Measuring Social Unrest Using Media Reports
  • 1 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund
  • | 2 0000000404811396https://isni.org/isni/0000000404811396International Monetary Fund

Contributor Notes

We present a new index of social unrest based on counts of relevant media reports. The index consists of individual monthly time series for 130 countries, available with almost no lag, and can be easily and transparently replicated. Spikes in the index identify major events, which correspond very closely to event timelines from external sources for four major regional waves of social unrest. We show that the cross-sectional distribution of the index can be simply and precisely characterized, and that social unrest is associated with a 3 percentage point increase in the frequency of social unrest domestically and a 1 percent increase in neighbors in the next six months. Despite this, social unrest is not a better predictor of future social unrest than the country average rate.

Abstract

We present a new index of social unrest based on counts of relevant media reports. The index consists of individual monthly time series for 130 countries, available with almost no lag, and can be easily and transparently replicated. Spikes in the index identify major events, which correspond very closely to event timelines from external sources for four major regional waves of social unrest. We show that the cross-sectional distribution of the index can be simply and precisely characterized, and that social unrest is associated with a 3 percentage point increase in the frequency of social unrest domestically and a 1 percent increase in neighbors in the next six months. Despite this, social unrest is not a better predictor of future social unrest than the country average rate.

1 Introduction

Social unrest is a major social issue in many countries across the world. During just the second half of 2019, major protests or other forms of disorder occurred in locations as diverse as Bolivia, Chile, France, Hong Kong, India, Iraq, and Lebanon. Such unrest has a natural connection to economic questions, as economic factors may contribute to social unrest. For example, increases in taxes and fuel prices were important triggers for recent protests in Lebanon and Iran, respectively. Moreover, disruption stemming from social unrest can have economic consequences, potentially interrupting trade and dissuading investment.

Of course, answering questions about the relationship between social unrest and economic or financial outcomes necessitates data on social unrest. Without knowing, at the very least, where and when major social unrest events have occurred, it is hard to make progress on such issues. Yet data on social unrest are largely unsatisfactory. There is – as far as we know – no transparent, high-frequency, timely indicator of social unrest with broad and consistent coverage across countries and periods. This paper is an attempt to fill that gap.

We introduce a new measure of social unrest, as measured by media reports, which we term the Reported Social Unrest Index (RSUI).1 Coverage is broad: we provide data on 130 countries from January 1985 to May 2020. Measurement criteria are consistent and transparent: we use the same set of sources and search terms throughout, and any country-specific adjustments are reported and justified. The index is high-frequency and timely: data are available for each month with only a few days lag. This makes the index useful not only for research but also for contemporaneous monitoring of cross-country social unrest episodes.

In this paper, we detail how we construct the RSUI and how large movements can be used to identify major events. By comparing our results to several case studies of major, well-known social unrest episodes, we argue that the index and the coded events reflect respected narrative descriptions of real events, and are not simply the product of media fads or biases. We also analyze the statistical properties of the index and present three key findings, all of which are essential context for future work. First, that the units of the RSUI have a natural interpretation, consistent across and within countries: that a one percent increase in the index reduces the fraction of higher observations by approximately two percent. Second, that social unrest events are typically associated with around a three percentage point increase in the probability of social unrest in the same country and a one percentage point increase in neighboring countries during the next six months. Third, that despite this correlation, past social unrest is not a good predictor of future unrest, which we interpret as reflecting the fact that social unrest is a low-probability event driven by disparate factors.

The next section discusses related economic literature. Section 3 outlines the calculation of the RSUI and how we identify major social unrest events from it. Section 4 discusses the internal and external consistency of our measures. And section 5 investigates the properties of the time-series and cross-country variation of this index.

2 Related Literature

Our work contributes to the rapidly growing literature on text search methods using newspaper archives. Initially, this literature focused on measuring economic policy uncertainty. The seminal paper is by Baker et al. (2016), who construct an index of economic policy uncertainty (EPU) for 12 major economies using newspaper archives back to 1985. They show that the EPU is associated with lower investment and employment, and higher stock price volatility. Similarly, Ahir et al. (2018) construct a World Uncertainty Index (WUI) for 143 individual countries on a quarterly basis from 1996 onwards. Their index measures the frequency of the word “uncertainty” in the quarterly Economist Intelligence Unit country reports. And in a country-specific setting, Jirasavetakul and Spilimbergo (2018) develop a news-based economic policy uncertainty (EPU) index for Turkey. The index measures the frequency of news articles about economic policy uncertainty and, as we do, uses Factiva as the primary source.

Other authors use text-based indices proxy for sentiment in financial markets, geopolitical tensions, and corruption. Manela and Moreira (2017) calculate a text-based measure of uncertainty starting in 1890 using front-page articles of the Wall Street Journal. They show that periods when their index is high are followed by periods of above average stock returns. To examine the effects that media sentiment has on equity prices, Fraiberger et al. (2018) create a daily news-based sentiment index for 25 advanced and emerging economies between 1991 and 2015. They restrict their sample to articles published by Reuters in English, and develop an algorithm which quantifies tone, counting the number of positive and negative words within financial, political and economic texts. Caldara and Iacoviello (2018) construct a monthly measure of geopolitical risk based on a set of newspaper articles covering geopolitical tensions since 1985. Like us, these authors use an algorithm which counts the frequency of articles that refer to geopolitical risks in leading newspapers published in the US, UK, and Canada, although in their case ProQuest Historical Newspapers and ProQuest News stream are the primary sources. Finally, Hlatshwayo et al. (2018) construct a cross-country news-based flow indexes of corruption and anti-corruption by reviewing 665 million international news articles provided by Factiva. They show that shocks in the corruption index are associated to negative impacts on asset prices and economic activity.

Clearly, our work is closely related to other measures of social unrest. There are three main alternatives to our approach: the Cross-National Time-Series Data (CNTSD) database by Banks and Wilson (2020); the Armed Conflict Location and Event Database (ACLED); and the Mass Mobilization in Autocracies Database (MMAD) by Hellmeier et al. (2019) are alternatives to measure social unrest worldwide. The temporal and spatial coverage of the CNTSD is comprehensive – it provides annual time series data since 1815 covering 200 countries for the number of riots and anti-government demonstrations. Yet despite its wide usage, the CNTSD suffers from three significant drawbacks. First, it is at annual frequency and is typically updated with a lag of at least a year. Second, it identifies only a small number of events per year so the marginal impact of a new event can be substantial. Third, the results can be somewhat hard to interrogate, with the relationship between increases in the CNTSD series and events on the ground not always compelling (see Appendix C for further illustration of this point).

The ACLED is a high-quality source with monthly observations. Within sub-Saharan Africa, coverage is broad and begins in the late-1990s. Coverage elsewhere, though, is rather limited. Finally, the MMAD has information about individual protest events generated by a combined machine learning and human coding process from three newswire agencies: The Associated Press, the Agence France-Presse and BBC Monitoring. However, the covers only a few countries from 2005 to 2012.

Several articles have used the CNTSD and the MMAD to proxy for social unrest in crosscountry statistical analysis of political and economic phenomena, and a few articles have exploited the ACLED’s more granular data in the context of Sub-Saharan Africa. The political science literature used these data extensively to study social unrest and related political issues in a crosscountry setting –see for example, Bodnaruk Jazayeri’s (2016) analysis of the impact of identity-based political inequality on protest in MENA using CNTSD, or Ciocan and Wu¨est (2016) for an application of the MMAD to analysis of media censorship in MENA countries during the Arab Uprisings of 2011. The economics literature tended to use CTSD more frequently for cross-country econometric analysis of long-run issues – see for example, Acemoglu et al. (2019), who use this data to explore the relationship between democracy and economic growth, or Barro’s (1991) classic paper on economic growth where variables from this dataset are used as controls for political instability. Some more recent studies look at higher frequency economic phenomena, such as empirical analysis on budget cuts, and to social unrest in Europe by Ponticelli and Voth (2020). The political science and economics literature has also worked with ACLED focusing on Sub-Saharan Africa and benefiting from its detailed coverage –see for example, Harari and Ferrara (2018), and Ali et al. (2019).

3 The Reported Social Unrest Index

In this section we explain how the RSUI is created and a method for coding peaks into major events. We discuss how concerns for robustness influence the design of the methodology and report basic properties of the index.

3.1 Data

The primary source is Dow Jones’ Factiva news aggregator. We restrict our sample to printed articles published by major English-language newspapers and networks in the USA, UK, and Canada. Specifically: the ABC Network, the BBC, the CBS Network, the Canadian Broadcasting Corp, the NBC Network, the Los Angeles Times, the Financial Times, the Boston Globe, the Globe and Mail, the New York Times, the Telegraph U.K., the Times U.K., the Chicago Tribune, the Telegraph, the Guardian U.K., the Wall Street Journal, the Washington Post, and the Economist. This is a very similar set of sources to Caldara and Iacoviello (2018).

