Liberation Technology: Mobile Phones and Political Mobilization in Africa

Liberation Technology: Mobile Phones and Political Mobilization in Africa∗ Marco Manacorda† Andrea Tesei‡ This version: April 2015 Preliminary. Ple...
2 downloads 0 Views 5MB Size
Liberation Technology: Mobile Phones and Political Mobilization in Africa∗ Marco Manacorda†

Andrea Tesei‡

This version: April 2015

Preliminary. Please do not circulate further

Abstract: Can digital information and communication technology foster mass political mobilization? We use a novel geo-referenced dataset for the entire African continent between 1998 and 2012 on the coverage of mobile phone technology - together with geo-referenced data from multiple sources on the onset of protests - to bring this argument to empirical scrutiny. We show that protests are anti-cyclical and that the availability of mobile phones in a local area amplifies the effect of economic downturns on the incidence of mass political mobilization: our most conservative estimates imply that a 5 p.p. fall in national GDP growth leads to an estimated increase in the differential rate of protests per capita in covered versus non-covered areas of between 2 to 10 percent. Our results lend support to the liberation technology argument that digital ICT fosters political mobilization.

Keywords: mobile phones, collective action, Africa, geo-referenced data JEL codes: D70, O55, L96



We thank XX and seminar participants at DFID, IADB, IMT-Lucca, Namur, Oxford, Oslo and UPF for many helpful comments. † Queen Mary University of London, CEP (LSE), IZA & CEPR; contact: [email protected]. ‡ Queen Mary University of London & CEP (LSE); contact: [email protected].

1

1. Introduction The recent spread of digital information and communication technology has fed a wave of optimism and a large amount of rhetoric about its use as a “liberation technology” capable of helping the oppressed and disenfranchised worldwide. According to this argument, famously articulated in ?, thanks to opportunity that these offer for two-way, multi-way and mass communication, their low cost and decentralized, open-access nature, mobile phones and internet, together with their innumerable applications (i.e. mass SMS, such as Frontline SMS, monitoring and information storage systems, blogs, on-line chat rooms, social media such as Facebook and Twitter, etc.) - offer the opportunity for citizens and grass-root organizations to more easily access and spread information, including to the outside world, and more easily communicate and organize. In turn, this has the potential to foster citizens’ political activism and even mass political mobilization when reasons for grievance arise and especially traditional, civic forms of political participation (e.g. voting) are de facto or lawfully prevented.1 Although collective, sometimes violent, action might not necessarily be desirable, as this is likely to impose some costs on other individuals and on society as whole, this can help advance the cause of excluded groups, leading to redistribution and possibly improved social welfare, especially in countries where the majority of citizens live in poverty and are deprived of political, economic and social freedom. This argument appears particularly appealing for Africa, the continent that has experienced the fastest rise in the spread of mobile phone technology worldwide: while in 1999 an estimated 80 million African citizens had access to a mobile phone, in 2008 this number was estimated on the order of 477 million, around 60 percent of the entire continent’s population (?). The spread of mobile technology across the continent has taken place against the backdrop of a practically non-existent fixed telephone line infrastructure, and because of this, it is claimed to have had unprecedented economic and social effects on the lives of its citizens, in particular the poor and very poor (?).2 Due to the lack of a fixed phone line and high-speed internet cabling, mobile phones are also the most used way to access the Internet and social media in the continent (?), greatly enhancing their information and communication potential. Although a new wave of optimism surrounds Africa’s recent development path, reasons for grievance have traditionally abounded in the continent, with Africa, and in particular SubSaharan Africa, ranking at the bottom of the world ranking along most indicators of economic, social and democratic development (?). Consistent with the liberation technology hypothesis, Africa has witnessed over the last decade 1

Already in 2007 The Economist highlighted the role of mobile phone technology in fostering political activism worldwide, launching the term of “mobile activism” (The Economist, 2007). Digital ICT and new media, including blogging and twitter, are also claimed to have been instrumental in what appears to be a recent surge of protests worldwide (?), from the Occupy Wall Street movement in the USA and the indignados in Spain to the “Arab Spring” in North Africa and the Middle-east (?). 2 The ubiquitous use of mobile phones in the continent has also led to the emergence of a number of creative applications and technological developments, such as SMS-based election monitoring, health and disaster prevention SMS-based information campaigns, disaster relief campaigns and mobile banking (?).

2

some of the most spectacular episodes of mass mobilization. Food riots swept the continent between 2007 and 2008 affecting 14 countries (?), while mass civil unrest (the “Arab Spring”) exploded in the northern countries (Morocco, Libya, Egypt, Tunisia, Algeria, and to a minor extent in Sudan, Mauritania and Djibouti) between 2010 and 2012, (?)). Despite a great deal of enthusiasm and plenty of anecdotal evidence on the role played by digital ICT - and in particular mobile phones - in fostering mass political mobilization in Africa, there are good reasons, discussed below, to be skeptical about the role effectively played by this technology, and the evidence remains admittedly scant. As far as we are aware there is no systematic study that establishes a convincing relationship between digital ICT and political mobilization in Africa. In this paper we bring the liberation technology argument to empirical scrutiny using several novel and by and large unexploited datasets for the whole of Africa, respectively on the spread of mobile phone technology and on the incidence of protests. What makes these different datasets particularly appealing is their level of geographical detail, which allows us to examine the spread of protests and mobile phone technology over time across small areas within countries. Data on local mobile phone coverage come from the Global System for Mobile Communications Association (GMSA) which collects this information for the purpose of creating roaming maps for use by customers and providers worldwide. These data provide information on the availability of signal, separately for 2G, 3G and 4G technologies, for the whole of Africa (with the only exception of Somalia) between 1998 and 2012 at a level of geographical precision between 1 to approximately 20 km2 on the ground, depending on the country. GSM technology accounts for around 80 percent of mobile technology worldwide and almost 100 percent in Africa. In order to measure protests, we use several datasets, all largely coming from compilation of newswires. First, we use micro data from a very large, open-source dataset, which relies on automated textual analysis of news sources, the Global Database on Event, Location and Tone (GDELT, (?)). As this is a largely yet unutilized dataset and since we have no control on the algorithm used to collect the data, we complement this information with data from two widely utilized, but much smaller, manually compiled datasets on unrest in Africa, the Armed Conflict Location Events Database (ACLED) and the Social Conflict in Africa Database (SCAD) (see for example ? and ?). We finally combine these data with data from a variety of sources about, among others, population, nature and use of land, infant mortality, natural resources, distance to cities, to the border and to the coast, kilometers of road etc. for approximately 10,000 (55 X 55 km) cells which compose the continent, which is ultimately the unit of observation in the analysis. This very detailed level of geographical disaggregation allows us to compare changes in the incidence of protests in areas within the same country that experienced differential changes in the coverage of mobile technology. By focusing on within rather than between countries variation in the incidence of protests and the spread of ICT, we hope to alleviate the obvious concern - and the ensuing bias that results in estimates of impact - that ICT adoption and the

3

incidence of protests are correlated due to country specific trends or shocks in unobservable variables, such as the state of the economic cycle or levels of economic growth. In the empirical analysis we start by showing that the three dataset on protests convey similar information offering some reassurance on their information content. We also show that protests are strongly anti-cyclical, i.e. more likely to occur during recessions. When we turn to the within-country analysis, we then show that the effect of recessions on protests gets magnified in areas with a greater spread of mobile technology: our estimates suggest that a fall in the rate of GDP growth of 5 percentage points (approximately one standard deviation) leads to a differential increase in the protests per capita between an area with full mobile phone coverage relative to an area with no coverage of around 10 percent (compared to an average of 1.24 protests per hundred thousand population per year), with an effect of 3G technology - which allows for data downloading and internet browsing - of more than five times as much. In order to control for the possibility that local economic shocks or other determinants of ICT adoption and protests might drive our results, we show that our estimates are robust to flexible specifications that condition for differential linear time trends across areas within the same country with the same large array of baseline characteristics. Although, by including interactions of a large array of cross sectional cell characteristics with country-specific time trends in the regressions, we attempt to control at best for the joint determinants of protests and mobile phone technology across areas, a concern remains that even conditional on these variables, mobile technology adoption remains correlated with unobserved trends in protests. To address this concern, we use an instrumental variable strategy that exploits the slower adoption of mobile technology in areas subject to high incidence of lightning. Lightning damages mobile phone infrastructures and in particular antennas on the ground that transmit the signal in their vicinity and negatively affects connectivity, acting on both supply (as power surge protection is costly) and demand (as poor connectivity makes the investment in technology less profitable and the risk of intermittent communications discourages adoption). Based on NASA satellite-generated data on the incidence of lightning for the entire Africa, we show that areas with higher average incidence of lightning display slower adoption of mobile phone technology. A one standard deviation increase in flash rates leads to a lower penetration rate of mobile phone technology of approximately 0.1 percentage points per year. The IV estimates - although typically less significant than the OLS - show even larger effects of economic downturns on the incidence of protests in high coverage relative to low coverage areas, with effects as large as 5 times the ones found based on the OLS, pointing into the direction of mobile phone technology adoption being negatively correlated with local trends in the incidence of protests due to unobserved factors or even reverse causality. Although this is the first paper that we are aware of that investigates empirically how digital ICT affects mass citizens’ political mobilization, there are different streams of literature in economics that our paper relates to and that can borrow insights from. In principle, it seems reasonable to speculate that mobile phone technology has the potential

