(2025) Retombées économiques des conflits sociaux : preuves issues des médias sociaux et des images satellite

Economic Fallout of Social Conflict: Evidence from Social Media and Satellite Images Matteo Grazzi Paola Llamas Giulia Lotti Werner Peña WORKING PAPER N o IDB-WP- 1731 Inter-American Development Bank Departamento de Países de Centroamérica, México, Panamá, República Dominicana y Haití November 2025 Economic Fallout of Social Conflict: Evidence from Social Media and Satellite Images Matteo Grazzi Paola Llamas Giulia Lotti Werner Peña Inter-American Development Bank Departamento de Países de Centroamérica, México, Panamá, República Dominicana y Haití November 2025 Cataloging-in-Publication data provided by the Inter-American Development Bank Felipe Herrera Library Economic fallout of social conflict: evidence from social media and satellite images / Matteo Grazzi, Paola Llamas, Giulia Lotti, Werner Peña. p. cm. — (IDB Working Paper Series ; 1731) Includes bibliographical references. 1. Social conflict-Economic aspects-Haiti. 2. Social media-Haiti. 3. Remote-sensing images-Haiti. 4. Economic development -Haiti. I. Grazzi, Matteo. II. LLamas, Paola. III. Lotti, Giulia. IV. Peña, Werner. V. Inter-American Development Bank. Country Office in Haiti. VI. Series. IDB-WP-1731 http://www.iadb.org Copyright © 2025 Inter-American Development Bank ("IDB"). This work is subject to a Creative Commons license CC BY 3.0 IGO (https://creativecommons.org/licenses/by/3.0/igo/legalcode). The terms and conditions indicated in the URL link must be met and the respective recognition must be granted to the IDB. Further to section 8 of the above license, any mediation relating to disputes arising under such license shall be conducted in accordance with the WIPO Mediation Rules. Any dispute related to the use of the works of the IDB that cannot be settled amicably shall be submitted to arbitration pursuant to the United Nations Commission on International Trade Law (UNCITRAL) rules. The use of the IDB's name for any purpose other than for attribution, and the use of IDB's logo shall be subject to a separate written license agreement between the IDB and the user and is not authorized as part of this license. Note that the URL link includes terms and conditions that are an integral part of this license. The opinions expressed in this work are those of the authors and do not necessarily reflect the views of the Inter-American Development Bank, its Board of Directors, or the countries they represent. EconomicFalloutofSocialConflict: Evidencefrom SocialMediaandSatelliteImages ∗ Matteo Grazzi † Paola Llamas ‡ Giulia Lotti § Werner Peña ¶ November 17, 2025 Abstract In this paper, we leverage a quasi-experimental design and innovative sources of infor- mation to examine the impact of rising social conflict and political instability in Haiti. By exploiting geographical heterogeneity and leveraging data from Facebook and satellite imagery, we show the impact of different types of violence on proxies of economic activity in the context of countries with limited data availability. In the short term, we find that one additional violent event reduces economic activity by approximately 3.1% within the ten-day window following its occurrence. In the medium term, one additional political or civil event in an arrondissement is associated with a decline of approximately 1.5% and 2.5%, respectively, in economic activity over the subsequent five-month period. Impor- tantly, the Facebook data also allows for a disaggregation of the effects by sector, with the sectors most impacted by rising insecurity being home services and professional services. The long-term estimates indicate that an additional violent event is associated with a 1% decline in economic activity, as proxied by nighttime light intensity, one year following the event. These results show a sharp initial decline in economic activity, followed by smaller but lasting contractions, indicating limited recovery after violent events. Keywords:Haiti, social conflict, economic performance, social media, activity quantile, nighttime lights. JEL Codes:C80, O54, Q34, E32, R11. ∗We are grateful to Jeffrey Wooldridge, Christian Volpe, Santiago Perez-Vincent, Amrit Amirapu, Agustín Fil- ippo, Randolph Gilbert, Mounir Mahmalat, anonymous referees and members of the Rapid Crisis Impact Assess- ment for Haiti, for useful comments and suggestions. The partnership between Meta and the Inter-American Development Bank (through the Data Partnership) facilitated access to the data generated by the company in Haiti. We used OpenAI’s ChatGPT to assist with refining language during the preparation of the