This choice of sources is motivated by several factors. First, this produces a very large sample over a long period of time, with our headline measure of relevant articles rising from around 7,000 per month in 1985 to over 20,000 between 2000 and the present, totaling over 7 million articles (see Figure 1). Second, by selecting well-known sources, we can be completely transparent about the possible biases in our sources. Of course, no sources are entirely free of bias, but these sources are ones that a large set of users is familiar with, and so can caveat the results as they see fit. This is not true about either less well-known sources or newer media, such as Twitter or Facebook, where one cannot easily understand the likely biases of individual commenters. This latter group of sources suffer from further challenges, as it is difficult to know how to account for changes in penetration and composition or the role of bots, trolls, and other nefarious actors. Similarly, by limiting our sample to only foreign news media, we hope to avoid the most nakedly political sources, as within-country publications might have specific agendas which influence their reporting of civil unrest.

Figure 1:
Figure 1:

Number of contemporary articles per month, zt

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Nevertheless, in Section 4.2 we also provide further robustness checks, examining the possible impact of these arguments for a selection of Middle Eastern countries, including using different English-language sources, alternate search terms, and French and Arabic language media. And when constructing the RSUI from the raw data, we also take account of how possible media bias might be mitigated by the specific details of the design of the index, discussed further below.

Using these sources, we collect three monthly article counts:

  • xit : Number of articles about social unrest in country i at period t

  • yit : Number of contemporary articles in country i at period t

  • Zt : Number of contemporary articles in period t

Our sample includes the 130 countries countries with at least one million inhabitants and 200 social unrest articles (i.e. Σtxit ≥ 200).

The search criteria defining these three series vary principally in the text strings tha must match in order for an article to be counted (see Table 1). This is most involved for xit, which has both inclusive and exclusive requirements. The inclusive requirements aim to pick up specific events related to civil unrest events, including protests, riots, major demonstrations, and other forms of unrest. The exclusive requirements aim to prevent false positives. These are either country-specific (more on this in Section 3.4), or aim to prevent matches from mis-use of search terms, 2 or articles related to commemorations of past unrest episodes.3 As the series yit and zt will be used principally to normalize the article count, we follow Jirasavetakul and Spilimbergo (2018) and add a further filter: that articles include the common and neutral term “today”.4

Table 1:

Article search Criteria

article image

We also add restrictions on location, subject, and word count on the articles. The first two of these are implemented via category tags provided by Factiva. The location tag helps to screen some articles about civil unrest directed at a given country rather than in a given country, such as foreign-policy protests. However, this is not always successful (see section 3.4 below).

Table 2 presents key summary statistics of the data and Figure 2 shows the cross-country distributions of articles. The cross-country distribution is highly skewed. For example, although the median country is featured in an average of 116 articles per month, the average across countries is 350. The within-country skew is also very large, with the median country attaining a peak of almost 900 articles per month. In general, the relative volatility of the social unrest series xit is much greater than the the country total count yit. For example, the median country has an average of three social unrest articles per month, but a maximum of 75. This suggests that there is considerable scope for large movements in the data which sharply discriminate between different observations.

Table 2:

Monthly article counts, summary statistics

article image
Figure 2:
Figure 2:

Cross-country average monthly article counts

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

3.2 Constructing the Index

From the raw article counts, we create two indices:

RSUIitA=xit112Σj=112ztj×100x¯i/z¯RSUIitB=xit112Σj=112yi,tj

Where x¯iandz¯ are the (country i) averages of xit and zt respectively over all time periods. That is:

x¯i=1TΣt=1Txitz¯=1TΣt=1Tzt

We call the two RSUI measures, unsurprisingly, the A index and the B index. These two indices have complementary strengths and weaknesses. The advantage of the B index over the A index is that it has naturally interpretable units – it is the share of contemporary articles in a given country about social unrest. In contrast, the A index uses the fraction of all articles which are about unrest in country i. This makes cross-country comparisons difficult using the A index, as a given country may receive more coverage from our sources on average for reasons unrelated to social unrest, such as size, proximity, or common historical or cultural ties. To emphasize this point, we define the A index as being rebased to have mean 100. Of course, scaling by average total country coverage should correct for this, as average interest in country i affects both numerator and denominator. Indeed, this is exactly what we do when constructing the B index. Yet this approach fails, as we discuss next.

The B index suffers from two flaws which mean that we chose the A index as our primary measure. First, many countries have very little coverage outside of social unrest events. And so the denominator 112Σj=112zi,tj is often very close to zero, vastly amplifying any noise in the numerator, xit. Second, coverage of a given country is often endogenous to social unrest or recent social unrest. Coverage of a given country in general will likely respond to social unrest, as journalists follow up on the consequences of major unrest. While the one-period lag in the denominator (as j runs from 1 to 12 not 0 to 11) can help address contemporaneous endogeneity, this is insufficient for major events where interest may run high for several months.

In contrast, the A index suffers from neither of these drawbacks, and so we use it as our headline measure, presented as an index scaled to a mean of 100 within each country. However, because this measure contains no useful cross-country information, we will will also use the B index as a filter for false positives when coding events in Section 3.3.

Figure 3 shows the A and B measures for a sample of six Middle Eastern countries during and after the Arab Uprisings of 2011, scaled to average to 100 over the whole sample. This illustrates two important points. First, the behavior of the index appears to be characterized by large intermittent spikes. This motivates our secondary characterization of the series in terms of major unrest events in the next section. Looking forward slightly, this Figure includes the events and labels that we derive using the methodology in Section 3.3. The purpose at this point is not yet to explain the event coding in detail but instead simply to use a well-known episode of unrest to introduce the series and discuss the relative merits of the A and B indices.

Figure 3:
Figure 3:

Arab uprisings, June 2010 – December 2014: Major events identified from RSUI event coding

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The second point this figure highlights is the endogeneity problem for the B index, in that it appears to understate the persistence of social unrest. Major spikes in the A index during 2011 and early 2012, such as the Tahrir Square protests in Egypt and the Tunisian Parliamentary elections of 2011, are almost invisible when using the B index. This difference is entirely due to short-term changes in the divisor of the B index. Because coverage of the country as a whole also rises following social unrest events, the denominator series yu increases, pulling down the B index and resulting in its failure to capture secondary unrest events. This means that the B index will badly mis-state the persistence and predictability of social unrest events, meaning that it is not suitable as a primary measure of social unrest. Accordingly, all subsequent charts and references to the RSUI are used to refer to the A index only.

3.3 Events

We define a social unrest event as a period i in a country t matching three criteria:

  • 1. The A index is a local peak: RSUIitA>max(RSUIi,t1A,RSUIi,t+1A).

  • 2. The A index satisfies at least one of the following extremum criteria:

    • (a) RSUIitA is in the top 2% of observations for country i;

    • (b) RSUIitA exceeds the mean for country i by at least four times the standard deviation for country i;

    • (c) RSUIitA exceeds the rolling 20-year mean for country i by at least four times the rolling 20-year standard deviation for country i.

  • 3. The B index exceeds 0.1, RSUIitB>0.1

Each of these criteria aims to select or exclude particular types of events. The first two conditions aim to select unusually high values of the RSUI relative to the distribution of within-country outcomes. By including both a percentile and a standard deviation threshold we aim to ensure that the event coding is robust to cross-country variation in distribution of the A index. The inclusion of part c) addresses the possibility of changing media focus over time and means that we still identify events consistently even if a given level of unrest in a given country becomes more or less interesting to our sources over time.5 The design of conditions 1 and 2 also strips out much cross-country variation in media bias when coding events. Because they are entirely relative measures, they will identify exactly the same set of events whether the outlets we use are highly focused on a given country or pay it little attention.

Because they depend on the full distribution of within-country outcomes, conditions 1 and 2 work well when there are at least a few major social unrest events within a country. However, in some countries, social unrest is very rare. Yet in these cases condition 2 will still select at least 2 percent of observations as candidate events. Thus we add condition 3, which leverages the natural units of the B index to impose a very small amount of cross-country information. This is a sufficiently weak restriction that major events are still identified despite the flaws in the B index, yet still still acts to filter out false positives which would otherwise be produced in countries with very little social unrest.

Of course, the specifics of how series and events are computed are somewhat arbitrary. To a large extent this is inevitable; there is no way to pick thresholds and measures for extreme points which is not arbitrary. So rather than seeking to defend these choices a priori, we instead justify them by their results. We turn to this in section 4, where we carefully assess the index and events against authoritative external sources for three major episodes of protest.

This methodology identifies some [679] monthly events, summarized in Figure 4, which shows the fraction of countries with social unrest events by region. One can clearly see several regional waves of protest. In the late 1980s, a number of events in South and Central America and the Caribbean drive the spike in events in the Western Hemisphere, including the democratic transition in 1986 in Guatemala, attacks on voters in the Haitian elections of 1987, the Chilean Plebiscite of 1988, the 1989 Paraguayan Coup d’état, and the 1990 Nicaraguan elections which ended the Sadanista regime. The fall of communism is reflected in an increase in the European series in the early 1990s due to events in Bulgaria, Czechoslovakia, Germany, Poland, Romania, and the Baltic countries.6 The largest recent wave of events is in the Arab Uprisings of 2011, which are clearly seen, driven by events across almost the whole Middle East (shown for six countries with event labels in Figure 3), as well as a simultaneous wave of anti-austerity protests in Europe. More recent protests in Bolivia, Chile, Colombia, Ecuador, and Peru contribute to a spike in unrest events in late 2019 and early 2020.

