12 AfroBarometer 126 Latinobarómetro 256 Eurobarometer 318 Asian Barometer 32 South Asian Barometer 10 Total 742 12 Not included in the sample size of 742 country-year observations 31 Appendix 2C – Tabulation of Sample Observations by Country-Year Total 2000 2002 2003 2004 2005 2006 2007 2008 2009 1996 1998 2010 2012 2013 2014 2015 2016 2017 2011 Country Algeria X X 2 Austria X X X X X X X X X X X X 12 Bangladesh X X 2 Belgium X X X X X X X X X X X X 12 Bolivia X X X X X X X X X X X X X X X X X 17 Botswana X X X X X X 6 Brazil X X X X X X X X X X X X X X X X X 17 Bulgaria X X X X X X X X X X X X 12 Burkina Faso X X X X 4 Burundi X X 2 Cambodia X X X 3 Cameroon X X 2 Chile X X X X X X X X X X X X X X X X X 17 China X X X X 4 Colombia X X X X X X X X X X X X X X X X X 17 Costa Rica X X X X X X X X X X X X X X X X X 17 Croatia X X X X X X X X X X 10 Cyprus X X X X X X X X X X X X 12 Czech Republic X X X X X X X X X X X X 12 Denmark X X X X X X X X X X X X 12 Ecuador X X X X X X X X X X X X X X X X X 17 Egypt, Arab Rep. X X 2 El Salvador X X X X X X X X X X X X X X X X X 17 Estonia X X X X X X X X X X X X 12 Eswatini X X 2 Finland X X X X X X X X X X X X 12 France X X X X X X X X X X X X 12 Gabon X X 2 Germany X X X X X X X X X X X X 12 Ghana X X X X X X 6 Greece X X X X X X X X X X X X 12 Guatemala X X X X X X X X X X X X X X X X X 17 Guinea X X X 3 Honduras X X X X X X X X X X X X X X X X X 17 Hungary X X X X X X X X X X X X 12 India X X 2 Indonesia X X X 3 Ireland X X X X X X X X X X X X 12 Italy X X X X X X X X X X X X 12 Japan X X X X 4 Kenya X X X X X X 6 Korea, Rep. X X X X 4 Latvia X X X X X X X X X X X X 12 Lesotho X X X X X X X 7 Liberia X X X 3 Lithuania X X X X X X X X X X X X 12 Luxembourg X X X X X X X X X X X X 12 32 Madagascar X X X X 4 Malawi X X X X X X 6 Malaysia X X X 3 Mali X X X X X X 6 Mauritius X X X 3 Mexico X X X X X X X X X X X X X X X X X 17 Mongolia X X X X 4 Morocco X X 2 Mozambique X X X X 4 Namibia X X X X X X 6 Nepal X X 2 Netherlands X X X X X X X X X X X X 12 Nicaragua X X X X X X X X X X X X X X X 15 Niger X X 2 Nigeria X X X X X X X 7 North Macedonia X X X 3 Pakistan X X 2 Panama X X X X X X X X X X X X X X X X X 17 Paraguay X X X X X X X X X X X X X X X X X 17 Peru X X X X X X X X X X X X X X X X 16 Philippines X X X X 4 Poland X X X X X X X X X X X X 12 Portugal X X X X X X X X X X X X 12 Senegal X X X X X X 6 Sierra Leone X X 2 Slovak Republic X X X X X X X X X X X X 12 Slovenia X X X X X X X X X X X X 12 South Africa X X X X X X 6 Spain X X X X X X X X X X X X 12 Sri Lanka X X 2 Sweden X X X X X X X X X X X X 12 Tanzania X X X X X X 6 Thailand X X X 3 Togo X X X 3 Turkey X X X X X 5 Uganda X X X X X X X 7 United Kingdom X X X X X X X X X X X X 12 Uruguay X X X X X X X X X X X X X X X X X 17 Venezuela, RB X X X X 4 Zambia X X X X X X 6 Zimbabwe X X X 3 Total 14 14 18 23 25 15 61 48 45 33 45 47 48 41 59 44 60 44 58 742 33