4

to trigger collective action during bad times by making citizens more informed on issues and reasons for grievance. ICT can help individuals communicate and spread information via twoway, multi-way and mass communication.3 Related to this argument, a number of studies for the USA show that media and newspapers foster political participation (??), although these studies focus on traditional media and on civic forms of participation in advanced democracies and it is unclear whether these findings extend to ICT in low income countries and in less mature democracies or in autocracies. This argument focuses on the role of information provision on citizens’ private incentives to participate in the political process, via its effect on the perceived individuals costs and returns. However, collective action has a public good nature: as many typically share the returns from mass mobilization, while costs are privately borne, free riding is likely to emerge, leading possibly to an under-provision of protests (?). In this respect, mobile phone technology, with its associated reduction in the costs of acquiring and transmitting information, is likely to affect mass mobilization through its ability to foster coordination among citizens when reasons or grievance emerge. This seems particularly relevant when strategic complementarities exist, i.e. when the returns to political activism increase - or the cost of participation decrease - the larger the number of others participating. Digital ICT has the potential to trigger collective action not only because it makes individuals privately more informed about reasons for grievance or about the private costs of participating, but also through the common nature of the signal received. A similar argument is advanced in ? who studies the role of government propaganda during the Rwandan genocide. The paper argues that greater local radio coverage fostered participation in mass killing not only through a direct information channel, that the government was unwilling to punish perpetrators, but also through the spread of common knowledge, i.e. the knowledge that others also knew (and knew that others knew), in turn solving the coordination problem that plagues collective action. Somewhat in a similar vein, recent work by ? shows that participation in rallies during the Tax day in the USA was lower in rainy locations. This is probably not only because individuals had greater private costs of participating but also because rain knowingly made others less likely to participate, hence reducing each individual’s private invective to participate. Consistent with this, ? also show that riots in the USA are more likely to occur in cities than in rural areas, an effect that they ascribe precisely to improved coordination when communication costs are lower. By granting access to unadulterated information digital ICT has also the potential to offset government propaganda, which - typically through traditional media - curbs discontent via misinformation and persuasion ?. This seems particularly relevant when the media are under the control of the government or in the hand of powerful interests groups. Indeed, a small but growing body of literature emphasizes the role of traditional media on voters’ political alignment (see ? seminal study of FOX news in the USA, ? study of the effect of government-sponsored 3

Mass SMS, political and information campaigns are indeed increasingly popular in Africa, see Aker et al. (2011).

5

radio propaganda on mass killing in Rwanda, ? study of Berlusconi’s TVs on voters’ outcomes in Italy), while other papers show that independent media can counteract these effects (see ? on the effect of independent TV on voting outcomes in Russia). As already mentioned, despite the popularity of the “liberation technology” argument, there is no lack of reasons for skepticism. First, governments can use this technology as a control, surveillance or propaganda tool, hence making protests less rather than more likely (?). This effect is enhanced by the nature of the technology, that makes centralized control likely or at least possible, an effect that is magnified by the circumstance that physical infrastructures as well as market regulation of ICT is, for obvious reasons, often directly in the hands of governments. A second often heard counter-argument against the liberation technology hypothesis is that digital ICT can discourage social capital accumulation and the establishment or “strong ties” (in favor of “weak ties”) which are thought to be instrumental to mass mobilization (?), ultimately leading to political apathy rather than mobilization. Indeed, there is some evidence that internet and new media can lead to greater political disaffection (?). These results echo findings for the USA on the negative effect of television on voters’ turnout (?) and results for Indonesia by ? showing that radio leads to reduced levels of trust and social capita.4 Although this literature suggests that digital ICT has the potential to deter political participation, it focuses on voting and civic forms of political participation in richer countries and mature democracies and it remains an empirical question of what effect this has on spontaneous, less codified and perhaps less civic forms of political participation in poorer countries and under authoritarian regimes or in more fragile democracies. Perhaps a more subtle argument of why digital ICT might not ultimately lead to the emergence of mass mobilization is that this technology has the potential to increase government accountability via information spread and greater transparency, in turn detracting from the rationale for mass political mobilization, which is widespread discontent with the perceived state of the economy and politics. An established body of literature shows that traditional media affect government through increased accountability (see ?, ?, ?, respectively for India, Uganda and the USA) and recent evidence for Russia shows that blog posts have the potential to reduce corruption in state-controlled companies (?). Our paper is also related to the literature on the relationship between mass mobilization and economic circumstances. A small but established body of evidence shows that individual participation in mass political movements is negatively correlated with economic conditions, as worse economic conditions are associated with lower private opportunity costs of participation and provide a rationale for widespread grievance (???).5 4

This appears in contrast to what found for radio in the USA (??). There is also an argument that in the long run new media can become platforms for political change (?). 5 This parallels findings that worse economic conditions are typically associated with greater incidence or risk of conflict and insurgency, see (?????)), although there is also an argument that economic growth can foster rather than discourage unrest through a rapacity effect (?). There is evidence that civil unrests is associated to variability in food prices but only in low-income countries (?). Focusing specifically on protests in Africa, ?

6

A small recent literature in political science investigates the role of mobile phone technology in affecting violent conflict and insurgency using data from Africa and Afghanistan (see respectively ? and ?), with mixed findings. Differently from our study, though, not only do these studies focus on violent and rare phenomena but they also largely exploit cross-sectional variation across areas, or differential time variation across areas but only over a short time span. In contrast, our data allow us to examine variation over time within countries during a long time period over which mobile phone technology adoption grew rapidly. In sum, there are a number of theoretical arguments, and some indirect evidence, suggesting that mobile phones and digital ICT in general can foster political mobilization. However, there are also good reasons to believe that the reverse is true. Ultimately, it remains an empirical matter, which - if any - of these forces in practice prevails and this paper precisely turns to investigating empirically this question. The rest of the paper is organized as follows. Section 2 presents the data. Section 3 lays out our empirical strategy and section 4 presents the empirical results. Section 5 concludes.

2. Data In this section we present the main sources of data used in the rest of the analysis. We start by focusing on geo-referenced data on mobile phone coverage and we then document the available geo-referenced micro data on protests. Our data refer to 48 African countries, all those on the mainland (with the exception of Somalia), plus the islands of Madagascar and Cape Verde, and extend over 15 years, from 1998 to 2012.6

2.1. PRIO grid cells In the analysis we match these different datasets at the level of cells of 0.5◦ x 0.5◦ degree resolution, approximately corresponding to 55 x 55 km at the equator (3,025 km2 ). The lattice of these cells is constructed and made available by the Peace Research Institute Oslo (PRIO) (Tollefsen et al., 2012). The advantage of focusing on grid cells (rather than, say, on administrative partitions within countries) is that for these cells we have data on a large array of socio-economic and other characteristics, including population. This allows us to examine the relationship between ICT adoption and the spread of protests across relatively fine geographical areas, while controlling for a large array of local characteristics.

show that protests are directly affected by extreme weather events and ? find a positive correlation between protests and droughts, while ? claim that droughts make the threat of protests more likely and, via this, they speed the transition to democracy. ? and ? also claim that the responsiveness of political participation to individuals” education is lower when the returns to education are higher and when the economy performs better. 6 In order to keep the dataset balanced we do not account for the creation of South Sudan in 2011, treating Sudan as a single country throughout the entire sample period.

7

In total, Africa is divided into 10,650 cells. If one excludes Somalia, for which we have no data on mobile phone coverage, this leaves us with a total of 10,419 cells - an average of 211 cells per country - that we observe during 15 years. At total continent population (excluding Somalia) of around 885 million over this period, each cell accounts for around 84,300 individuals. This is shown in row 1 of Table A.6, which reports information on average population by cell.7,8 Clearly, there is very large heterogeneity in the population of cells. While cells covering urban areas (as for example those overlapping with Lagos in Nigeria, population in the metropolitan area according to the most recent estimates 21 million over around 1,000 km2 , and Cairo in Egypt, population in the metropolitan area 17 million over 1,709 km2 ), host more than 5 million individuals, other cells in desert areas (in countries in Northern Africa such as Morocco, Egypt and Libya, and in Sub-Saharan Africa, such as Botswana, Mauritania, Namibia and Kenya) host less than 50 people. Since the contours of cells do not correspond typically to countries’ political borders, whenever we perform country level analyses we assign cells spanning over more that one country to the country which occupies the greatest area in any given cell.

2.2. Mobile Phone Coverage 2.2.1. Data sources: GSMA Data on mobile phone coverage are collected by the Global System for Mobile Association (GSMA), the association representing the interests of the mobile phone industry worldwide, in partnership with Collins Bartholomew, a digital mapping provider. The data come from submissions made directly from mobile operators and are used to construct the Collins Mobile Coverage Map, a roaming coverage map service used by network operators and their customers worldwide. The coverage refers to the GSM network, which is the dominant standard in Africa with around 96 percent of the market share (?).9 The data that have been licensed to us collate, for all years between 1998 and 2012 (but with the exception of 2005 and 2010), the most recent submission during that year from all member operators in each country, and provide geo-located information on mobile phone coverage separately for 2G, 3G and 4G technologies, aggregated across all operators. These technologies are incremental: 3G coverage is available only if 2G coverage was previously available. There is virtually no 4G technology in Africa over the period we consider. The extent of geographical precision of the coverage data depends on the quality of the operators’ submissions and ranges between 1 km2 on the ground (for high quality submission 7

The original data for population are available at 5-years interval, starting from 1995. We interpolate linearly - using logs - across subsequent years in order to obtain a measure of population in each year. 8 For comparison, these cells are similar to USA counties both in terms of population and extension. There are 3,143 counties in the USA over an extension of around 9.85 million km2 . At population of around 320 million, each county accounts for around 100,000 individuals and just over 3,000 km2 . 9 Based on restricted-use data from Collins Bartholomew we estimate that the operators submitting their data represent 86 percent of the total market share of African mobile operators.