Figure 4:
Figure 4:

Fraction of countries with social unrest events, 12 month moving average

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

3.4 Event Screening

Other studies using measures of media attention often run extensive checks on the matching articles that they generate (see, for example, Baker et al. (2016)). This provides a valuable screen to assess the validity of the results. Instead of checking every month for every country by hand, we check only the events identified using the methodology of the preceding section. For each event, we check the matching articles in Factiva to 1) verify that events are not egregiously mis-identified, and 2) generate a label for each event. While the event dates are coded systematically, the labels are purely descriptive, aiming to provide a little narrative color to large movements in the RSUI. Note in particular that the label descriptions do not endorse a particular interpretation of historical events.

Mis-identified events do occur, for five main reasons. First, because events relating to a given country occur elsewhere. For instance, January 2019 protests in Greece over the name of North Macedonia are labeled as occurring in the latter. Second, because past social unrest events may be relevant context to a current event and therefore much-mentioned. For example, elections in Armenia in September 2018 followed mass protests in April. Accordingly, articles in September 2018 mention these protests even though they are not contemporaneous. Third, because the general search terms match certain country-specific features. For example, many Iranian institutions include the phrase “Islamic Revolution,” and so match the civil unrest search term “revolution.” Fourth, sometimes alternate usages of the search terms pollute the results – for example, elections in Denmark in 1987 for “Tax protest” parties. And fifth, occasional diplomatic events or disputes can result in events being misidentified. One such example was a diplomatic dispute in June 2014 between the Slovenian and Croatian governments over the possible entry of a Croatian patrol boat into Slovenian waters, where diplomatic interactions are described as “protest.”

We identify some [100] events where articles made no specific mention of contemporaneous civil unrest events (see Table 3) and follow a two-step strategy to address mis-identified events. We start by making the smallest country-specific changes we can to the search criteria to exclude articles related to reduce mis-identification. However, this is not always possible. For example, removing references to specific past unrest events will suppress the earlier event. So in some cases, we flag mis-labeled events by hand (less than one-third of the mis-identified events) and exclude them from the set of event individually (see Table 3).

Table 3:

Results of event scereening

article image

4 Assessing the Index

Many of the choices we make in defining the index and events in the preceding section are to a greater or lesser extent arbitrary. Justifying these choices is a key objective of this paper. This justification relies on two exercises.

First, we show that the resultant index and event listing agrees almost exactly with authoritative narrative sources for four episodes of major unrest in diverse locations and times: the Arab Uprisings of 2011; the sequence of protests, coups, and constitutional crises in Thailand during 2006–2014; protests in Venezuela in 2015–2019; and the color revolutions of the 2000s. This exercise is the acid test for our work and is something of an end-run around questions of bias or coverage or language or other details. It shows that despite the many critiques one might make of it, our approach seems to work; it provides timely and accurate identification of major events in diverse countries and times. In case this evidence is insufficient, we also provide several further narrative sources in Appendix B.7

Second, we show that the results are not sensitive to the particular choices we make. We vary language, sources, and search terms and –where applicable – get very similar results for a subset of Middle Eastern Countries since 2011. We interpret this as evidence that using media reports to identify social unrest is relatively forgiving; major events are sufficiently stark and obvious that they are revealed no matter the precise details of the methodology used to identify them.

One exercise that we do not put a lot of weight on is comparing to alternate measures, such as the Cross-National Time-Series Database (CNTSD, Banks and Wilson (2020)) or The Armed Conflict Location & Event Data Project (ACLED). This is because comparison to these alternatives is a relative test, and cannot distinguish between shortcomings of our measures and those of the alternatives themselves. For example, do differences with the CNTSD simply reflect the small number of sources used in that measure? And are differences with ACLED merely a function of that source’s emphasis on conflict? In contrast, the tests we value are absolute ones. Our measure matches authoritative, external reports of social unrest and it is not unduly sensitive to small changes.8

4.1 External Validity

We compare the index and our event coding to four major waves of protest about which we have alternative authoritative sources. In the main text we necessarily have to summarize these sources’ descriptions. To permit the skeptical reader to check our interpretation, Appendix E provides full verbatim descriptions of these events form the primary sources.

The four episodes discussed below represent a wide diversity of circumstance. They vary widely by location and time period; some are country-specific and others are region-wide; the types of government vary widely from vibrant democracies to more repressive regimes; and some result in major political changes and others do not. Yet across all of them, the index and the implied events match up very closely with alternate narrative sources.

4.1.1 The Arab Uprisings of 2011

For the Arab Uprisings of 2011, we take as our primary external source Worth (2016). This book-length account, written by a former New York Times journalist, is a comprehensive summary of the events and causes of the Arab Uprisings between June 2010 and December 2014, and is commonly cited as an overview reference on the subject. Most relevant for our purposes, this source includes an extended timeline with daily events in six countries: Bahrain, Egypt, Libya, Syria, Tunisia, and Yemen.

The downside of using Worth (2016) as our point of external comparison is that it is not peer-reviewed. This is not by design. Despite a wide political science literature on the Arab Spring, finding a single authoritative peer-reviewed summary timeline or event listing seems to be rather difficult; we could not find one. The reasons for this are beyond the scope of the current paper, but might be that academic study of the Arab Uprisings by political scientists focuses on individual episodes rather than a comprehensive history. In Appendix B.3, we also present the details of country timelines for a slightly different set of Middle Eastern countries during January 1999 to June 2019 assembled by the United States Institute of Peace and find broad agreement between their list of events and ours.

The events identified in Worth (2016) are shown as the vertical lines in Figure 5, along with summary descriptions.9 The period and countries is identical to those in Figure 3, which includes the RSUI event codings; the two can thus be directly compared.

Figure 5:
Figure 5:

Arab uprisings, June 2010 – December 2014: Countries with major events identified by Worth (2016)

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The most striking aspect of Figure 5 is that the cross-country timing of the initial round of protest almost exactly coincides with the initial spikes in the RSUI early 2011. The first outbreak of unrest during the Arab Uprisings were protests in Tunisia following the self-immolation on 17 December 2010 and subsequent death on 4 January 2011 of Mohamed Bouazizi. This is reflected in the large spike in the RSUI and corresponding RSUI-implied event in January 2011.10 Prior to this there are no RSUI-implied events in any of these six countries.

The first RSUI-implied event of the Arab Uprisings in Egypt is dated to February 2011. This again agrees almost identically with the external timeline, which dates the start of protests to the last week of January and extending through February 2011. Similarly, the RSUI dates the start of the Arab uprisings for Bahrain, Libya, and Yemen to February, consistent with Worth’s timing of February or very late January. And in Syria, the RSUI first picks up sharply in March, also consistent with the Worth timing. Overall, the index and event coding match the cross-country order of the start of the Arab Uprisings exactly, and the timing very closely.

Worth’s timeline following the initial outbreak of unrest also lines up very closely with the index and implied events. For instance, the major turning points of the Egyptian revolution all show as index spikes and satisfy event definitions, albeit sometimes with a one-month lag when an event occurs in the second half of a month. This includes running battles for control of Tahrir Square in November 2011, the election of Muhammed Morsi and subsequent unrest in June 2012, the collapse in November 2012 of the constituent assembly to draft a new constitution following President Morsi’s decree of personal immunity, and Morsi’s eventual overthrow in July 2013. In Libya, Tunisia, and Yemen the major post-revolutionary events are similarly identified by the RSUI.

There are only two major difference between the RSUI events and the Worth timeline, both of which are easily explained. The first is that the RSUI distinguishes separately the deaths of Tunisian politicians Chokri Belaid and Mohamed Brahmi, which Worth mentions together only following the death of the latter. The second major difference is for Bahrain in April 2012 when a bomb blast injured several policemen and the Bahrain Grand Prix was subject to domestic protest. These are arguably both major unrest events and so their inclusion does not seem unreasonable.

4.1.2 Thailand 2006–2014

The second narrative case study we look at is the series of related protests, coups, and constitutional crises in Thailand between 2006 and 2014. Our narrative sources are a series of annual synopses of Thai current events published in a peer-reviewed political science journal – Prasirtsuk (2009, 2010, 2015) and Dalpino (2011) – as well as two further papers from the Asian Studies literature, one published in a peer-reviewed journal – Baker (2016) – and one an institutional working paper, Keyes (2006). From these, we identify the timing of key unrest events. Verbatim extracts are reproduced in Appendix E.

Following a landslide re-election in 2005, anti-government protests erupted in January 2006 following revelations that the Prime Minister’s family had benefited from the tax-exempt sale of assets. In late February, snap elections were announced, although protests continued until the elections themselves in April. These events are reflected in a sharp spike in the RSUI in February and March 2006, corresponding to the peak of the protests. The results of the election were disputed, leading to a political vacuum which was resolved only by a military coup in September, reflected by a second spike in the RSUI in the same month (see Figure 6).