manuscript. All final content is the authors’ own. †Inter-American Development Bank (IDB). Email:matteog@iadb.org. ‡Northwestern University. Email:paola.llamas@kellogg.northwestern.edu. §Inter-American Development Bank (IDB). Email:glotti@iadb.org. ¶University of Kent. Email:wp62@kent.ac.uk. 1 Introduction Economic activity and social conflict are deeply interconnected, both exerting powerful in- fluences that shape the trajectory of societal development. In the interplay of both elements, social conflict is often perceived as a threat to economic stability. There is a large body of literature documenting the strong association between social conflict and poor economic outcomes. While some studies directly focus on the negative impact of social conflict on the growth rate of output (Alesina and Rodrik,1992;Rodrik,1999;Collier,1999;Abadie and Gardeazabal,2003;Fang et al.,2020;Le et al.,2022), others look at the relationship between social conflict and a wide set of economic outcomes, such as economic inequality (Esteban and Ray,2011;Genicot and Ray,2017), investment and human capital accumulation (Ben- habib and Rustichini,1996;León,2012;Ray and Esteban,2017), consumption, trade and fi- nancial markets (Barro and Ursua,2008;Amodio and Di Maio,2018;Guiso et al.,2009;Novta and Pugacheva,2021). The literature highlights several key channels through which social conflict disrupts pro- ductive processes. Social conflict discourages human capital accumulation by limiting ac- cess to education and training opportunities (Bodea and Elbadawi,2008;León,2012;Cook, 2014;Ray and Esteban,2017;Brück et al.,2019). It also increases firms’ operating costs, act- ing as a barrier to innovation and entrepreneurship (Amodio and Di Maio,2018;Prete et al., 2023;Couttenier et al.,2024). Similar patterns are found in Latin America, where violence has been shown to discourage investment, reduce productivity, and lead to business closures (Perez-Vincent et al.,2024). Spikes in social conflict reduce competitiveness and deters both national and foreign investment (Benhabib and Rustichini,1996;Knight et al.,1996;Rodrik, 1999;Novta and Pugacheva,2021). Labour demand often declines due to rising operational costs and firms ceasing operations, while labour supply may shrink as certain jobs or locations become too dangerous to sustain employment (Fernández et al.,2014;Ksoll et al.,2022;Maio and Sciabolazza,2023). Social conflict can also incentivise brain drain, as skilled individuals seek safety and better opportunities elsewhere (Docquier and Rapoport,2012). Finally, pub- lic resources are often diverted to managing social instability, crowding out investments that would otherwise go toward improving human capital or enhancing the productive capacity of the economy (Gupta et al.,2004;d’Agostino et al.,2016).¹ The multidimensional disruption caused bysocial conflict creates a complex environment ¹The literature on the relationship between crime and economic growth have identified similar channels by which economic performance is affected. However, crime also have specific impacts. For instance, individuals may reduce their participation in the labour market and turn to illegal activities if the marginal benefit of en- gaging in such activities exceeds the marginal cost (Becker,1968;Anderson,1999). Also, high crime rates are strongly linked to weak enforcement of property rights, which discourages innovation initiatives (Goulas and Zervoyianni,2015). 1 where economic progress becomes difficult to sustain, effectively increasing the risk of be- coming a conflict trap.²This is especially true for countries with high levels of institutional fragility, where social conflict exacerbates pre-existing vulnerabilities in terms of governance and public institutions (Besley and Persson,2011). In such contexts, the state’s ability to effec- tively manage resources, implement policies, and provide basic services is severely weakened, exacerbating the negative impact on economic growth. Therefore, understanding the specific dynamics between social conflict and economic activity is even more urgent in fragile states, where the risk of self-reinforcing cycles is heightened and targeted initiatives are essential to break this vicious circle. While some studies have conducted research on the relationship between social conflict and economic performance in fragile states (Fang et al.,2020;Diwakar,2015;Akresh et al., 2012;Rizvi,2022;Nkurunziza,2019;Ouedraogo,2024), a recurring challenge has been data availability. In fact, not only these countries have limited statistical collection and production capacity, but the outbreak of conflict can further disrupt data collection efforts. Therefore, researchers have increasingly turned to non-traditional data sources, which are less reliant on local statistical capacities and less affected by conflict-related disruptions. Since the influ- ential paper ofHenderson et al.