8

based on GIS vector format) and 15-23 km2 (for submissions based on the location of antennas and their corresponding radius of coverage) (??). This allows us to measure the adoption of mobile phone technology at a very disaggregated geographical level since the onset. Our data improve considerably over similar data used in previous studies. Most cross-country studies typically use measures of mobile subscription density or penetration, which vary only at the country level (??). Studies at greater level of geographical detail, on the contrary, typically focus only on one country (??). By and large, these studies also use measures of mobile phone usage rather than availability and a concern here is that usage is potentially endogenous to the outcomes of interest. The only studies we are aware of that use detailed information on mobile phone availability at a fine level of geographical detail are ? and ? although these studies only cover a limited time span (respectively 1999-2006 and 2007-2009), hence being unable to capture the historical expansion in mobile phone technology across the African continent. 2.2.2. Descriptive statistics Figure 1 shows mobile phone coverage over the entire continent at 5-years intervals. While, as of 1998, only 3 percent of the African territory was covered by the mobile phone signal, by 2012 this figure was 27 percent. Figure A.1 in the Appendix zooms onto Nigeria, superimposing the lattice of 0.5◦ x 0.5◦ grid cells. One can visually appreciate the level of geographical detail allowed by our data together with the very rapid expansion in mobile phone infrastructure over the period. These figures clearly do not provide information on the fraction of population covered, as coverage is higher in more populated areas. We use information on the share of each cell’s area that is covered by mobile phone technology (and we assume that population is uniformly distributed within cells) in order to compute the fraction of individuals reached by the mobile phone signal in each cell. In the rest of the paper we use this measure as our primary measure of mobile phone penetration. We aggregate across cells using population weights to obtain country-level or continent-level measures of mobile phone penetration. Row 2 of Table A.1 reports the average population-weighted 2G coverage across the 19982012 period, showing that the probability that a random African citizen is in reach of the signal is 43 percent. Starting from an initial value of 9.2 percent of the population, the fraction of population covered is 63 percent in 2012. Looking at figures for the 3G technology, over the entire period only 2 percent of the population is in reach of the signal, a number that grows to about 6.3 percent in 2012. This very fast continental growth, clearly, masks large differences across countries. Figure2 shows that among early adopters, such as Egypt, Morocco, South Africa and Burkina Faso, coverage was virtually ubiquitous by the end of the period. This is in contrast with countries like Ethiopia, Guinea-Bissau and Mali where still less than 10 percent of the population was covered as of 2012. Our measure of coverage is the fraction of population that lives within range of a mobile

9

network signal, regardless of whether they actually subscribe to the service or use it. In order to gauge an understanding of what this implies in terms of mobile subscriptions, we compare our measure of mobile phone penetration with data on the of number of subscribers by country and year from ITU (2015).10 A regressions of the fraction of log subscribers over total population on the log fraction of individuals covered by the 2G signal (plus .0001 to allow for zeros), controlling for country and year fixed effects leads to an estimated coefficients of 0.30 (s.e. 0.02), implying that a 10 percent increase in coverage is associated with an increase in mobile phone subscriptions of 3 percent. This is possibly a lower-bound estimate of the true relationship between coverage and actual usage, given the widespread practice of phone sharing in Africa (?).

2.3. Political Mobilization 2.3.1. Data sources: GDELT, ACLED and SCAD Our first source of data on political mobilization come from the Global Dataset on Events, Location and Tone (GDELT) (?), an open-access database which collects information on the occurrence of political events, including protests, worldwide.11 Data in GDELT refer broadly speaking to political events in the area of verbal and physical mediation and conflict (Make a public statement, Consult, Threaten, Disapprove, etc.), including protests but excluding events that make part of the routine political process, such as those pertaining to elections, the legislative debate, and government actions that do not fall into the category of mediation or conflict. The full dataset contains more than 300 million fully geo-coded records of daily events from 1979 to nowadays for the entire world. The elementary observation in GDELT is an event per day. Out of the 20 primary event categories in the data, we focus on “Protests”, defined as “civilian demonstrations and other collective actions carried out as a sign of protest against a target”.12 Within the broad category of protests, the data allow us to distinguish between four categories: demonstrations, riots, strikes and others. As for many other similar and more widely used datasets discussed below, information is gathered from newswires.13 Compared to other datasets, though, which are typically based on manual coding, the data collection and the geo-referencing process in GDELT are executed through an automated coding of digitalized newswires.14 10

Average coverage rate for the sample of country/year observations for which ITU data are available is 49 percent versus 43 percent for the entire set of countries/years for which GSMA data are available. The fraction of subscribers over population for this sample is 30 percent. 11 Data are available at www.gdeltproject.org/data. 12 These exclude verbal protests. 13 Events in GDELT come from both digitalized newspapers and news agencies (Africa News, Agence France Presse, Associated Press Online, Xinhua, BBC Monitoring, The Washington Post, New York Times...) as well as from web-based news aggregators such as GoogleNews, which gathers around 4,000 media outlets (?). 14 The data are extracted using an open-source coding algorithm, TABARI, or Textual Analysis by Augmented Replacement Instructions, that sifts through news articles in search of actions and actors available in CAMEO, the Conflict and Mediation Event Observation, a widely used coding system in the field of political science

10

For each event the data report the exact day of occurrence and precise location, as well as the number of sources and articles across sources that refer to the same event. Importantly, the data do not provide any information on the issue at stake, the number of participants or the original news sources.15,16 Appendix Figure A.2 reports GDELT data on protests in Cairo in 2011 to show the level of geographical detail allowed by our data. There are as many as 70 different landmarks, with the size of the circles indicating the number of episodes in each precise location. Events in Tahrir Square and Cairo University are easily recognizable, but other episodes and locations that are probably less familiar to readers, such as the recurrent strikes in the industrial district of Helwan in the southern suburbs of the city, are also identified. Since GDELT is a yet largely unutilized dataset and in order to probe the robustness of our analysis to the measures of protests used, we complement the analysis with data from two manually complied sources. ACLED, or the Armed Conflict and Location Event Dataset (v.4), is a database reporting information on the exact location, date, and other characteristics of politically violent events in Africa.17,18 Similarly to GDELT, the data report the precise information on the date and location of occurrence of a protest but no information on the issue and the number of participants (although they do report the original news source). Differently from GDELT, the dataset does not report information on the exact typology of event (e.g. whether a rally or a riot). The last dataset we use is SCAD, the Social Conflict Analysis Database.19 A joint effort by researchers at the University of North Texas and the University of Denver, SCAD is a specialized datasets on social) conflict.20 As in the other datasets described above, the data report the precise date and location of occurrence of each event. As in GDELT but differently from ACLED, the data allow users to distinguish subcategories of events (Demonstrations, riots that provides a list of around 15,000 actions and 60,000 political actors. A precise location at the level of city or landmark is assigned to the event using the GeoNames gazetteer, which includes over 8.5 million toponyms for 6.5 million places with 2 million alternate names in up to 200 languages (?). Multiple references to the same event in one or more articles from the same source are collapsed into a single record and then events are further de-duplicated across different sources. 15 The data also report information on the actors involved - both the source and the target - although this is missing for many events and for this reason we do not to use this information in the analysis 16 In a comprehensive study of protests worldwide ? cite, in order, the following reasons for protests which occurred during 2006 and 2013: Economic justice and Anti-austerity, Failure of political representation, Global justice, People’s rights. Most protests are against national governments. 17 Data are available at www.acleddata.com/data/. The dataset focuses primarily on political violence that occurs during a civil war, instability or state failure, and as such it has been used widely in the literature on civil conflict (e.g. ???). However, events that are potential precursors or critical junctures of conflict, like protests and riots during peaceful times, are also recorded. We focus on these events, which represent around 20 percent of the total number of records in ACLED, in order to construct a measure of political mobilization separate from civil conflict 18 Events in ACLED are manually compiled from local, regional, national and continental media and are supplemented by NGO reports. The number of different sources used in ACLED has increased from 72 in 1998 to 232 in 2012 (ACLED, 2012). As in GDELT, events are atomic in that they are coded by day. 19 Data are available at https://www.strausscenter.org/scad.html. 20 Information is manually compiled starting from Associated Press and Agence France Presse newswires and verified against external sources.

11

etc.).21 In the rest of the analysis, we focus on episodes that occurred in Africa between 1998 and 2012 (2011 for SCAD), the period over which information on mobile phone coverage is available to us. We focus on events for which we have precise geo-location, typically at the locality or landmark level.22 2.3.2. Descriptive statistics Table A.1 reports descriptive information on micro data on protests from the different sources. There is a marked difference in the number of observations across datasets. While GDELT records a total of 168,890 days of protest over the entire period, this number is only 11,322 and 22,848 for ACLED and SCAD respectively. Demonstrations account for around two thirds of total protests in GDELT, with the second largest category being riots. The share of demonstrations in SCAD though is only 25 percent, with the residual category (which includes also violence from the government or militant organizations and which perhaps is not classified as protests in GDELT) accounting for 44 percent of all observations. One reason why the number of observations in GDELT is much larger than in ACLED and SCAD is that this is less likely to suffer from type-1 error, whereby truly occurring protests fail to be reported or are misclassified. Indeed, compared to manually complied datasets, machine coded datasets have typically low rates of false negatives (?) and an independent appraisal of GDELT suggests that this performs particularly well in this respect even compared to other automated coded datasets (?). On the other hand it is possible that GDELT data suffer from higher rates of type-2 error compared to the other datasets, whereby events which are not protests are incorrectly classified as such. A related and possibly even more serious problem is that, although an attempt is made to collapse multiple reports of a unique event into a single record, the algorithm might fail to do so if the variables that uniquely identify an event differ across newswires or they are missing. This problem is likely to be more serious in later years, due to the proliferation of electronic newswires. In order to deal with this issue and minimize the amount of false positives, in the following we also present results where instead of using raw GDELT data, we aggregate into a single event all those in the same category (e.g., demonstrations, riots etc.) happening in the same location in the same day (irrespective of the actors involved and whether actors are reported). This gives a total number of protests in GDELT of 78,533 (as opposed to 168,890 in the raw data). In order to combine information on protests with information on coverage of mobile phone technology, we assign protests to the 0.5◦ x 0.5◦ cells where they occurred and we compute 21

Differently from GDELT and ACLED, SCAD reports the start and end date of each event. For consistency with the other datasets, we consider each day of protest as a separate event. 22 This leads us to drop 28 percent of the observations in GDELT, 5 percent of the observation in ACLED and around 15 percent of observations in SCAD.