Figure 6:
Figure 6:

Thailand: Major events identified from external sources, January 2005-December 2014

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The events of 2006 set the stage for further unrest in 2008–2010. Military power ended and elections were held in late 2007, bringing allies of the now-exiled Prime Minister to power. His return in February prompted further political disputes, culminating in violent clashes in Bangkok between the “yellow” and “red” shirt factions (in September) and the takeover of two major airports by protestors (in November). These are reflected in peaks in the RSUI in the corresponding months.

Following a change in government in December 2008, protestors held major anti-government demonstrations in April 2009 in Bangkok. This coincided with a major regional summit (ASEAN) in nearby Pattaya and led to its cancellation and a further local peak spike in the RSUI.

After a Supreme Court ruling validating the seizure of the former Prime Minister’s assets, massive anti-government protests again broke out in March 2010. By April, the ongoing protests had turned violent, with the new Prime Minister declaring a state of emergency as anti-government protestors occupied parts of the city. In May, efforts by the army to re-establish control were accompanied by widespread disorder and several fires, including the burning of the Stock Exchange of Thailand and a major department store. Dozens died, including protestors, army officers, and an Italian journalist. The RSUI reflects the depth of the crisis at this point, with very high values in April and May.

Several years of relative calm followed and the transfer of power in 2011 to a government led by ex-Prime Minister’s sister prompted relatively little unrest, at least by the standards of preceding years. Yet the introduction of a bill in October 2013 providing amnesty for protestors and politicians upset the delicate balance of power and led to mass anti-government demonstrations in November. Unrest dissipated in December around public celebrations for the King’s birthday. But anti-government protests once again broke out in January, as opposition parties demanded reforms prior to the February elections. The elections went ahead as scheduled but the results were disputed and the returning Prime Minister ousted by the constitutional court, prompting a constitutional crisis. This was resolved in May when the military announced the imposition of martial law and, two days later, a coup ousted the civilian government. These events are closely matched by spikes in the RSUI in November 2013 (amnesty law protests), January 2014 (pre-election protests), and May 2014 (military coup).

Overall, the external narrative accounts of events in Thailand during this period line up very closely with the RSUI. The RSUI event-coding procedure identifies seven events in this period, corresponding closely to the narrative evidence discussed above. The events of 2006 do not register as RSUI-identified events, as later events received much more coverage, reflecting their increased severity and more frequent violence.

4.1.3 Venezuela 2014–2019

Following the presidential election of April 2013, Venezuela experienced several major waves of protests as opposition forces attempted to oust him from office. We use Gutiérrez (2017) and Briceño-Ruiz (2019) as alternate narrative sources for this period. Both sources are published in a peer-reviewed Latin American political science journal. Between them, these papers document the major events events during 2013–2019, several of which include to major social unrest events.

The first peak in the RSUI in this period corresponds to the first major wave of anti-government protests in February 2014. At the time, it was not obvious that the new President had the full support of a government dominated by supporters of the previous president, Hugo Chavez, and these protests – coded as an RSUI event – were a serious threat to the incumbent president. This is verified by the alternate narrative sources, which date major protests to February 2014 as well.

During 2015 and 2016, the RSUI remains at lower levels, with no implied events, and a small peak around the opposition’s parliamentary victories in late 2015. This also agrees with the narrative source. The events they record between mid-2014 and mid-2016 are all important political events, but no major protests or other civil disruptions. This changes in April 2017, when the RSUI picks up sharply and remains elevated until August. This agrees almost exactly with the description in Briceño-Ruiz (2019) who reports “a new wave of protests from April to August 2017”. That the RSUI can pick up an elevated period of protest such as this suggests that media “protest fatigue” is not a major source of bias in the index.

The final series of spikes for the RSUI in this period is in January 2019, when a constitutional crisis resulted in a power struggle and widespread protests, prompting a humanitarian crisis. These events are reflected both in the narrative descriptions and in RSUI-implied event for January 2019.

Figure 7:
Figure 7:

Thailand: Major events identified from RSUI event coding, January 2005-December 2014

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The narrative sources stop at this point. However, it is worth noting a further successful event identification of the RSUI in May of the same year, when a failed military uprising is reflected in another RSUI peak and an associated event.

Overall, the events in Venezuela line up very closely with those described in two peer-reviewed articles which include narrative descriptions of this period.

4.1.4 Color Revolutions

The color revolutions were a series of non-violent mass participation pro-democracy movements in former communist countries during the early 2000s resulting in the displacement of previous, often long-standing rulers. The term “color revolutions” refers to the fact that many protests were associated with particular colors, such as orange in Ukraine.11

Our external source for comparison is Tucker (2007), a peer-reviewed publication of the American Political Science Association, which focuses on the four most prominent color revolutions: Yugoslavia’s “Bulldozer revolution” in October 2000;12 Georgia’s 2003 “Rose revolution”; the 2004 “Orange revolution” in Ukraine; and the 2005 “Tulip revolution” in Kyrgyz Republic. In comparing these four episodes, Tucker provides a narrative description of each with key events listed by day.

Figures 10 and 11 present the RSUI and event coding for these countries during 2000–2005. In all four cases, the event coding identifies exactly the month of each of the four color revolutions. Secondary events following the main revolutionary episodes are also associated with spikes in the RSUI in Georgia and Kyrgyz Republic, in both cases delayed elections. That the event coding is somewhat conservative here is a manifestation of the trade-off between false positives and false negatives. Identifying more minor follow-up events would come at the cost of introducing erroneous events at other times in the sample.

Figure 8:
Figure 8:

Venezuela: Major events identified by Gutierrez (2017) and Briceño-Ruiz (2019), January 2013-June 2019

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 9:
Figure 9:

Venezuela: Major events identified from RSUI event coding, January 2013-June 2019

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 10:
Figure 10:

Color Revolutions: Major events identified by Tucker (2007), August 2000-December 2005

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 11:
Figure 11:

Color Revolutions: Major events identified from RSUI event coding, August 2000-December 2005

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

4.2 Internal Validity

In this section we report several results from several alternative measures of the RSUI, where we vary a number of parameters which define the RSUI, including language, sources, and search terms.

4.2.1 Sensitivity to language

To examine the hypothesis that regional and country-specific media might have different biases or could identify major social unrest events faster than western leading English-language media outlets, we recalculate the RSUI for a subset of Middle Eastern and North African countries using searches in French and Arabic. Arabic is, of course, the predominant language in the region and French – due to lengthy historical and cultural connections – is widely spoken in several countries. Naturally, this requires a different set of sources. For French, we used leading newspaper as well as newswire agencies in French: Le Figaro, Le Monde, Libération, Agence France Presse, Reuters, The Associated Press and The Canadian Press.

In Arabic, restricting our sample to only well-known international news sources did not provide wide enough coverage through Factiva to generate useful indices, with only a handful of articles available from major Arabic-language publications and a very low match rate to social unrest terms.13 We therefore cast a much wider net when defining our Arabic sources, allowing all search hits from all sources. This increases the number of articles available but the cost of a more varied set of sources, including news websites, blogs and less well-regarded outlets.

The French and Arabic search terms are reported in Table 9 in Appendix A.

Table 4:

Alternate search terms

article image
Note: Alternate searches for xit. All searches are derivatives of the headline search. Those labeled “Extension” include OR operators for the extra search terms, so return extra matches which may include terms listed. Those labeled “Restriction” append the listed terms using an AND operator, so return only those headline matches which also include the terms listed.
Table 5:

Within-country correlations with baseline, varying search terms

article image
Table 6:

Full-sample maximum-likelihood estimation of cross sectional density, N=6926, m¯=200. Standard errors in parentheses.

article image
Table 7:

Conditional probabilities of social unrest.

article image
Note: p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 Double-clustered standard errors shown in parenthesis.
Table 8:

Measures of predictability of index and events. Maximum within countries, then averaged across countries

article image
Table 9:

Alternate language search terms

article image

Figure 12 reports the coverage of three datasets by language as measured by the number of contemporary articles, zt, from which a key difficulty is immediately apparent: French language coverage is limited throughout and Arabic language content grows rapidly following 2011.

Figure 12:
Figure 12:

Number of contemporary articles per month, zt, by language

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 13 presents the A index for six Middle Eastern and North African countries where French is either widely spoken or with close historical or cultural ties to France. The major spikes are in exactly the same months for Algeria, Lebanon, Syria, and Tunisia. For Morocco there is some disagreement, but this is mostly during relatively quiet periods; major events such as the Arab Spring and the Hirak Rif protests of late 2016 and early 2017 show up clearly in both languages. Mauritania shows some differences but this is likely a function of the very small number of articles in French, with on average less than one article per month.

Figure 13:
Figure 13:

RSUI, June 2010 – present: Select Middle Eastern Countries with French widely spoken, by language

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 14 presents the A index for the six largest Arabic-speaking countries for both English-and Arabic-language sources. The Arabic-language sources do a poor job of distinguishing between countries. For example, Sudan and Saudi Arabia show large spikes in early 2011 despite very limited unrest at those times (and, in the case of Sudan, a much smaller response during the overthrow of the government in April 2019). Even within countries, the timing is inconsistent with the narrative evidence presented in the preceding section, dating the start of the uprising in Egypt to November 2010 – clearly too early to be plausible.