(2012) and seminal contributions made byDoll et al.(2006), Sutton et al.(2007) andGhosh et al.(2009), the use of innovative sources of information, in particular satellite imagery, has become an invaluable tool that enables the analysis of the effects of social conflict on productive activities where data availability is scant (Haslam and Tanimoune,2016;Racek et al.,2024;Levin et al.,2018;Joseph,2022;Guo et al.,2023;Tähti- nen,2024). Yet, most studies exploring this relationship focus on national or highly aggregated subnational levels, often overlooking the heterogeneous impacts of social conflicts on more granular settings, and having more difficulties claiming causality. In this paper, we aim to offer a more comprehensive analysis of the heterogeneous and granular impacts of social conflict on economic activity in fragile states. We focus on Haiti, a paradigmatic case of a country deeply affected by violence and social unrest. By leverag- ing satellite imagery, social media data and exploiting geographical heterogeneity, we inves- tigate how social conflict causally influences economic outcomes across both regions and in- dustries. This innovative approach enables us to uncover spatial and sectoral variations in economic performance, offering a nuanced perspective on how the effects of social conflict differently shape the development trajectory of regions and industry-specific performance. ²A conflict trap can be defined as a self-reinforcing cycle where low levels of development and economic setbacks increase the likelihood of social conflicts, and, conversely, social conflict hinders development and economic recovery. In this trap, repeated cycles of conflict and economic damage make it progressively harder for a country to escape, as each phase of social unrest further erodes economic stability and raises the risk of future conflicts (Collier et al.,2003). 2 Social conflict has been long rooted in the history of Haiti since its foundation as an inde- pendentcountry(Girard,2005). Overthepastfewdecades, Haitihasfacedaseriesofcrises, in- cluding natural disasters, frequent changes in leadership, corruption and a weak institutional framework, all of which have contributed to a deepening sense of instability and deterioration of public order. Particularly, since 2018 Haiti has endured a new cycle of political instability and social conflict, marked by a series of violent events that have significantly undermined governance and exacerbated pre-existing social and economic challenges. The situation of violence reached a critical point with the assassination of President Moïse in July 2021 (Con- gressional Research Service,2023). Notably, this period has seen an increase and consolida- tion of criminal groups, in particular gang related violence. Power struggles between political actors increased political instability, an environment in which gangs increased their control over the capital. The surge in gang violence in Port-au-Prince has compelled numerous resi- dents to flee their homes and seek safety in other areas. Importantly, gang control is no longer confined to the capital, it has expanded (although with less intensity) into other regions (Ber- telsmann Stiftung,2024). While the Covid-19 pandemic played a role in the decline of production in 2020, insecurity appears to be the single most important factor influencing the poor economic performance in the last 7 years. Despite the aggregate evidence on the negative association between social conflict and Haiti’s economic performance (as suggested by recent escalating levels of vio- lence and a sustained decline in production), a deeper analysis is needed to understand the causal and heterogenous impact of the former on the later. Indeed, it is important to esti- mate the effect of insecurity on the economy in Haiti to show how violence and instability are blocking investment, closing businesses, and weakening economic growth. By measuring the economic impact, leaders can prioritize actions that not only improve safety but also create the conditions needed for jobs, education, and development.