12

the total number of events falling within each cell boundary in each year. We then divide this number by the cell population to obtain a measure of intensity of political mobilization in each cell in each year. Rows 3 to 5 of Table A.6 reports summary statistics on the number of protests per 100,000 population during the period 1998-2012 based on the data sources. Data are aggregated across cells using population weights. On average, over the entire continent between 1998-2012, GDELT records 1.24 protests per year every 100,000 population. Data from ACLED and SCAD provide an estimate of the intensity of protests per 100,000 individuals in the order of 0.08 and 0.17 respectively. In order to investigate the relationship between GDELT, ACLED and SCAD, in figure 4 we report the evolution in protests per capita measured in the three datasets across each country separately. As the scale of the different series varies across countries, we report the residuals from regressions of the log of protests per capita (plus one, to account for zeros) on country dummies and year dummies. In practice these figures report the evolution of protests within countries net of continent-wide trends in protests. One can clearly see, that there is a very strong correlation between the three series, although this is far from being true in all countries.23 In countries like, for example Algeria, Burkina Faso, Ethiopia, to name a few, one can see that the series line up remarkably well. This is less true in countries like the Central African Republic or Lesotho. One can also see an increase in protests around 2008 in countries like Mauritania or Guinea which experienced food riots. The variation in the data is - in all cases -dwarfed by the very rapid surge in protests at the beginning of the current decade, with clear spikes in countries like Algeria, Egypt, Libya, Morocco and Tunisia, where the Arab Spring took place.24 Perhaps trends in protests across the continent can be better appreciated in Figure 3 that reports the evolution of protests per capita (in logs) over the entire continent during the fifteen years of observation. One can see a pronounced positive trend in the incidence of protests, with an overall increase of almost 30 log points over the period. One can notice a temporary increase in 2008-09, when the food riots exploded and a very strong increase in 2010-12 when the Arab Spring swept part of the continent.25 23

Note that the scale of the graphs differs across countries. The series in Figure 4 refer to average protests per capita in each country/year. As ultimately our analysis focuses on cells within countries, we also explore the correlation between protests per capita from the different sources across these cells. Figure A.3 reports on the vertical axis the intensity of protests per capita measured in GDELT and on the horizontal axis the intensity of protests from ACLED. Both series are obtained as residuals of logs of the relevant variables (plus 1 to account for zeros) on cell and country fixed effects, separately for each country. Regressions are weighted by population size. We superimpose to the data an estimated regression line, separately for each country. The pooled regression coefficient across all countries alongside the associated standard error clustered at the level of cell is reported at the bottom of the figure. One can clearly see, that even within countries, there is a very clear positive correlation between the two series. This is true in almost all countries, and the pooled regression coefficient of protests from GDELT on protests from ACLED is of 1.870 (s.e. 0.140). A very similar picture emerges from an analysis of the correlation between GDELT and SCAD (in Figure ??). Taken together these figures suggest that, despite some unavoidable measurement error, the three series convey very similar information. 25 This increasing tends in protests is continent with evidence from ? that the number of protests and protesters worldwide has increased in recent years. As said, though, one cannot rule out that the secular increase in protests is also likely to reflect an increase in reporting due to the rapid surge of media coverage and the 24

13

Alongside trends in log protests per capita, Figure 3 reports average GDP growth (the dotted line) across the continent over this period. This is weighted by country population size. One can clearly see that protests are strongly anti-cyclical, consistent with the evidence cited in the introduction that protests are more likely to occur when reasons for grievance abound and when the opportunity cost of participating falls, both of which are more likely to occur during recessions.26 We revert to the effect of economic conditions on the incidence of protests later on in the paper when we present our regression analysis. Figure 5 finally reports the cross-sectional correlation between 2G coverage and protests per capita (again expressed in logs) across all countries. The data illustrate a clear positive correlation between these two series, with countries with full coverage, such as South Africa, showing rates of protests per capita around 50 log points higher than countries with virtually no coverage, such as Ethiopia. Clearly, one cannot infer any causality from these data, and in the next section we turn explicitly to the econometric analysis in an attempt to identify the role of mobile phones in explaining differences in mass mobilization across countries.

2.4. Additional Cell-Level Characteristics We finally complement our data with a number of variables which are also available at the 0.5◦ x 0.5◦ grid-level. These data broadly characterize local economic conditions, availability of infrastructure, characteristics of the land, access to the capital and other close major cities, distance from the coast and the border, and climatic characteristics of each cell. Table ?? reports for each of these variables their definition and the original source, as well as a short description of the data, while rows 8 and below in Table A.6 report their corresponding summary statistics, again weighted by country population. Note that these variables (except income and population, which are available every five years) only vary cross-sectionally.27 As anticipated in the introduction, in some of the regressions below we attempt to circumvent the potential spurious correlation between protests and mobile phone technology adoption using local lighting activity as instrument for mobile phones adoption.28 The data that we use are average flash density - the number of ground strokes per km2 per year - between 1995 and 2010 explosion of new media during the period. Although, given the nature of the data, we cannot tell these two effects apart, it is important to realize that in the regressions below we do rely on the pure time series variation to identify the effect of mobile phone coverage on the incidence of political mobilization, so this should not be a source of major concern. 26 A regression of log protests per capital on GDP growth plus a linear trend delivers a coefficient of -3.894 (s.e. 0.093), meaning that a 1 percentage fall in in the rate of GDP growth is associated to an increase of protests per capita of around 4 percent 27 Indeed, we have information on temperature and precipitation at the yearly level although not for all years in the data. For these two variables we use the average across all available years. 28 Data on lightning activity come from the Global Hydrology and Climate Center (GHCC), which makes publicly available the data collected by the National Aeronautics and Space Administration (NASA) through spacebased sensors (?). Flashes are recorded along with their spatial location (latitude, longitude) with a level of resolution of at least 10 km on the ground. Data are available at www.thunder.msfc.nasa.gov/data/data_lisotd-climatology.

14

in each of the 0.5◦ x 0.5◦ cells.29 Figure A.4 in the Appendix illustrates the density of flash activity for the whole of Africa. The continent has the highest flash density on earth, with an average of 13.7 flashes per km2 per year, compared to a world average of 2.9 (?).30 The last row of Table A.6 reports populationweighted summary statistics for flash density for the entire continent.31

3. Regression analysis In the previous section we have shown that protests respond to the state of the economic cycle, increasing during recessions and falling during booms. As said, worsening economic conditions might increase the incidence of protests because they provide reasons for grievance and because the opportunity cost of participating in mass mobilization might decrease when economic condition deteriorate. In this section we use regression analysis to investigate whether the response of protests to the state of the economic cycle depends on the availability of mobile phones. Although one would not expect mobile phones per se to affect the incidence of protests, our discussion in the introduction suggests that mobile phones might amplify the effect of economic downturns on the incidence of mass mobilization. In order to measure the state of the economic cycle, we use - as before - the growth rate in the country’s GDP. Our unit of observation in the regression is a country cell per year. Note, importantly, that we focus on country- rather than cell-level economic shocks, and we study how these shocks translate into differential local protest activity as a function of local mobile phone coverage.32 If we denote a generic cell by j, with j ∈ c, where c denotes a country and t denotes a year and ignoring other covariates, our main regression model is:

29

The data have been used before by ?, who also show that flash density is very persistent across areas. The peak annual flash rate is 160 flashes per km2 per year in eastern Congo, corresponding to almost half a million flashes per year in each cell of the area. A broad region of central Africa exceeds 30 flashes per km2 per year while most land regions in the tropics and subtropics - except for arid regions - exceed 10 flashes per km2 . 31 The population weights are responsible for the difference between our average flash density and that in ?. 32 There are two reasons for focusing on country - as opposed to local - economic shocks. First, it is difficult to recover measures of economic activity at a level of detail of the PRIO-grid cells. Although ? provides measures of local GDP, these measures are only available every five years and there are good reasons to believe that these are affected by considerable measurement error. Others in the literature - reviewed in the introduction - focus on variations in local precipitation or temperature, which is available in PRIO-Grid. One exception is ? who construct their own measure of local economic shocks based on the main crop cultivated in each 1◦ x 1◦ degree cell. Second, and perhaps more important, there is a concern that local economic shocks and local trends in mobile phone coverage are simultaneously determined. Even if one has a sources if variation in local economic conditions that is credibly exogenous to both coverage and protests, there is still a concern that coverage might itself respond to local economic development, making it and an endogenous variable in the model, hence affecting the consistency of our estimates. For these reasons and in order to simplify the analysis we decided to focus on macro-economic shocks. 30

15

Yjct = β0 + β1 Covjct + β2 Covjct ∆lnGDP ct + djc + dct + jct

(3.1)

where Yjct denotes the log number of protests (plus 1 to account for zeros) per 100,00 people in each cell, Covjct denotes the fraction of each cell’s population that is covered by mobile phone signal, ∆lnGDPct is the country-level rate of growth of GDP and djc and dct are respectively cell fixed effects and unrestricted country X time effects. jct denotes the error term. The coefficient β1 captures the effect of coverage on protests at zero GDP growth while β2 measures how protests vary as a function of coverage at different points of the economic cycle.33 The coefficient β2 is the coefficient of primary interest: this measures how economic downturns, translate into protests in areas with different mobile phone coverage. For mobile phones to magnify the effect of economic downturns one would expect the coefficient β2 to be negative. In this specification the coefficient β1 identifies how protests vary across areas with different levels of coverage at zero economic growth. Note that, in the spirit of an ITT approach, we use coverage as opposed to actual mobile phone usage or subscriptions in a certain area. This should alleviate concerns that take-up is itself endogenous, a problem that as discussed in the introduction plagues other studies in the area. Ignoring for now other covariates, identification of the model is based on a differences in differences strategy that compares changes in the incidence of protests across cells within the same country experiencing differential trends in the adoption of mobile phone technology. By conditioning on country X time effects, our regressions implicitly control for any time varying country specific determinant of protest and ICT adoption that would contaminate our regression coefficients. Consistency of the estimates relies on the assumption that, other than for differential trends in mobile phone coverage, trends in protests per capita would be similar across areas within a country. Clearly there are good reasons to be skeptical about this assumption, as unobserved determinants of mobile adoption might affect the incidence of mass mobilization even within a country. In order to deal with this issue, we also present regression results where we include in the model the few time-varying characteristics that are available at the level of cells over the entire period (namely log population and income) plus a very large array of observable cross sectional cell-level characteristics (see Table A.6) that we interact with country-specific linear trends.