Figure 14:
Figure 14:

RSUI, June 2010 – present: Largest six Arabic-speaking countries, by language

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

There may be several reasons why the Arabic-language index struggles to distinguish country-specific unrest events. One factor could be the common language itself, with the resulting close cultural connections leading to an inherently more pan-Arabic outlook and thus reporting the Uprisings as a single common occurrence rather than as a series of country-specific events. The difficulty in matching social unrest articles also undoubtedly plays a role, with the peak in November 2010 a result of just 18 matching social unrest articles (versus a February 2011 English-language peak of some 1600 articles). This means that a few incorrect matches can have an outsize impact on the index.

Overall, the comparison across languages suggests either that our headline series agrees well (with French) or that there are more fundamental limits imposed by availability of sources (in Arabic).

4.2.2 Sensitivity to search terms

We also explore how our results vary with the choice of search term, for two reasons. First, we want to check that our headline search is not sensitive to small changes to the search terms. Second, we want to better understand what other issues might be associated with social unrest events.

To this end, we conduct four extra searches, all derivatives of the headline search for xit (see Table 4). The first two are extensions of the headline search, adding extra matches on related terms: one using a broader set of descriptions, including those related to potentially more violent outcomes (the “Expanded” search); and one which also admits a wider description of social tension rate than just unrest (the “Tension” search). The second set of extra searches are restrictions of the headline search, highlighting two of the key factors often cited as a driver of protest during the Arab Uprisings of 2011: economic underperformance and alienation of youth.14

We conduct these searches for a sub-sample of 19 Middle East and North African countries during the sample period. The results are presented in table 5, which shows the distribution of important correlation statistics of the different measures (and Figure 15 presents the alternate indices for the same countries as in Figure 14). Given our focus on major unrest events, our preferred measure is the overlap of the largest observations, which we limit to 2 percent, but we also present standard correlation coefficients. The overlap for the extension searches is very high, with three-quarters of countries having at least 85 percent of observations in common in the top 2 percent. This suggests that the headline index is not driven by specific or spurious matches related to the particular search terms we include but is instead a broad-based indicator of the underlying discussion.

Figure 15:
Figure 15:

RSUI, June 2010 – present: Largest six Arabic-speaking countries, by search term

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The searches restricted to include results driven by economic outcomes or youth activities also shine some light on the potential drivers of protest. In general, the largest observations of the RSUI tend to include just two key economic terms with notable frequency; in the median country, nearly four-fifths (79%) of the largest headline matches also included a economic term, and the agreement is over half (55%) in fully 90 percent of countries. While we cannot conclude that economic factors necessarily drove protest, this does mean that they were at least a common simultaneous topic of discussion. In contrast, terms related to young people come up much less frequently, with a common fraction of top matches in median country of only 45%.

5 Statistical Properties of the Index

In this section we discuss the statistical properties of the index. These are useful both for understanding how to interpret the index, as well as for understanding the temporal and spatial correlation of major unrest events and their predictability.

5.1 Cross-sectional distribution

We start by proposing a functional form for the cross-sectional distribution of the RSUI index. We assume that for values in excess of some cutoff m¯, the full sample RSUIitA can be described by:

(RSUIitA<M)=F(M)(forMm¯)

Where:

F(M)=1(a(Mm¯)+1)b

We think of m¯ as a filter for noise; below this threshold there is no meaningful information in the index. An appealing property of this distribution is that the fraction of observations above M is proportionate to Mm¯+1/a. That is:

log(1F(M))=blogablog(Mm¯+1/a)

In other words, if m¯+1/a is small, then a one percent increase in M above the threshold m¯ implies a b percent decrease in the number of observations at least this large. If this describes the data well then this provides a meaningful interpretation of changes in the A index.

We estimate this density by maximum likelihood, and report the point estimates and standard errors in Table 6 with m¯=200. Both parameters are sharply identified by the data, with small standard errors. And, as shown in Figure 16, the fitted and empirical distributions match very closely.

Figure 16:
Figure 16:

Cross-sectional RSUI density, empirical and estimated, m¯=200

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Appendix D provides further robustness checks, showing that 1) estimates for the parameters of the cross-sectional distribution are unaffected by the choice of the cutoff m¯ and that 2) within-country estimates of the parameters are not significantly different from the aggregate estimates.

These results give a useful interpretation of large observations of the headline index. Imagine one wishes to compare two observations in the upper tail of the distribution, say M1 and M2 with M1 < M2 and that both are large relative to m¯1/a130. Then we can compare them using a simple rule of thumb: that if M2 is x% larger than M1 then there usually will be approximately 2x% fewer observations larger than M2 than M1 (using here that b ⋍ 2). This rule of thumb holds both within individual countries, and across the sample as a whole.

5.2 Correlation of unrest events in time and space

At first glance, social unrest appears to be correlated across both space and time, with regional waves of unrest – such as the Arab Uprisings or recent protests in Latin America – appearing to be a relatively common phenomenon. Yet isolated unrest can and does occur. For instance, mass protests in Korea during 2016 leading to the impeachment of President Park Guen-hye and protests in France over public spending and pensions during late 2019 and early 2020 are not part of regional waves of unrest. So, to what extent is social unrest a correlated phenomenon? Our aim in this section is to examine this question, and document the conditional correlations of social unrest both in time in space. Note that our aim here is purely descriptive. Notions of causality -such as whether protests beget more protest or simply reflect persistent underlying drivers – are beyond the scope of this paper.

To analyze the dynamic and spatial correlations in social unrest we define:

  • Si,t = Months since last unrest event in country i

  • ui,t = Months since last unrest event in neighbors of country i

Neighboring countries are defined as those which share borders. Figure 17 shows the distribution of the times since social unrest events. As individual countries often have several neighbors, the time since neighboring country events is typically smaller, averaging 32 months versus 60 months for own-country events.

Figure 17:
Figure 17:

Distribution of time since last event

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

We can then create indicators for these times, dividing them into K distinct intervals (t0,t1], (t1,t2],..., (tK−1,tK]:

Si,tk=si,t(tk1,t1]Ui,tk=ui,t(tk1,t1]

Then we estimate the following panel regression:

eventi,t=αi+ηt+ΣkκβkdSi,tk+ΣkκβknUi,tk+eit(1)

where eventi,t is an indicator for an unrest event and αi and ηi are country and time fixed effects. Using this framework, we can interpret βd and βn as the incremental probability of an unrest event conditional on there being either a domestic or neighboring event during the last (tk-1,t1] months.

Table 7 presents the results of this exercise. The first column simply estimates the conditional average the unrest event indicator over the full sample, a little over one percent of months feature an unrest event. In column (2) we restrict estimation to the sub-sample where there has been at least one prior unrest event, resulting in an increase in the unconditional average event probability to nearly two percent. This specification is relevant null model for specifications (3) through (5) -note the same sample size as in (2) – as this specification can only be estimated on observations with at least one prior event.

Specifications (3) through (5) estimate the probability of a social unrest event within a country conditional on one occurring previously. As shown in specification (3), this probability rises sharply but briefly during the months following a social unrest event, rising by nearly 4 percent immediately following an unrest event.15 This is both statistically significant and quantitatively large, representing a quadrupling of the rate over that in times far from past unrest events.16 This impact decays slowly over the next nearly two years.

Of course, these effects may simply reflect selection, as countries with higher rates of average unrest will constitute a larger fraction of observations, particularly in the periods immediately after an event. We therefore correct for this in specification (4), which includes country fixed effects and, along with specification (8), represents the headline results in this section. These figures thus represent the marginal increase in the conditional probability of unrest over a county’s own average rate. Although the immediate increase falls to around 3.5 percent, this still represents a meaningful increase, with an approximate tripling over the sample average. However, accounting for average effects results in a much less persistent response, with no statistically discernible impact at horizons longer than six months. Specification (5) includes time fixed effects, which account for correlated average global variation in event probabilities, and which looks little different from specification (4).

Specification (6) performs a similar exercise but using time after neighboring-country events only. In this case, time fixed effects make little sense, as they will absorb any variation in neighboring countries which is correlated with global trends. The effect of neighboring-country events on the probability of social unrest is smaller, at a little over half the own-country effect, but still statistically significant. However, this effect seems to be slightly longer-lasting than the effect of purely domestic events. Specification (7) is identical to (6) except restricted to the sub-sample used for specification (8).

Along with specification (4), specification (8) is the other headline estimate. It reports estimates for the full specification in equation (1). This broadly confirms the findings of the separate regressions: within-country unrest is associated with an approximate 3 percentage point increase (an approximate quadrupling) in the rate of social unrest in the short term, but the effect is insignificant after six months; unrest events in neighboring countries are associated with roughly a doubling of the rate of unrest.

The fit of the regressions is in general fairly poor, with a naive R2 is no better than five percent.17 In causal settings this might be a sign of mis-specification and thus a concern. But in a purely descriptive context this is not relevant. Instead, it simply means that there is considerable variation in the realized outcomes beyond the averages conditional on time since last event.18 Moreover, the Akaike information criterion also suggests that the richer models are valid. For example, specifications (2) through (5) show a decline in the AIC and hence an improvement in the fit beyond that expected purely through the inclusion of extra parameters.19 Likewise, for specification (8) relative to (7).