³ To shed light on this issue, we leverage a quasi-experimental design and innovative sources of information. The central hub of power in Haiti is Port-au-Prince, where escalat- ing social turmoil has been mirrored by increasing gang violence as various groups compete for control of the capital. Consequently, violence has surged in Port-au-Prince to a much greater extent than in other regions of the country. We exploit this geographical heterogeneity to compare the economic disruptions experienced in Port-au-Prince due to the spike in vio- lence with those in other regions where gang presence is more reduced. We leverage Facebook data and satellite imagery from the National Aeronautics and Space Administration (NASA) to ³An example of these mitigating efforts is the Rapid Crisis Impact Assessment for Haiti (RCIA) launched by the Government of Haiti in May 2024. The objectives of the RCIA were to evaluate the 2021-2024 crisis impact in key regions and sectors, develop a recovery framework and investment plan for FY2025-2026, and enhance coordination between the government and partners, supported by international institutions. 3 show the impact of different types of violence on economic activity in the context of countries with limited data availability. Specifically, we use Facebook’s Business Activity Trends (BAT) – aggregate and by industries– and NASA’s Black Marble night-time lights (NTL), both available at daily and monthly intervals, as proxies for economic activity.⁴⁵The BAT data are disag- gregated to the administrative level 2 (arrondissementsin Haiti, or districts), whereas the NTL data can be aggregated at thearrondissementandcommunelevels, the last one corresponding to administrative level 3. We complement this information with annual sub-national (com- munes) figures on crop and textile production.⁶This way, through a two-way fixed-effects model and using daily, monthly and yearly data for Haiti from the Armed Conflict Location & Event Data (ACLED)⁷(Raleigh et al.,2010), we can explore how increases in total violent events, political (violent) events , civil (violent) events , and related fatalities can affect eco- nomic activity in the short-, medium- and long-term. Our results indicate that 1 more violent event in a district, is associated with a short-run decrease of economic activity (measured by Facebook’s daily BAT) of 3.1 percent in the follow- ing week-time window. In the medium term, an extra political event decreases the production activity between 1.5 (Facebook’s BAT) and 6.2 (NTL) percent in the following five months. In the longer term, we observe a decrease of economic activity of approximately 1 percent in the following year-time window. These results point to a sharp initial decline in economic activity, followed by persistent though smaller contractions over longer horizons, suggesting limited recovery dynamics after violent events. Thus, in the medium-long term the persistency of so- cial conflict might leave long-lasting scars on production.⁸Importantly, the Facebook data also allows for a disaggregation of the effects by sector, with the most impacted sectors by ris- ing insecurity being homer services and professional services. The public good sector instead exhibited greater resilience and did not experience significant changes in economic activity driven by increase in political or civil events. Our paper is partially related toYousuf and Muller(2022). These authors look at the ef- fect of political violence on economic activity in Bangladesh by using ACLED’s database and NASA’s Black Marble night-time lights. Their results indicate that there is an immediate impact ⁴The BAT data covers the period from March 2020 to November 2022, while our NTL data spans the same time frame and extends further, covering January 2018 to December 2023. ⁵As we explain in Section3.2, we decided not to use daily NTL data due to their strong autoregressive compo- nent, which largely stems from NASA’s gap-filling procedure. This method fills missing values based on the most recent high-quality observation, introducing persistence that may distort temporal dynamics. This significantly limits attempts to establish a unidirectional causal link between violence and disruptions in economic activity, as the temporal sequence of events may be artificially reversed. ⁶This is proprietary data produced byGeoAdaptive(2024), and is based on sectoral production statistics, firms’ spatial location, and satellite data. ⁷Which can be aggregated at thearrondissementandcommunelevel. ⁸Indeed,Masri et al.(2024) highlight that social conflict can lead to persistent negative economic impacts. 