33

Note that due to the inclusion of country X time effects, the coefficient on ∆lnGDP ct cannot be separately identified.

16

3.1. Main regression results: OLS Table 2 presents the main regression results, separately for protests measured in GDELT (columns 1 to 4), ACLED (columns 5 to 8) and SCAD (columns 9 to 12).34 We present specifications (in odd-numbered columns) where we exclude the interaction term between coverage and GDP growth (i.e. we constraint β2 to be equal to zero), meaning that we investigate whether coverage per se affects the incidence of protests, and specifications (in even-numbered columns) where we also include the interaction term. The latter is the specification in equation 3.1. For each dataset, we report regressions with no controls other than cell and country X time effects in the first two columns, and with all additional controls in the two remaining columns. Additional controls include log population and log per capita income plus distance to the closest cell with a drought in that year, as well as a large number of cross-sectional cell-characteristics interacted with a linear time trend. Cell characteristics include: fraction of the cell’s area covered by mountains, forests and irrigated; dummy for the presence of mines, diamonds and oilfields in the cell; distance of cell centroid to the capital, the coast and the border; travel time to the closest city with more than 50,000 inhabitants; km. of primary roads, of paved primary roads and primary road in good conditions, km. of secondary roads and km. of electrical grid in the cell; infant mortality rate; average temperature and precipitation, percentile of population size, plus fixed effects for second administrative level units (typically districts). All regressions are weighted by population size and standard errors are clustered by cell. Results in even-numbered columns consistently show no correlation between coverage per se and protests. When we turn to the interaction effects, in even numbered columns, one can clearly see that this has a negative coefficient and is statistically significant across all datasets, although magnitudes vary. Results are virtually unaffected by the inclusion of additional controls. Estimates based on GDELT, with all controls (-2.038, s.e. 0.447) imply in particular that a 5 p.p. fall in GDP growth (around one standard deviation, roughly the change from zero growth to a deep recession) is associated to an increase in the differential in the yearly protests rate between an area with full coverage and an area without coverage of around 10 percent. Point estimates are smaller in magnitude for ACLED and SCAD. Estimates from ACLED and SCAD imply that a 5 p.p. fall in GDP growth is associated to an increase in the differential protest rate across areas with different rate of mobile phone coverage of between 2.5 and 2 percent, around a fourth of what found in GDELT. Although still statistically significant at conventional levels, these estimates are typically less precise than estimates based on GDELT. This should be no surprise given the much larger number of observations in the latter data set compared to the former. One possible reason why estimates based on GDELT are sensibly 34

Given the very high number of fixed effects and controls we perform our regressions via partitioned regression. In practice we regress both the dependent variable and the independent variables on year dummies, cell fixed effects plus - if applicable - the controls, separately by country using GLS. We derive residuals from these regressions and we regress the residuals of the dependent variable on the residuals of the dependent variables, again using GLS. We use this procedure as opposed to readily-available routines in Stata to handle multiway fixed effects as the latter typically do not allow to perform weighted regressions.

17

larger than those based on the other two datasets is measurement error. The concern here is that GDELT might be over-reporting protests (type-2 error) - when they occur and this might itself be correlated with the availability of mobile phones. By the opposite token, ACLED and SCAD might suffer from higher rates of type-1 error , i.e. fail to report protests that actually occurred, and this might itself be correlated with the availability of mobile phones. One way to address this issue is to use our alternative definition of protests in GDELT that de-duplicates events occurring in the same day in the same location. This is a conservative approach that possibly leads to an undercount of protests in GDELT. Regression results on this alternative measure of protests (reported in Appendix Table A.4) show indeed that the estimated coefficient falls modestly in value, from -1.9 in Table 2 to -1.6 and it remains highly significant at conventional levels. Although this suggests that indeed type-2 error might be a source of concern in GDELT, point estimates are still around 3 times as large as those from ACLED and SCAD pointing into the direction of considerable type-1 error in these two manually compiled data sets.

35 36

3.2. Additional results: OLS Table ?? presents a variety of additional estimates meant to probe the robustness of our results and investigate the heterogeneity in effects across different samples. All regressions include the entire set of controls as in columns (3) and (4) of Table 2 and again regressions are weighted by population size while standard errors are clustered by cell. Columns (1) to (3) of Table ?? refer to GDELT data. Column (1) presents results for cells other than those containing the centroid of the each country’s capital city. The concern here is that protests might be concentrated in capital areas, where also mobile phone coverage adoption is higher and speedier. A related concern is that protests might take place in capital cities but that individuals might be living elsewhere. In this case one might erroneously assign high coverage to those who live in low coverage areas. By focusing on areas outside the capital city we attempt to ease these concerns. Column (2) excludes data for 2011 and 2012. We want to reassure a reader that our results are not driven by the Arab Spring, when protests exploded in urban areas of countries in Northern Africa, areas that also witnessed a very rapid expansion in mobile phone coverage over the period. Again one might be concerned that these observations are fully driving our identification and this would raise suspicion that our results depend on unobserved heterogeneity across cells. Column (3) investigates whether there is any differential effect of 2G versus 3G technology. 35

Although we express protests per capita in the regressions in logs as a way partly control for the presence of a few very large outliers (see Table 1), as additional robustness checks we have also run our regressions using as a dependent variable, in turn, the log total number of protests, rather than the log total number of protests per capita, the square root of protests per capita (which behaves approximately like logs) ad total protests per capita (in levels) regression results, not reported but available upon request, are qualitatively very similar, implying that our choice of functional form for the dependent variable does not drive our results. 36 We have also experimented with standard errors clustered at second administrative level.

18

In principle, one would expect 3G technology to further magnify the effects of recessions on protests, as in addition to phone calls and SMS messaging, 3G technology allows users to browse the internet and access social media. Recall that 3G technology is additive with respect to 2G technology, meaning that the effect of the former must be interpreted as the effect of upgrading from 2G to 3G. Recall also that 3G adoption was still very low in Africa during this period (see Table 1), so our ability to identify the effect of 3G technology is limited by the modest variability in the data. Columns (4) to (6) and (7) to (9) of Table ?? report the same specifications for ACLED and SCAD, respectively. Point estimates in column (1) show that results also hold outside capital cities. Although the coefficient on the interaction term is about half of what found when one focuses on the entire set of cells (-1.039 versus -2.038), this is still statistically and economically significant. Column (2) shows that results are not driven by year-country cells where episodes of mobilization typically associated to the Arab Spring took place. Column (3) finally shows that, consistent with what postulated, 3G technology further magnifies the effect of variations in the rate of GDP growth, with an effect around 7 times as large as the effect of 2G technology, although the point estimate is not statistically significant at conventional levels. Results for ACLED and SCAD in the subsequent columns are roughly in line with those from GDELT, although, unsurprisingly typically less precise. There is evidence that results still hold true outside capital cities and that 3G technology has a much larger effect than 2G technology.

3.3. 2SLS estimates Although in our model we exploit the diffusion of mobile phone across relatively small areas within countries and we include a very large set of baseline cell characteristics interacted with country specific time trends in an attempt to ease concerns that our estimates are biased, the concern remains that our estimates are contaminated by the endogenous adoption of mobile phones across areas. Ex- ante it is difficult to sign the direction of the bias. If local economic progress discourages participation in protests, either because there is less rationale for redistribution or because the opportunity cost of mass mobilization increases, one will expect OLS estimates to be upward biased. In other terms one will expect them to be conservative estimates of the differential effects of economic downturns on the incidence of protests. This effect might be reinforced by measurement error that is possibly affecting our measure of local mobile phone coverage. On the other hand, if local economic progress leads to citizen’s economic and social empowerment and this creates incentives for participation, perhaps because human and social capital are complements or because those in direr poverty are unable or unwilling to be involved in social movements, then the reverse is true. In an attempt to generate a source of arguably exogenous variation in the spread of mobile technology across small geographical areas, we exploit the differential adoption of mobile tech-

19

nology in areas subject to difference incidence of lightning. Lightning damages mobile phone infrastructures and in particular antennas on the ground that transmit the signal in their vicinity and negatively affects connectivity, acting on both supply (as power surge protection is costly) and demand (as poor connectivity makes the investment in technology less profitable and the risk of intermittent communications discourages adoption) (??). One will expect hence to see a slower adoption of mobile phone technology in areas with high intensity of lighting. As there is substantial variation in lighting across areas (see section 2.4), this instrument has the potential to generate variation in the rate of mobile phone adoption across cells. In particular we instrument mobile phone coverage using the interaction between average lighting activity in a cell, i.e. the average flash rate, denoted by F lashjc and a linear time trend. In formulas the instrument is:

Zjct = F lashjc Xt

(3.2)

where F lashjc is number of flash rates per km2 and t is a linear time trend. and the first stage equations are:

Covjct = δ0 + δ1 Zjct + δ2 Zjct ∆lnGDP ct + djc + dct + ηjct

(3.3)

Covjct ∆lnGDP ct = θ0 + θ1 Zjct + θ2 Zjct ∆lnGDP ct + djc + dct + µjct

(3.4)

For this approach to deliver consistent estimates of impact, one requires differential trends in lightning across areas to be uncorrelated with unobserved determinants of protests. Although lighting might itself affect economic activity and population growth, and this exclusion restriction is likely not to hold unconditionally, we speculate that this is likely to hold once we condition on the interaction of a large array of cell level characteristics also interacted with trends. Table A.7 in the appendix reports the first stage estimates for the impact of flash rate density on mobile phone coverage. We only present regression results with the entire set of controls. Columns (1) presents first stage estimates for coverage (equation 3.3) where we restrict the coefficient δ2 to be equal to zero. This provides a first indication of the correlation between mobile phone penetration and flash rate activity. Columns (2) and (3) report estimates of equations 3.3 and 3.4 respectively where we now also include the interaction term between the instrument and the rate of growth in GDP. Note that for the model to be well specified one will expect the coefficient θ2 to be equal to the coefficient δ1 . Note that also that all coefficients are multiplied by 100. Consistent with what postulated above, column (1) shows that a 1 s.d. increase in the flash