Overall, the results of the panel regression exercise suggest that there is considerable spatial and temporal correlation of social unrest. Although social unrest is rare, occurring usually in around 1.3 percent of countries per month, this rate typically triples following an event in the same country, and doubles following one in a neighboring country. Nevertheless, these effects are short-lived, decaying to half a percent or less within 9 months.

5.3 Predictability

In the previous section we discussed whether social unrest was more likely after a social unrest event in the same or a neighboring country. Here, we turn to a related but complementary question: can we reliably predict outbreaks of social unrest before they occur? Of course, many factors may potentially predict social unrest including economic, financial, or other social variables. We leave investigation of these questions for other authors and instead ask a more limited question: can the RSUI itself predict social unrest?

Figure 18 illustrated the main point of this section and highlights its comparison with the previous one. The left-hand panel shows the conditional average of the RSUI index around an event. Unsurprisingly, the average RSUI spikes dramatically in the month of the event and declines slowly thereafter. This is essentially the same phenomenon described in the preceding section, just presented a little differently. Yet the RSUI displays a moderate increase (around 174 index points) in the period prior to the event. This raises a natural question: can such an increase be used to reliably predict social unrest?

Figure 18:
Figure 18:

Mean RSUI around events and large increases

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

The right-hand panel suggests not. It shows the mean RSUI conditioned observing increases of at least this magnitude (i.e. 174 index points). This has essentially no predictive power for social unrest events; the RSUI declines sharply back towards its long-run average in the subsequent months. In other words, while almost all social unrest events are preceded by moderate increases in the RSUI, moderate increases in the RSUI are usually followed by a swift regression to the mean.

We investigate this issue more formally by estimating lasso regressions for the index and events. A lasso (least absolute shrinkage and selection operator) regression is a simple machine learning algorithm which selects a model to minimize the out-of-sample forecast error. In this case, we use cross-validation for the out-of-sample exercise and estimate separately for each country:

υi,t+h=Σj=1Mβijυi,tj

Where M = 18, h = 1, ... , 24 and vit is either the RSUI or event indicator for country i in period t.

This approach therefore calculates the best country- and horizon-specific predictive model for the RSUI using up to a year-and-a-half of prior outcomes data. The purpose of using a cross-validated lasso is to try to estimate the most reliably informative predictive model.

For each country-horizon predictive model we compute three measures of predictability: the well-known Granger-Newbold measure, which is the fraction of variance of the outcome explained by the forecast, analogous to the R2; the Theil uncertainty statistic, which is less than one when the forecast is better than the long-run average; and the Diebold and Kilian (2001) statistic which measures predictability relative to the long-run error variance.

Table 8 compiles these measures by selecting the most predictive horizon for each country and then taking the cross-country average. Despite choosing the most successful predictions for each country, the predictability of each series is low. Indeed, the lasso regression typically selects at most only one lag of the RSUI when predicting future values of the index, and typically at very short lags. Likewise, the Granger-Newbold statistic shows that the best forecast explains barely three percent of the variation in outcomes.20 The Theil measure implies that the best forecast is little better than guessing the long-run average (which would produce a score of 1). Figure 19 presents the average and country-specific results by horizon for two measures. For RSUI events, predictability is even lower, with almost all country-horizon pairs selecting zero lags, i.e. the best predictive model in almost all cases actually is identically the country average.

Figure 19:
Figure 19:

Measures of predictability of the RSUI series

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

How should we interpret this lack of predictability, particularly in the light of the regression results of the previous section? To start with, note that these are not inconsistent. Just because two variables are correlated does not mean that either is a good predictor of the other. Social unrest indicates a higher probability of future unrest, but social unrest is sufficiently rare and driven by so many other factors that using past social unrest to predict future unrest results in such a small improvement in forecasting power as to be statistically useless. To make the point differently, the probability of winning a lottery rises conditional on buying a ticket. But if lottery wins are sufficiently rare, holding a lottery will not be a reliable predictor of winning.

A more constructive interpretation of these results is that social unrest likely happens for incredibly varied reasons, and can be triggered by seemingly minor events (e.g. increased subway fares). The rareness of events, combined with this apparent susceptibility to triggers, make individual social unrest events simply very hard to predict. As a result, average effects (however sharply identified) explain little of the variation in outcomes. Thus we can make statements both about average impacts (as in the preceding section) but still struggle to predict individual events.

6 Conclusions and possible applications

This paper does three things. First, it introduces a novel method for measuring and monitoring social unrest using article counts from international press, and presented a way of coding sharp spikes in the index into events. Second, we have shown that the events and the index correlate well with independent narrative courses during three diverse case studies. We also showed that the index is robust to choice of language, sources, and search terms. From this evidence, we conclude that the index and events likely measure realized social unrest rather than simply variation in media interest over time. Third, we provide some basic statistical analysis essential for interpreting the index, concluding that: the log of the index is proportional to its percentile rank in both within-and cross-country distributions; that social unrest is temporally and spatially correlated – social unrest raises the average probability of further unrest over the next six months by about three percentage points in the same country and about one percentage points in neighbors; but that that social unrest is not itself a reliable predictor of future social unrest.

These results have two main applications: providing reliable monitoring of social unrest; and in research on the causes and consequences of social unrest. To illustrate this first point, Figure 20 presents the latest RSUI data for select Middle Eastern and South American countries, indexed to 100 within the time periods presented. In both cases, the RSUI provides a timely, high-frequency data source for summarizing the timing and relative importance of episodes of social unrest. In particular, the timing and frequency of unrest across the two regions is very clearly differentiated. The timing of social unrest events is weakly correlated across the Middle Eastern countries, with sporadic outbursts throughout 2019, and some countries showing multiple unrest events (Algeria, Iran). This contrasts with recent protests in South America, which shows up as a single regional outburst from October 2019 onwards.

Figure 20:
Figure 20:

RSUI January 2018 – present

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

As for future research, the RSUI and the event listing – which are both freely available online – provide a rich battery of documented examples against which theories about the economic (and non-economic) causes and consequences of social unrest can be tested. The global coverage and monthly frequency of the RSUI data provide the granularity to investigate important issues such as the impact of unrest on financial markets and investment, the disruption to trade due to unrest, and the effect of corruption, governance, or poverty on unrest. A key challenge in all these areas is the issue of endogeneity: if social unrest and economic performance might each affect the other, or are jointly determined by a third factor. The analysis presented here suggests some ways of solving this problem. In particular, the results presented here on predictability and spillovers suggest that the exact timing of social unrest is a near-random event and that regional waves of unrest might be a useful instrument for domestic unrest.21 We leave these questions to future work.

References

  • Acemoglu, D., S. Naidu, P. Restrepo, and J. A. Robinson (2019). Democracy does cause growth. Journal of Political Economy 127 (1), 47100.

    • Search Google Scholar
    • Export Citation
  • Ahir, H., N. Bloom, and D. Furceri (2018). The world uncertainty index. Mimeo.

  • Ali, M., O.-H. Fjeldstad, B. Jiang, and A. B. Shifa (2019). Colonial legacy, state-building and the salience of ethnicity in sub-saharan africa. The Economic Journal 129 (619), 10481081.

    • Search Google Scholar
    • Export Citation
  • Baker, C. (2016). The 2014 thai coup and some roots of authoritarianism. Journal of Contemporary Asia 46 (3), 388404.

  • Baker, S. R., N. Bloom, and S. J. Davis (2016). Measuring economic policy uncertainty. The Quarterly Journal of Economics 131 (4), 15931636.

    • Search Google Scholar
    • Export Citation
  • Banks, A. and K. A. Wilson (2020). Cross-national time-series data archive.

  • Barro, R. J. (1991). Economic growth in a cross section of countries. The quarterly Journal of economics 106 (2), 407443.

  • Bodnaruk Jazayeri, K. (2016). Identity-based political inequality and protest: The dynamic re¬lationship between political power and protest in the middle east and north africa. Conflict Management and Peace Science 33 (4), 400422.

    • Search Google Scholar
    • Export Citation
  • Briceño-Ruiz, J. (2019). The crisis in Venezuela: A new chapter, or the final chapter? Latin American Policy 10 (1), 180189.

  • Caldara, D. and M. lacoviello (2018). Measuring geopolitical risk. FRB International Finance Discussion Paper (1222).

  • Choe, H. and J. Kim (2012). South korea’s democratization movements, 1980–1987: political structure, political opportunity, and framing. Inter-Asia Cultural Studies 13 (1), 5568.

    • Search Google Scholar
    • Export Citation
  • Ciocan, D. and B. Wüest (2016). Censorship of the arab spring in mena media.

  • Dalpino, C. (2011). Thailand in 2010: Rupture and attempts at reconciliation. Asian Survey 51 (1), 155162.

  • Diebold, F. X. and L. Kilian (2001). Measuring predictability: theory and macroeconomic applications. Journal of Applied Econometrics 16 (6), 657669.

    • Search Google Scholar
    • Export Citation
  • Fraiberger, S. P., D. Lee, D. Puy, and R. Rancier (2018). Media sentiment and international asset prices. The World Bank.