4 of political violent protests on luminosity of -0.9 percent on daily night lights. The nationwide monthly impact is approximately 1.7 percent, which becomes evident within a 1-month time frame. While we also use ACLED’s database and NASA’s Black Marble imagery, there are im- portant differences. First, our objective is to establish a unidirectional causal link between violence and disruptions in economic activity, whereasYousuf and Muller(2022) does not elaborate on a design that could allow for strict causal inference. Second, we look at the case of Haiti, where the nature and reach of social unrest is more violent and more widespread than the case of Bangladesh. Third, we make a more detailed analysis of the different expressions of social conflict (civil and political events and related fatalities) and their differentiated impact on economic activity. Notably, by using Facebook’s BAT we are able to analyse the impacts of social conflict on industries. Our paper contributes to several strands of the literature. First, we add to the papers that have used nighttime lights for the specific case of Haiti. For instance,Mitnik et al.(2018) use communal-section and pixel level annual nighttime lights to approximate the impact of trans- port infrastructure investments on economic activity in Haiti, whereasJoseph(2022) uses an- nual nighttime lights to assess the differentiated subnational impact on economic activity of the 2010 earthquake. Owing to the need of using long annual time series, both papers com- bine harmonised nighttime light data coming from satellites with different resolution levels and saturation issues in brightly lit areas. Due to our time span, we rely exclusively on high- quality nighttime light data (NASA’s Black Marble) coming from the Visible and Infrared Imag- ing Radiometer Suite (VIIRS), whichGibson et al.(2021) demonstrate provides a more accu- rate approximation of economic activity at finer spatial resolutions. Second, our paper contributes to the growing literature on the causal impact of social conflict on economic activity in fragile countries. By leveraging novel data sources and ex- ploiting regional heterogeneity in both the intensity of violence and its differential impact on economic activity, our paper allows a transparent discussion of causal attribution. Third, the use of daily, monthly and annual data on social conflict and variables highly correlated with economic activity allows us to evaluate the short-, medium- and long-term negative impacts of social conflict on productive activities. Fourth, our paper also highlights the usefulness of social media data to measure economic performance. While satellite data has been widely used to measure the economic impact of social turmoil, to the best of our knowledge, our paper is the first to leverage the use of Facebook’s BAT for this purpose.⁹This points out the ⁹Despite its relative recent release, there are studies that have taken advantage of Facebook’s business infor- mation.Eyre et al.(2020b), which constitutes the seed of the BAT, use Facebook data to asses the recovery of small businesses after natural hazard events in Nepal, Puerto Rico and Mexico. Whereas,Díaz and Henríquez(2024) use the BAT data to examine how the economic activity of small businesses influenced mental health outcomes across five Latin American countries during the initial phase of the Covid-19 pandemic. 5 usefulness of social media data to measure economic performance. Fifth, Facebook’s BAT data enable us to examine the impacts of social conflict on various industries across different regions. This level of detail sheds light on the differential effects of social unrest, providing valuable information for policymakers and stakeholders aiming to support sectoral resilience in conflict-prone regions. This type of analysis is almost non-existent in studies focusing on fragile countries, even when satellite data is used. The paper proceeds as follows. In Section2we describe the ACLED database and the dif- ferent types of events it measures. Next, in Section3we describe in detail our two sources for approximating economic activity: Facebook’s Business Activity Trends from Meta datasets and NASA’s Black Marble night-time lights. Section4presents our empirical strategy and in Section5the results. Section6contains robustness checks applied to our baseline results and extensions of our analysis, while the last section concludes. 