20

rate (13.84) leads to a differential growth in coverage of around 0.2 percentage points a year, i.e. a differential growth of 3 percentage points over the entire period. The estimates remains roughly similar (around a 4 percentage points differential growth across areas a 1.s.d. apart in flash rates when we further include the interaction between the instrument and the growth rate in GDP in column 2 (although that greater the growth in the country’s GDP the lower the gap between areas with high and low flash rate but this effect is relatively modest in magnitude). As expected, column (3) shows that the effect of the interaction of the instrument with GDP growth on the interaction between coverage and GDP growth (-0.029) is similar to the effect in levels in column (2) (-0.021). In sum the data present cellars evidence that flash rates tend to lead to slower adoption of mobile phone technology even controlling for a very large set of baseline covariates that are like to affect protests. 2SLS estimates are reported in Table 3. Again we restrict to specifications with all controls. The estimated coefficients on the interaction terms increases by a factor of around five (-10.163), suggesting - if anything - that the OLS estimates provide a lower bound for the effect of interest. We find qualitatively similar results for ACLED, with the estimated coefficient increasing from -0.423 in Table 2 to -3.516 when we instrument for flash rates. Results for SCAD are positive but small and statistically insignificant, consistent with the observation above that SCAD data lead typically to less precise estimates. Table 4 reports additional results, as in Table ??, using the IV approach. Results, although typically less precise, square well with the OLS estimates and are typically larger in magnitude.

4. Conclusions In this paper we use novel geo-referenced data on the coverage of mobile phones across the whole of Africa between 1998 and 2012 together with a variety of geo-referenced data on protests (i.e. marches, riots, strikes) from compilation of newswires plus auxiliary data from a variety of sources to test whether the availability of mobile phones in a local area leads to greater political mobilization. Our analysis illustrates that, although on average greater availability of mobile phones does not translate into more protests, protests are anti-cyclical, i.e. more likely to occur during bad economic times and that this effect is magnified by the availability of mobile phone technology in a given area. The OLS estimates suggest that a five percentage point - effectively a one s.d. - fall in a country’s rate of GPD growth leads to an increase in the incidence of annual protests per capita in an area with full coverage relative to an area with no coverage of between 2 to 10 percent, depending on the source of data. Results are one order of magnitude larger for 3G technology, which allows for data downloading and access to social media, compared to 2G technology. Results are not driven by capital cities or episodes of mass mobilization during the Arab spring.

21

The IV estimates that exploit the differential rate of adoption of mobile phone technology across areas with different exposure to lightning activities point into the direction of the OLS estimates providing lower bound estimates for the effect of interest, with a bias as large as 80 percent. We ascribe this to the fact that greater economic growth, which is likely to be positively correlated with local adoption of mobile phone technology, is also associated with lower protests per capita, explaining why the OLS estimates tend to be conservative. Although we remain agnostic about the economic and political consequences of protests, there are good reasons to speculate that they can lead to redistribution and promote democracy, with returns among those who are the most deprived in society. We also remain agnostic on the precise channels through which the availability of mobile phone technology acts as an amplifier for protests during bad economic times. We have speculated in the introduction that both information provision - including through the public nature of the signal that individuals receive - and improved coordination, whereby mobile phones allow to communicate one’s and learn about others’ willingness to participate, are likely to explain our effects. One caveat to our analysis is that, due to the nature of the protest data, greater incidence of protests in areas with higher coverage might depend on a pure reporting effect. This is the case if mobile phones increase the probability that an event is reported in the news. Although the evidence that we provide in the paper suggests that possibly this is not a source of major concern, we are planning to integrate our analysis with micro data from the Afro-Barometer on actual (as opposed to reported in the news) political mobilization. The Afro-Barometer in particular provides information on citizens’ participation in marches and protests (in addition to voting, social capital, as well as political views) for 27 (out of 48) African countries for up to three waves. One advantage of the Afro-Barometer data is that they provide information on a household district of residence: this allows us again to link information on local mobile phone coverage to actual political participation and explore heterogeneity across different individuals. This analysis is next on the agenda.

22

Figure 1 2G Diffusion, Africa 1998-2012

Notes. The figure reports geo-referenced data on 2G coverage for the entire of Africa at 5-year intervals between 1998 and 2012. Source: GSMA.

23

24

Notes. The figure reports the fraction of the population covered by 2G technology by country and time. Series are obtained as a population-weighted averages of the fraction of each 0.5◦ x 0.5◦ degree cell that is covered by the signal in each year. Cells belonging to different countries are assigned to the country that spans the largest area within that cell.

Figure 2 Fraction of population covered by 2G by country and time

Figure 3 The evolution of GDP Growth and Protests over time - Africa

Notes. The figure reports continent-wide (log) protests per 100,000 people (dashed line) and the rate of GDP growth (dotted line) as a function of time. Continent-wide GDP growth is obtained as a population-weighted average of GDP growth in each country.

25

26

Notes. The figure reports log protests per 100,000 individuals in GDELT (solid line) and ACLED (short-dashed line) by country and year. Residuals from regressions on country and year fixed effects reported.

Figure 4 Trends in protests per capita

Figure 5 Cross-Sectional Relationship between Coverage and Protests

Notes. The figure reports log protests per 100,000 individuals based on GDELT on the vertical axis and the fraction of the population covered by 2G signal on the horizontal axis. Averages between 1998 and 2012 by country reported.

27

Table 1 Descriptive statistics Avg.

Std. Dev.

Min.

Max.

P opulation (1000s)

84.32

266.78.5

0

12,860

M obile P hone 2G Coverage (percent) M obile P hone 3G Coverage (percent)

0.43 0.02

0.42 0.09

0 0

1 1

P rotests per 100, 000 pop. − GDELT P rotests per 100, 000 pop. − ACLED

1.24 0.08

17.29 0.688

0 0

3,000,000 1,146.13

Country GDP growth (percent)

0.049

0.041

-0.33

0.63

Border Distance (100 km) Capital Distance (100 km) Coast(dummy) P rimary Roads (100 km) P rimary Roads P aved (100 km) P rimary Roads Good Conditions (100 km) Secondary Roads (100 km) Electricity N etwork (100 km) T ravel T ime nearest city pop. ≥ 20K(hours) T ravel T ime nearest city pop. ≥ 50K(hours) Inf ant M ortality Rate (h) M ountain (percent) F orest (percent) Irrigated (percent) Diamonds (dummy) M inerals (dummy) Oil (percent) T emperature (Celsius degrees) P recipitation (mm.) Drought (n. of years) Distance f rom drought (100 km) F lashrate (per Km2 per year)

1.73 3.57 0.15 0.87 0.49 0.26 1.42 0.86 4.42 4.21 8.91 0.23 0.23 0.08 0.03 0.22 0.13 23.12 876.2 1.44 1.74 17.32

1.47 3.35 0.36 0.99 0.72 0.49 1.10 1.18 3.77 3.69 3.71 0.32 0.25 0.17 0.18 0.42 0.33 4.25 487.5 1.25 0.56 13.80

0 0.04 0 0 0 0 0 0 0.16 0 1 0 0 0 0 0 0 4.06 69.39 0 0 0

10.54 19.48 1 5.22 4.66 3.80 6.40 7.55 106.9 102.2 20.31 1 1 0.87 1 1 1 31.41 3,296.4 11 4.56 163.1

Notes. The table reports descriptive statistics for each of the 10,419 cells of 0.5◦ x 0.5◦ degree resolution that compose Africa (excluding Somalia). All data, except population in row 1, are weighted by cell population. Row 2 and 3 report the fraction of each cell area covered by 2G and 3G technology respectively. Rows 4 and 5 report the average number of protests in a year per 100,000 people, from GDELT and ACLED respectively. Row 6 reports the country growth in GDP per capita. The residuals rows report cross-sectional physical, climatic, geographical and economic characteristics of each cell. Table A.2 reports a short description as well as the original source of each cell characteristic.

28

Table 2 Baseline Least Squares Regressions GDELT

Coverage

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

0.007 (0.022)

0.093*** (0.033)

-0.007 (0.020)

0.091*** (0.032)

-0.006 (0.009)

0.015 (0.015)

-0.004 (0.006)

0.017 (0.013)

Coverage ∗ ∆lnGDP

Additional Controls

ACLED

-1.623*** (0.478)

-1.891*** (0.449)

-0.384* (0.233)

-0.397* (0.230)

No

No

Yes

Yes

No

No

Yes

Yes

N. of cells

10,368

10,368

10,368

10,368

10,368

10,368

10,368

10,368

Observations

152,415

152,415

152,415

152,415

152,415

152,415

152,415

152,415

Notes. The table reports separate regressions of log protests per 100,000 people (plus one to account for zeros) by cell and year. Columns (1)-(4) refer to GDELT; columns (5)-(8) refer to ACLED. Coverage is the fraction of each cell area covered by mobile phone signal in a given year. Coverage ∗ ∆lnGDP is the interaction of this variable with the country yearly GDP growth rate in that year. Columns (1), (2), (5), (6) only control for country X years effects plus cell fixed effects. Columns (3), (4), (7), (8) also include log population and log per capita income plus distance to the closest cell with a drought in that year, as well as cross-sectional cell-characteristics interacted with a linear time trend. Cell characteristics include: fraction of the cell’s area covered by mountains, forests and irrigated; dummy for the presence of mines, diamonds and oilfields in the cell; distance of cell centroid to the capital, the coast and the border; travel time to the closest city with more than 50,000 inhabitants; km. of primary roads, of paved primary roads and primary road in good conditions, km. of secondary roads and km. of electrical grid in the cell; infant mortality rate; average temperature and precipitation, percentile of population size, plus fixed effects for second administrative level units (typically districts). Dummies for missing values of each of the controls also included. All regressions are weighted by the cell population. Standard errors clustered at the level of cell reported in brackets. ∗ Significantly different from zero at the 90% level, ∗∗ 95% level, ∗∗∗ 99% level.