  • Gutiérrez, A. (2017). Venezuela’s economic and social development in the era of chavism. Latin American Policy 8 (2), 160188.

  • Harari, M. and E. L. Ferrara (2018). Conflict, climate, and cells: a disaggregated analysis. Review of Economics and Statistics 100 (4), 594608.

    • Search Google Scholar
    • Export Citation
  • Hellmeier, S., E. G. Rød, and N. B. Weidmann (2019). Coding instructions for the mass mobilization in autocracies database, version 2.0.

  • Hlatshwayo, S., A. Oeking, M. M. Ghazanchyan, D. Corvino, A. Shukla, and M. L. Y. Leigh (2018). The Measurement and Macro-Relevance of Corruption: A Big Data Approach. International Monetary Fund.

    • Search Google Scholar
    • Export Citation
  • Jirasavetakul, L.-B. F. and M. A. Spilimbergo (2018). Economic Policy Uncertainty in Turkey. International Monetary Fund.

  • Keyes, C. (2006). The destruction of a shrine to brahma in bangkok and the fall of thaksin shinawatra: the occult and the thai coup in thailand of september 2006.

    • Search Google Scholar
    • Export Citation
  • Kim, N. (2017). Candlelight and the yellow ribbon: Catalyzing re-democratization in south korea. Asia-Pacific Journal: Japan Focus 15, 117.

    • Search Google Scholar
    • Export Citation
  • Klopp, J. M. and E. Zuern (2007). The politics of violence in democratization: lessons from kenya and south africa. Comparative Politics, 127146.

    • Search Google Scholar
    • Export Citation
  • Manela, A. and A. Moreira (2017). News implied volatility and disaster concerns. Journal of Financial Economics 123 (1), 137162.

  • Ponticelli, J. and H.-J. Voth (2020). Austerity and anarchy: Budget cuts and social unrest in europe, 1919–2008. Journal of Comparative Economics 48 (1), 119.

    • Search Google Scholar
    • Export Citation
  • Prasirtsuk, K. (2009). Thailand in 2008: Crises continued. Asian Survey 49 (1), 174184.

  • Prasirtsuk, K. (2010). Thailand in 2009: Colored by turbulence. Asian Survey 50 (1), 203210.

  • Prasirtsuk, K. (2015). Thailand in 2014: another coup, a different coup? Asian Survey 55 (1), 200206.

  • Schock, K. (1999). People power and political opportunities: Social movement mobilization and outcomes in the philippines and burma. Social problems 46 (3), 355375.

    • Search Google Scholar
    • Export Citation
  • Tucker, J. A. (2007). Enough! electoral fraud, collective action problems, and post-communist colored revolutions. Perspectives on politics 5 (3), 535551.

    • Search Google Scholar
    • Export Citation
  • Waldmeir, P. (1997). Anatomy of a miracle: The end of apartheid and the birth of the new South Africa. WW Norton & Company.

  • Worth, R. F. (2016). A Rage for Order: The Middle East in Turmoil, from Tahrir Square to ISIS. Macmillan.

A Further details on construction

Here, we include further details on the construction of the alternative indices and events. Table 9 reports the search terms used for the alternate languages, and Table 10 the listing of modified searches and their justifications.

Table 10:

Search modifications and reasons

article image
article image
article image
article image
article image

B Other narrative approaches

In this appendix we compare the RSUI to narrative sources for five further times and places:the “people power” revolution in the Philippines in 1986; the pro-democracy Korean “June struggle” of 1987; protests surrounding the impeachment of President Park of Korea in 2016; and unrest in five Middle-Eastern countries between 1999 and 2019 (this overlaps with the analysis of the Arab Uprisings in the main text but uses an alternative narrative source).

B.1 Philippines, 1986

In 1986, mass protests and a contested election resulted in the resignation of longtime Filipino president Fernando Marcos, known as the “people power” or “EDSA” revolution. Our main narrative source for this is Schock (1999), a peer-reviewed article comparing social movements in the Philippines and Myanmar.

The events of this period occur within a relatively short period of time. Schock documents “rising protests” in 1985 (prior to the start of the RSUI) following the assassination of opposition leader Benigno Aquino. In November 1985, President Marcos announced snap elections for 7 February the following year. These were marred by allegations of violence and fraud, and a week later results were rejected by the new opposition leader, Corazon Aquino (Benigno’s wife). The following days saw an attempted coup, strong condemnation of the incumbent regime by several influential cardinals, and mass military defections to the opposition. By 25 February, Corazon Aquino had been sworn in and the Marcos family had departed into exile.

The rapid pace of events during this episode is reflected in singular peak in the appropriate month in the RSUI, February 1986, which s also identified as an RSUI event (Figures 21 and 22).

Figure 21:
Figure 21:

Philippines, January 1986 – January 1987: Major events identified by Schock (1999)

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 22:
Figure 22:

Philippines, January 1986 – January 1987: Major events identified from RSUI event coding

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 23:
Figure 23:

Korea, January 1986 – November 1987: Major events identified by Choe & Kim (2012)

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

B.2 Korea, 1986–1987

Our primary source for the events of the “June struggle” of 1987 in Korea is Choe and Kim (2012), a peer-reviewed article published in a cultural studies journal. The primary focus is on understanding the multiple democratization movements in Korea from 1980–1987 and so includes an extended history of the events leading up to the June struggle. These authors identify an intensifying circle of protests and repression starting with the student protests at Geonguk University in October 1986. The torture and death of a student demonstrator in January 1987 while in government custody was initially suppressed, but once information became public in late May, mass public protests started on 10 June and grew with each passing week, peaking with the largest demonstrations on 26 June. Three days later, President Chun agreed to free elections and his nominated successor General Roh acceded to the protestors’ major demand. Protests subsequently subsided and largely peaceful elections took place in December 1987.

The RSUI reflects these events, showing a very pronounced peak in exactly June 1987, and a corresponding RSUI-implied event. Combined with the evidence presented in Section B.1, this illustrates how the index can identify precisely to the month very fast-moving protests events, event very early in the sample.

Figure 24:
Figure 24:

Korea, January 1986 – November 1987: Major events identified from RSUI event coding

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

B.2.1 The End of Apartheid

The final test case for the RSUI is the end of the Apartheid regime in South Africa. For this episode we compare our approach to two in-depth analyses. One is a peer-reviewed article from a political science journal, Klopp and Zuern (2007), who provide a narrative account of the democratization of South Africa with a daily record of major events. The other, Waldmeir (1997), is a book-length journalistic account of the transition away from Apartheid based on interviews with major figures involved and which provides a monthly timeline of major events.

The end of Apartheid was a unpredictable and fast-moving event, and so the two external sources identify a large number of major events in this period. For the sake of transparency Figure 25 includes all those mentioned by the two external sources. Many events are important, but principally political rather than describing actual unrest. We therefore highlight with asterisks descriptions which are more plausibly described as social unrest. Nevertheless, by including all the events in our sources we allow the reader to assess for themselves whether this distinction is reasonable, and to form their own opinion on whether non-unrest events are being (wrongly) reflected in the index.

Figure 25:
Figure 25:

South Africa: Major events identified by Waldmeir (1997) and Klopp & Zuern (2007), October 1989-December 1994. Asterisks denote likely unrest events.

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

We interpret Figure 25 as convincing evidence that the RSUI reflects well the narrative histories of Waldmeir (1997) and Klopp and Zuern (2007). The highest two peaks of the index match the two events which bookend the period, both of which were accompanied by massive public demonstrations: the release of Nelson Mandela in February 1990 and the first non-racial elections in late March 1994 (again, triggering a peak in April due to the monthly frequency). Likewise, key intermediate events which prompted or were due to unrest are also reflected by peaks in the RSUI, including: the deaths of 17 protesters at Sebokeng in March 1990; the breakdown of multi-party talks (CODESA) and mass ANC demonstrations starting in June 1992; the Bisho Massacre in September 1992; and the assassination of Chris Hani (the head of the ANC’s armed wing) in April 1993.

The large number of events in these sources also provides an opportunity for several falsifications checks. A number of undoubtedly important but entirely political (i.e. not social unrest) events barely register on the RSUI. These include: the start of multi-party talks in December 1990, the Record of Understanding which restarted negotiations following the failure of CODESA; and agreements on the constitution and governance prior to the 1994 election.

Figure 26 presents the same period with RSUI-implied events overlaid. This again highlights that the event coding is accurate but a little conservative – identifying arguably the most important events of the period to the correct month, but only identifying three events.

Figure 26:
Figure 26:

South Africa: Major events identified from RSUI event coding, October 1989-December 1994

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

B.2.2 Korea 2016–2017

The protests surrounding the impeachment of President Park are often known as the “Candlelight revolution”. In late October 2016, reports emerged of abuse of power by one of the President’s aides. This, and subsequent revelations, led to mass protests during November 2016. These events are documented in Kim (2017), a peer-reviewed article in a political science journal. This is reflects in the RSUI, which picks up sharply in November, a month coded as an RSUI event (see Figure 28).