2 MeasuringSocialConflictandPoliticalInstabilityinHaiti: ACLEDDatabase To quantify the various manifestations of social conflict in Haiti, we leverage the geographic and temporal granularity of the Armed Conflict Location & Event Data Project (ACLED) database. ACLED is a detailed data repository tracking political violence, demonstrations, and conflict events globally. By drawing from a variety of sources, including local and interna- tional news outlets, reports from non-governmental organizations, and international bodies, ACLEDoffersalmostreal-timeinsightsintovariousspheresofsocialandpoliticalviolenceand associated events, detailing their nature, participating actors, geographical location, dates, and other relevant attributes. Its emphasis on granular, location-specific data allows users to explore trends and patterns in violence and political activity at subnational levels. To get a detailed perspective on the surge in social conflict that has been impacting Haiti since mid- 2018, we take advantage of this last feature and obtain daily, monthly and annual subnational data (at the level ofarrondissementandcommune¹⁰) on total violent events, political (violent) events, civil (violent) events and related fatalities (see Table1for definitions). We follow the classification used by the United Nations Office for the Coordination of Humanitarian Affairs to distinguish between political and civil events, as well as the associated fatalities. ¹⁰Haiti is divided into 10 departments, each of which is further subdivided into severalarrondissements, giv- ing a total of 42arrondissements. Anarrondissementtypically comprises multiplecommunes(totaling 146 com- munes), which in turn are divided into communal sections. 6 Table 1: ACLED’s Definitions of Violent Events and Related Fatalities Category Description Total Events A distinct incident reported to have occurred at a specific time and location, involving either the use of force by one or more actors, a demonstration, or a strategic political development. There are six types of events: battles, protests, riots, explosions/ remote violence, violence against civilians, strategic developments. Political Events Political events are single altercations where force is used by one or more groups to- ward a political end. These include ACLED’s battles, violence against civilians, and explosions/remote violence event types, as well as the mob violence sub-event type of the riots event type. Civil Events Civil events involve civilians as the main actor or target of an altercation. Accord- ing to ACLED’s codebook, civilians, being unarmed by definition, lack the capacity to participate in acts of political violence. These incidents are asymmetrical, with the perpetrator being the sole party employing force. Civilian targeting events include vi- olence against civilians and explosions/remote violence where civilians were directly targeted. Total FatalitiesFatalities occurring as a consequence of any of the six events captured by total violent events. Political Fatalities Fatalities occurring as a consequence of a political event. Civil FatalitiesFatalities occurring as a consequence of a civil event. Counts of “civilian fatalities” exclude civilians unintentionally killed during combat between armed groups or as a by-product of actions targeting those groups remotely, such as airstrikes on militant positions. Notes: Strategic developments are defined as events that provide contextual insights into actions and develop- ments involving groups that, while not classified as political violence or demonstrations, may influence future unrest or shape broader political trajectories within or between countries. Note that total events is not simply the sum of political and civil events, because some events fall into both categories and thus overlap.Source:Armed Conflict Location & Event Data (ACLED) Codebook. Table2shows different moments of the monthly distribution of these six violence cate- gories in Haiti.The distributions exhibit a right-skew, indicating the presence of relatively low counts in somearrondissementsin comparison with fewarrondissementswhere violent inci- dents are more widespread. An interesting finding is that, on average, political events tend to be more deadly than total and civil events. Specifically, during the period 2018-2023, po- litical events resulted in an average of 1.68 fatalities per event, compared to 1.21 fatalities per civil event and 0.94 fatalities per total number of events.¹¹Figure1illustrates that over the observed period, all six categories of violence progressively increased following the assassina- tion of President Moïse in July 2021. This escalation notably led to a peak in political violence ¹¹This pattern also appears when considering the share of events with at least one fatality: 58% of political events report at least one fatality, compared with 46% of civil events. 