29

Table 3 Instrumental Variable Regressions GDELT

Coverage

(1)

(2)

(3)

(4)

-0.790 (0.520)

-0.272 (0.488)

0.087 (0.115)

0.294* (0.150)

Coverage ∗ ∆lnGDP

Additional Controls

ACLED

-7.976** (3.562)

-3.195** (1.555)

Yes

Yes

Yes

Yes

N. of Cells

10,368

10,368

10,368

10,368

Observations

152,145

152,145

152,145

152,145

Notes. The table reports similar regressions to those in columns (3), (4), (7), (8) of Table 2 where the variables Coverage and Coverage ∗ ∆lnGDP are instrumented respectively for average flash rates in a cell interacted with a linear time trend and the interaction of this variable with GDP growth. See also notes to Table 2.

30

Table 4 Heterogeneity Institutions

City Size

Arab Spring

Media

(1) Democracy

(2) Autocracy

(3) Small

(4) Large

(5) Pre-2011

(6) 2011-2012

(7) Free

(8) Captured

Coverage

-1.258 (0.809)

-0.661 (0.623)

-1.258** (0.639)

-0.724 (0.627)

-0.916 (0.558)

-0.591 (1.045)

0.849 (1.268)

-1.140* (0.662)

Coverage ∗ ∆lnGDP

-3.467 (4.028)

-11.518*** (3.913)

1.006 (1.322)

-9.239*** (3.309)

-6.381** (3.152)

-12.772** (5.642)

-5.322 (9.213)

-8.216*** (3.141)

Additional Controls

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

N. of cells

10,368

10,368

10,368

10,368

10,321

10,368

10,368

10,368

Observations

151,921

151,921

152,145

152,145

152,145

152,145

152,145

152,145

Notes. The table reports similar regressions to those in columns (2) and (4) of Table 3 where the variables Coverage and Coverage ∗ ∆lnGDP are instrumented respectively for average flash rates in a cell interacted with a linear time trend and the interaction of this variable with GDP growth. Columns (1) and (2) split the sample by the Polity Index. Autocracy is defined for Polity scores less or equal to zero. Columns (3) and (4) split by city size. Small cities are those with population below the median. Columns (5) and (6) split temporally. The “Arab Spring” is fro years 2011 and 2012. Columns (7) and (8) split by media freedom. The definition is based on the Reporters Without Borders World Press Freedom Index. Countries with captured media are below the median worldwide level of the index. See also notes to Tables 2 and 3.

31

Table 5 Afrobarometer. Individual Correlates of Protest Participation (1)

∆lnGDP

(2)

(3)

-0.149 (0.120)

(4)

(5)

(6)

-0.126** (0.055)

Disapproves of president

0.021*** (0.002)

0.022*** (0.002)

0.023*** (0.002)

Economy worsened

0.006*** (0.002)

0.007*** (0.002)

0.007*** (0.002)

Age/100

0.200*** (0.043)

0.196*** (0.043)

0.194*** (0.043)

0.187*** (0.029)

0.181*** (0.029)

0.180*** (0.029)

Age/100 sq.

-0.280*** (0.045)

-0.272*** (0.045)

-0.271*** (0.045)

-0.266*** (0.032)

-0.256*** (0.032)

-0.255*** (0.032)

F emale

-0.036*** (0.002)

-0.036*** (0.002)

-0.036*** (0.003)

-0.037*** (0.002)

-0.036*** (0.002)

-0.036*** (0.002)

Inf ormal ed.

-0.004 (0.007)

-0.003 (0.006)

-0.004 (0.006)

-0.007 (0.005)

-0.006 (0.005)

-0.006 (0.005)

Incomplete primary

0.007 (0.005)

0.008* (0.004)

0.008* (0.005)

0.005 (0.003)

0.006** (0.003)

0.007** (0.003)

Completed primary

0.014*** (0.005)

0.016*** (0.005)

0.016*** (0.005)

0.012*** (0.003)

0.014*** (0.003)

0.014*** (0.003)

Incomplete sec.

0.039*** (0.006)

0.041*** (0.006)

0.041*** (0.006)

0.036*** (0.003)

0.038*** (0.003)

0.038*** (0.003)

Completed sec.

0.032*** (0.006)

0.033*** (0.006)

0.033*** (0.006)

0.029*** (0.004)

0.031*** (0.004)

0.031*** (0.004)

Some tertiary (not college)

0.048*** (0.007)

0.049*** (0.007)

0.049*** (0.007)

0.047*** (0.004)

0.048*** (0.004)

0.048*** (0.004)

Some College

0.134*** (0.013)

0.135*** (0.013)

0.134*** (0.013)

0.134*** (0.006)

0.135*** (0.006)

0.135*** (0.006)

Completed college

0.075*** (0.011)

0.076*** (0.010)

0.076*** (0.010)

0.072*** (0.007)

0.073*** (0.007)

0.073*** (0.007)

P ost − graduate

0.115*** (0.027)

0.116*** (0.027)

0.114*** (0.027)

0.114*** (0.013)

0.114*** (0.013)

0.114*** (0.013)

Country X year

No

Yes

Yes

No

Yes

Yes

Cell-level controls

No

No

Yes

No

No

Yes

154,731

154,731

154,731

145,356

145,356

145,356

Observations

Notes. The dependent variable in all columns is a dummy variable equal to 1 if the respondent has attended a demonstration or protest march during the previous year. Columns (3) and (6) include all cell characteristics reported in notes to Table 2.

32

Table 6 Afrobarometer. Protest and Opinion about President Protest

Coverage Coverage ∗ ∆lnGDP

Cell-level Controls Observations

Distrust President

Economy Worse

(1)

(2)

(3)

(4)

(5)

(6)

-0.011 (0.023)

-0.033 (0.024)

0.024 (0.029)

0.007 (0.033)

0.061 (0.045)

0.013 (0.050)

-0.574*** (0.188)

-0.361* (0.214)

-1.050*** (0.342)

-0.966*** (0.344)

-1.202** (0.517)

-0.798 (0.601)

No

Yes

No

Yes

No

Yes

153,571

153,571

151,159

151,159

155,306

155,306

Notes. The dependent variable in columns (1) and (2) is a dummy variable equal to 1 if the respondent has attended a demonstration or protest march during the previous year. In columns (3) and (4) the dependent is a dummy variable equal to 1 if the respondent does not trust at all the president. In columns (5) and (6) is a dummy variable equal to 1 if the respondent thinks that the economy has worsened compared to the previous year.

33

A. Appendix Figure A.1 2G Diffusion, Nigeria 1998-2012

Notes. The figure reports the spread of 2G coverage in Nigeria between 1998 and 2012 at 5-years intervals. Source: GSMA.

34

Figure A.2 Geo-located events of protest, Cairo 2011

Notes. The figure reports the occurrence of protests in Cairo in 2012. Larger dots correspond to more days of protests in a certain location. Source: GDELT.

35

Figure A.3 Within-Country Correlation between GDELT and ACLED

Notes. The figure reports the correlation between protests in GDELT and ACLED within each country. Each point refers to a cell X year observation. All series are expressed in logs (plus one to account for zeros). Residuals from a regression on cell fixed effects and year X country fixed effects reported. A best-fit regression line obtained via GLS with weights equal to the population in each cell in each year also reported.

36

Figure A.4 Flash density in Africa

Notes. The figure reports average flash density across Africa. Source: NASA.

37

Figure A.5 Lightning and Protests, by Subperiods

Notes. The figure reports the estimated coefficients of F lash∗year ∗∆lnGDP and the corresponding 90% confidence intervals from pooled OLS regression.

Table A.1 Descriptive statistics - Protests micro data Obs.

Fraction GDELT (1998-2012)

Demonstrations

112,905

66.85

Riots

24,802

14.69

Strikes

10,645

6.30

Others

20,538

12.16

T otal

168,890

100

T otal

11,322

ACLED (1998-2012) 100

Notes. The table reports descriptive statistics on the number of protests in Africa - total and by type - between 1988 and 2012, respectively from GDELT and ACLED.

38

Table A.2 Cell-level covariates Variable

Source

Short Description

PRIO-GRID

Border distance is calculated from the cell centroid to the border of the nearest neighboring country, regardless of whether the nearest country is located across international waters. Capital distance is calculated from the cell centroid to the national capital city in the corresponding country. Geographical coordinates for the capital cities were derived from the cShapes dataset and captures changes over time wherever relevant. Coast is a dummy for the cell being coastal.

Africa Infrastructure Country diagnostic (ADB)

Geo-referenced roads files are downloaded separately for each country. Data usually refer to network in 2007, and are discussed at length in ?. For each cell we calculate the road network and its conditions in GIS. Country-specific data are missing for North-African countries. In these cases data are obtained from the Roads Africa dataset, but lack information on roads conditions.

Africa Infrastructure Country diagnostic (ADB)

Geo-referenced electricity files are available from the ADB dataset for all countries, except Egypt, Libya and Morocco. For these countries, data are obtained from the Openstreetmap project.Data usually refer to network in 2007. For each cell we calculate the electricity network in GIS.

Harvest Choice (IFPRI)

Calculated at the centroid of each cell in GIS. The raw data estimate accessibility using a cost distance function for each 1km pixel. The adjusted speed is based on road locations and type, elevation, slope, water bodies, and land cover.

PRIO-GRID

The cell-specific infant mortality rate is based on raster data from the SEDAC Global Poverty Mapping project. The variable is the average pixel value inside the grid cell. The value unit is the number of children per 10,000 that die before reaching their first birthday. The indicator is available for the year 2000.