Figure 27:
Figure 27:

Korea: Major events identified by Kim (2017), June 2016-December 2017

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 28:
Figure 28:

Korea: Major events identified from RSUI event coding, June 2016-December 2017

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Protests continued through late 2016 (identified by Kim (2017) as reaching their “heights” in December”) before easing when the President was suspended by Parliament for 180 days. This suspension corresponded with an easing in tensions while the Korean Constitutional Court heard argument and evidence. However, the verdict in March 2017 corresponded with further large protests, resulting in three deaths. This event is also coded as an RUI event.

The subsequent elections in May 2017 are also reflected in an increase in the RSUI. However, this is not coded as an RSUI event, as it fails our event screening criteria. There are no contemporaneous reports of protest or other forms of social unrest in this month. Instead, matching articles include mention of prior protests as context for the current elections. Thus, there is no RUSI event recorded despite the local spike.

B.3 Middle East, 1999–2019

In addition to our analysis of the Arab Uprisings of 2011, we also compare our results to a broader analysis of unrest in the Middle East between 1999 and 2019. For this, we use timelines assembled by the United States Institute of Peace (“USIP”) of five Middle Eastern countries (Egypt, Iran, Iraq, Libya, and Tunisia) between January 1999 and June 2019. The events identified in the USIP timelines are shown as vertical lines in Figure 29. The period and countries are identical to those in Figure 30, which includes the RSUI event codings.

Figure 29:
Figure 29:

Middle East, January 1999 – June 2019: Major events identified by USIP

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 30:
Figure 30:

Middle East, January 1999 – June 2019: Major events identified from RSUI event coding

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

For the four Arabic-speaking countries, the time series is dominated by the Arab uprisings of 2011. Accordingly, the initial spikes in the RSUI in early 2011 in each of these five countries coincides with the initial round of protests as identified by USIP. In Tunisia where the Arab Spring started, protests began spreading after the self-immolation on 17 December 2010 and subsequent death on 4 January 2011 of Mohamed Bouazizi. This is reflected in the RSUI-implied event in January 2011. In Egypt where the first wave of unrest started a few weeks later on January 25th, the first RSUI-implied event of the Arab Spring in Egypt is dated to February 2011. Similarly, the RSUI dates the start of the Arab Spring in Iraq and Libya to be February 2011 and Syria to be April 2011, which coincides with the external timeline.

The RSUI also captures civil unrest documented by the USIP timelines in the period following the initial outbreak of the Arab Uprisings, with secondary protests coinciding with spikes in the Index, and major turning points of the revolution matching the implied events. For instance, key protest episodes in the USIP timeline for Egypt, such as post-election unrest in June 2012, protests following President Morsi’s decree of personal immunity, and his eventual overthrow in July 2013 all agree with RSUI-implied events.

In addition to the four narratives Arabic-speaking countries, the USIP also provides a timeline of recent events for Iran. Similarly, the RSUI does a good job of capturing key unrest episodes in that country, including the widespread unrest in Janurary 2018 and the June 2009 Green Movement protest, described by USIP as “ne of the most serious challenge to the theocracy of the revolution”.

There are only three major differences between the RSUI events and the USIP timelines. First, the RSUI-implied events include the parliamentary and presidential election in Tunisia in 2014 and presidential elections in June 2013 in Iran. It is not clear why these are omitted from the USIP timelines. Second, the USIP captures more protest events than our event coding, simply because it identifies separate within-month events, such as demonstrations between April 1 and April 12 demanding Mubarak’s prosecution and faster reform in Egypt. As discussed in the main text, however, these relatively smaller unrest episodes are reflected in elevated levels of the index, while key episodes are identified as events. The event coding is somewhat conservative as we face a tradeoff between false positives and false negatives. In this study, we opt to minimize the former. Finally, there are three major RSUI-implied events that are not picked up by USIP: August 2011 and November 2011 in Libya , and July 2018 in Iraq. The events in Libya reflect the onset of the civil war there. In Iraq, the USIP dates the Basra protests “over unemployment, shortages of clean water and electricity, and widespread corruption” to September 2018. However, those protests appear to be a continuation of those a few months earlier, triggered by electricity cuts on July 8th, in line with our event dating.

C Comparison to other measures

Here we compare the RSUI to other commonly used measures of social unrest. The first is the Armed Conflict Location and Event Database (ACLED) which includes counts of individual demonstrations and riots at the country level. While coverage has expanded dramatically since 2015, the primary focus of this data source is on African countries, of which a sample of 27 have data starting in 1998. We therefore create an index for each of these countries analogous to the RSUI, summing the number of events in each month and rescaling so that the index averages to 100 since 2000, which we combine into a single summary measure by taking a simple average across countries in each month. This is presented in Figure 31, along with the average RSUI for the same 27 countries (also rescaled to average 100 over the same time period).

Despite measuring different phenomena using different methodologies, there is broad agreement between the two measures. Both are relatively low in the first decade of the 2000s and then exhibit a substantial and persistent increase starting in 2011. Within this, both series also show broad qualitative agreement in the last few years, with a rise in unrest around 2015, a subsequent decline during 2016–2017 and a pick-up since. At an annual frequency, the correlation is 0.7. At a monthly frequency, the two indices unsurprisingly diverge a little more, with a correlation nearer to one half.

The second alternate data source is the Cross-National Time-Series Data (CNTSD) database by Banks and Wilson (2020), used in other studies of social unrest and economics, such as [Fuceri paper]. This has very broad coverage, including 117 countries starting in 1995, but at annual frequency. This index reports the number of riots and anti-government demonstrations, which we sum and compute country indices. From these, we compute a simple average to give an aggregate annual index.

This index is reported in Figure 32. The CNTS seems to show an almost monotonic increase in unrest events from around 2009 onward with very little variation prior. This is somewhat hard to reconcile with the evidence from the RSUI and the ACLED. Unfortunately, there is no way to further interrogate the data to uncover which particular events are driving this. Moreover, given the annual frequency of the data one cannot easily make an educated guess as to the source of this increase. One possible explanation is that the CNTS uses a raw count of reports of unrest. If media coverage overall has increased during this time (perhaps as a result of reduced cost of printing or increased interest in foreign news) then such a pattern might emerge. As an extra cross-check, we also compare to the average rate of RSUI-implied events. This does not seem to explain the behavior of the CNTS index any better, with increases in the early 2000s and 2010s apparently uncorrelated with the CNTS index.

As fluctuations in long-term multi-country averages may reflect changes in composition and measurement over long periods of time, we also compare the RSUI to the ACLED and CNTS data during specific events for which we have already validated the RSUI using narrative data. To this end, Figures 3336 compare the CNTS and (where available) ACLED data to the RSUI. Here, the agreement between the various measures is very good. Note in particular the close alignment between the timing of spikes in the RSUI and the ACLED index during times of extreme stress.

Figure 31:
Figure 31:

RSUI vs. ACLED, 27 countries, January 2000 – March 2020, simple average

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 32:
Figure 32:

RSUI vs. CNTS, 117 countries, January 1995 – March 2020, simple average

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 33:
Figure 33:

RSUI vs. ACLED & CNTS, Tunisia, 2000 – 2014

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 34:
Figure 34:

RSUI vs. ACLED & CNTS, Egypt, 2000 – 2014

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 35:
Figure 35:

RSUI vs. CNTS, Venezuela, 2013 – 2019

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 36:
Figure 36:

RSUI vs. ACLED & CNTS, Thailand, 2005 – 2015

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

D Robustness of Statistical Properties

Figures 37 and 38 show two further robustness checks for the estimates of the cross-sectional density of the RSUI. The first presents country-specific estimates of the two parameters, compared to their full-sample estimates. Here, the same estimation exercise is conducted using only the RSUI index for a given country. Error bars show two-standard-error ranges percent confidence intervals. These show that in most cases we cannot reject the hypothesis that the country-specific estimates equal the full-sample ones. The full-sample estimates fall within the country-specific in over 80 and 90 percent of countries for the estimates of a and b respectively. This is somewhat surprising, particularly given the scope for cross-country variation in media coverage and patterns of unrest. This further justifies the interpretations of the units of the RSUI given in Section 5.1.

Figure 37:
Figure 37:

Country-specific parameter estimates. Some countries omitted due to MLE convergence failures. Full-sample estimates shown in blue.

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 38:
Figure 38:

Parameter estimates with changing threshold cutoffs

Citation: IMF Working Papers 2020, 129; 10.5089/9781513550275.001.A001

Figure 38 reports the same estimates as the value of the arbitrary cutoff m¯ changes. Changing m¯ has a mechanical impact on a, proportionate to 1/a, and so panel (a) shows estimates transformed into this space (with standard errors computed via the delta method). The parameter a is reasonably stable, tracking the full-sample-implied value for 1/a reasonably closely. However, the estimate for b – arguably more important for the interpretation of the RSUI units – is very stable. In fact, the standard error for b actually decreases despite the reduction in sample size as m¯ rises. We interpret this as evidence that the RSUI is particularly informative around major events. There may be considerable noise when there is little activity in a given country. But when a large event of unrest occurs, the signal becomes increasingly precise. As a result, the upper tail of the distribution of RSUI values is particularly easy to interpret.

E External event descriptions

E.1 Arab Uprisings

Table 11:

Verbatim descriptions of Arab Uprising events from Worth (2016)

article image
article image
article image