M ountain (percent) F orest (percent) Irrigated (percent)

PRIO-GRID

Mountain is the proportion of mountainous terrain within each cell. This indicator is based on high-resolution mountain raster data from the UNEP’s Mountain Watch Report 2002. Forest is the percentage forest cover in a cell extracted from the Globcover 2009 dataset. Irrigation is the proportion of area equipped for irrigation within each cell from the FAO Aquastat irrigation raster. The original data were imported and modified to fit into the 0.5◦ x 0.5◦ PRIO-GRID data structure.

Diamonds (dummy)

Diamond dataset PRIO

A diamond occurrence is defined as any site with known activity, meaning production (either commercial or artisan) or confirmed discovery. For each cell we calculate the presence of a diamond mine in GIS.

Oil (percent)

Petroleum dataset PRIO

The petroleum dataset groups oil fields in polygons within a buffer distance of 30 km. Each polygon is assigned the geographic coordinates of its centroid. For each cell we calculate the percentage that is covered by an oil-field in GIS.

U.S. Geological Survey

The dataset reports mineral resource occurrence worldwide. The data include the exact location and the type of mineral, as well as the magnitude of production at the site. For each cell we calculate the presence of a mine in GIS.

Border Distance (100 km) Capital Distance (100 km) Coast (dummy)

P rimary Roads (100 km) − T otal − P aved − Good conditions Secondary Roads (100 km)

Electricity (100 km)

T ravel T ime nearest city : pop. ≥ 20K (hours) pop. ≥ 50K (hours)

Inf ant M ortality Rate (h)

M ineral (dummy)

T emperature (Celsius◦ ) P recipitation (mm.) PRIO-GRID Drought (n. of years) Distance f rom drought (100 km)

F lashrate (per Km2 per year)

GHCC/NASA

Temperature and precipitation are, respectively, the yearly mean temperature and total amount of precipitation in the cell, based on monthly meteorological statistics from the University of Delaware (NOAA 2011). We calculate the average for the period 1946-2008. Drought is calculated from within-year deviations in precipitation based on monthly data. The annualized measure is coded 1 if there were at least three consecutive months more than 1 s.d. away from the average monthly values. Distance from drought measures the distance to the nearest cell incurring a drought in the current year. Data refer to lightning activity calculated from the Optical Transient Detector (OTD) and the Lightning Imaging Sensor (LIS). Each flash is recorded along with its spatial location (latitude, longitude) with a level of resolution of 5-10 Km on the ground. 39 The GHCC calculates the average flash density in 0.5◦ x 0.5◦ grid cells over the period 1995-2010.

Table A.3 First Stage Regressions (1)

(2)

(3)

Dep. variable

Cov

Cov

Cov ∗ ∆lnGDP

F lash ∗ year

-0.025*** (0.008)

-0.026*** (0.008)

0.001 (0.001)

0.026 (0.038)

-0.031** (0.012)

Yes

Yes

Yes

F-stat

10.53

6.186

6.059

N. of cells

10,368

10,368

10,368

Observations

152,415

152,415

152,415

F lash ∗ year ∗ ∆lnGDP Additional Controls

Notes. The table reports regressions of Coverage and Coverage ∗ ∆lnGDP on average flash rates in a cell X year interacted with a linear time trend and the interaction of this variable with GDP growth. Regressions include entire set of controls. See also notes to Tables 2 and 3.

Table A.4 Least Squares Regressions. GDELT - one event per day

Coverage

(1)

(2)

(3)

(4)

0.009 (0.016)

0.080*** (0.025)

0.007 (0.015)

0.085*** (0.024)

Coverage ∗ ∆lnGDP

Additional Controls

-1.341*** (0.372)

-1.515*** (0.344)

No

No

Yes

Yes

N. of Cells

10,368

10,368

10,368

10,368

Observations

152,145

152,145

152,145

152,145

Notes. The table reports similar regressions to those in Table 2 (columns 3, 4, 7 and 8 where the dependent variable is the number of protests per cells in a given cell X year (in logs) obtained by aggregating into a single event all those in the same category (e.g., demonstrations, riots etc.) happening in the same location in the same day. See also notes to Table 2.

40

Table A.5 Afrobarometer Country-Rounds Availability

Afrobarometer Round 3

Afrobarometer Round 4

Afrobarometer Round 5

Benin

1190 [37]

1184 [33]

592 [33]

Botswana

1182 [53]

920 [42]

880 [41]

Burkina-Faso

-

968 [53]

576 [40]

Burundi

-

-

1200 [15]

Cameroon

-

-

656 [55]

765 [5]

656 [10]

719 [7]

Ghana

1165 [71]

960 [60]

1376 [66]

Guinea

-

-

1136 [42]

Ivory-Coast

-

-

1136 [57]

Kenya

1246 [45]

960 [34]

2135 [31]

Lesotho

1161 [10]

1192 [9]

1197 [9]

Liberia

-

797 [28]

873 [25]

Madagascar

1333 [191]

1152 [183]

1012 [216]

Malawi

1199 [34]

1152 [23]

1523 [40]

Mali

1187 [101]

960 [115]

986 [94]

Mozambique

1198 [111]

1088 [85]

1936 [99]

Namibia

1139 [82]

1024 [49]

1097 [52]

Nigeria

2200 [193]

1781 [197]

1936 [182]

Senegal

1200 [47]

1030 [25]

1176 [34]

Sierra-Leone

-

-

550 [28]

South-Africa

2171 [212]

2220 [188]

1400 [130]

Swaziland

-

-

456 [7]

Tanzania

1203 [102]

1024 [68]

2144 [94]

-

-

368 [14]

Uganda

2400 [60]

2431 [46]

1444 [57]

Zambia

1200 [103]

1200 [68]

1176 [71]

914 [44]

1000 [42]

1888 [48]

Cape-Verde

Togo

Zimbabwe

Notes. The table reports the number of individuals by country in rounds 3 to 5 of Afrobarometer. In parenthesis is the number of cells identified for each country in each round, based on the geographic location of the village of residence of the respondent.

41

Table A.6 Descriptive statistics Afrobarometer Avg.

Std. Dev.

Min.

Max.

Individual Characteristics Age P rimary Education Gender Employed N o water N o medicine Christian M uslim Other religion M ultiple cells N umber cells P rotest M obile daily V oted F ree to vote F ree to join F ree to speak T rust el.comm T rust president T rust parliament Happy with president Happy with mp Happy with council

36.66 0.61 0.5 0.34 0.38 0.42 0.64 0.2 0.17 0.39 2.06 0.12 0.58 0.73 0.75 0.66 0.54 0.56 0.64 0.58 0.70 0.55 0.56

14.65 0.49 0.5 0.47 0.48 0.49 0.48 0.4 0.37 0.49 3.97 0.32 0.49 0.45 0.44 0.47 0.5 0.5 0.48 0.49 0.46 0.5 0.5

18 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0

130 1 1 1 1 1 1 1 1 1 93 1 1 1 1 1 1 1 1 1 1 1 1

Cell Characteristics P opulation (1000s) M obile P hone 2G Coverage (percent) P rotests per 100, 000 pop. − GDELT P rotests per 100, 000 pop. − ACLED P rotests per 100, 000 pop. − SCAD Country GDP growth (percent) Border Distance (100 km) Capital Distance (100 km) Coast(dummy) P rimary Roads (100 km) P rimary Roads P aved (100 km) P rimary Roads Good Conditions (100 km) Secondary Roads (100 km) Electricity N etwork (100 km) T ravel T ime nearest city pop. ≥ 50K(hours) Inf ant M ortality Rate (h) M ountain (percent) F orest (percent) Irrigated (percent) Diamonds (dummy) M inerals (dummy) Oil (percent) T emperature (Celsius degrees) P recipitation (mm.) Distance f rom drought (100 km) F lashrate (per Km2 per year)

528.64 0.82 2.86 0.27 0.23 0.06 1.44 2.25 0.17 1.4 1.3 0.73 1.37 1.53 2.3 8.68 0.2 0.19 0.02 0.04 0.53 0.09 23.28 920.74 212.29 17.00

872.40 0.28 8.4 0.87 1.37 0.02 1.68 2.48 0.38 1.06 1.06 0.85 1.28 1.29 1.55 2.79 0.28 0.16 0.03 0.2 0.5 0.28 4.08 348.2 113.82 9.00

Notes. Upper panel weighted by weight. Lower panel weighted by population.

42

0 0 0 0 0 -0.18 0 0.05 0 0 0 0 0 0 0.38 1 0 0 0 0 0 0 5.26 74.58 0 0

7,841 1 540.41 217.63 34.31 0.21 10.54 19.11 1 4.66 4.66 3.8 5.98 7.36 35.53 16.7 1 0.99 0.32 1 1 1 30.8 3299.92 499.16 69.00

Table A.7 Afrobarometer. Mobile Phone Usage and Knowledge about the Economy

Coverage

Additional Controls Observations

Mobile Usage

Perceived vs Actual Growth

(1)

(2)

(3) Match

(4) Over

(5) Under

0.213** (0.098)

0.149* (0.085)

0.055* (0.032)

-0.011 (0.037)

-0.044 (0.043)

No

Yes

Yes

Yes

Yes

102,724

102,724

154,201

154,201

154,201

Notes. The dependent variable in columns (1) and (2) is Mobile Phone Usage, a categorical variable equal to 0 if the respondent never uses the mobile phone; 1 if she uses it less than daily; 2 if she uses it daily. Columns (3) to (5) compare perceptions about the state of the economy and its actual state. We construct the 10-years moving average and standard deviation of GDP growth for each country and define “very good” a year which is 2 s.d. above the average growth; “good” 1 s.d. above the average; “normal” between 1 s.d. below and 1 s.d. above the average; “bad” 1 s.d. below the average; “very bad” 2 s.d. below the average. We match this classification with the respondents’ answer to the question: “How do you rate the economic conditions in this country compared to twelve months ago?”.

43

Suggest Documents