7 around March 2023, marked by 110 political events that resulted in 590 fatalities. Table 2: Summary Statistics of Monthly Events and Fatalities in Haiti Variable Obs Mean Std. Dev. Min Max Total Events3,024 2.13 9.14 0 127 Political Events 3,024 1.14 5.69 0 101 Civil Events3,024 0.62 3.23 0 66 Total Fatalities3,024 2.00 14.22 0 386 Political Fatalities 3,024 1.92 14.10 0 386 Civil Fatalities 3,024 0.75 5.46 0 112 Notes: This table presents basic descriptive statistics for total events, political events, civil events and related fatalities for the period January 2018 -– December 2023 at the level ofarrondissement. ACLED data were down- loaded on 14 November 2024.Source:Raleigh et al.(2010), authors’ own calculations. Figure 1: Time Series Evolution of Monthly Events and Fatalities in Haiti Notes: This figure presents the time series evolution of total events, political events, civil events and related fatal- ities for the period January 2018 –- December 2023. ACLED data were downloaded on 14 November 2024.Source: Raleigh et al.(2010), authors’ own calculations. As suggested by Table2, this surge in violence is unevenly distributed acrossarrondisse- ments(see Figure2). The six ACLED’s categories of social conflict show a higher incidence in twoarrondissements: Port-au-Prince and Croix-des-Bouquets (both located in the depart- ment of Ouest). Particularly, Port-au-Prince, the central hub of power in Haiti, has experi- enced a sharp escalation in gang violence over the last five years, reflecting the increasing so- cial turmoil. As various groups vie for control of the capital, violence has intensified there far more significantly than in other parts of the country. Our identification strategy aims to take 8 advantage of this geographical heterogeneity to identify the causal impact of social conflict on economic activity. Figure 2: Heat Map of the Spatial Distribution of Monthly Events and Fatalities in Haiti (a) Total events (b) Political events (c) Civil events (d) Total fatalities (e) Political fatalities (f) Civil fatalities Notes: This figure presents heat maps with the spatial distribution (arrondissement) of total events, political events, civil events and related fatalities for the period January 2018 –- December 2023. The label ranges represent quintiles of the corresponding variable, calculated excluding zero values. Areas with a value of zero are left blank. ACLED data were downloaded on 14 November 2024.Source:Raleigh et al.(2010), authors’ own calculations. 3 Innovative Data to Measure Business Activity and Eco- nomicPerformance To analyse the short-, medium-, and long-term economic impacts of social conflict, we re- quire to integrate ACLED’s detailed spatial and temporal conflict data with subnational figures 9 on economic activity.¹²This, however, poses significant challenges. First, Haiti’s highest fre- quencyindicatorofeconomicactivity–IndicateurConjonctureld’ActivitéEconomique(ICAE)– is only available on a quarterly basis, thus, limiting short- and medium- term analyses. More- over, given the data-collection challenges, theInstitut Haïtien de Statistique et d’Informatique (IHSI) does not produce an indicator measuring economic activity at the subnational level. These constraints would prevent us from leveraging the spatial heterogeneity in violence and economic activity to identify the causal impacts of the former on the latter. While the lack of high frequency subnational data is often a limitation in fragile countries such as Haiti, with the advent of groundbreaking sources of information this is no longer a binding constraint. In- deed, since the influential paper ofHenderson et al.(2012) and seminal contributions made byDoll et al.(2006),Sutton et al.(2007) andGhosh et al.(2009), the use of innovative sources of information, in particular satellite imagery, has become an invaluable tool that enables the quantitative analysis of economic issues where data availability is scant. In this paper, we take advantage of these new sources of information and use data highly correlated with economic activity to approximate the economic performance at the subna- tional level. For the short and medium term analyses, we leverage data from META-Facebook and satellite imagery from the National Aeronauticsand Space Administration (NASA). Specif- ically, we use Facebook’s BAT –aggregate and by industries– and NASA’s Black Marble night- time lights, both available at daily and monthly intervals and disaggregated to the administra- tive level 2 and level 3 (arrondissements–districts– andcommunes), as proxies for economic activity. For the long term analysis, we obtaincommune-level annual values of the Black Mar- ble NTL and complement this information withcommune-level yearly indicators on real agri- cultural and textile production. The latter two indicators are proprietary data sourced from GeoAdaptive(2024). In the following paragraphs we describe each dataset in more detail.¹³ 3.1 Facebook’s Business Activity Trends (BAT) Facebook’s BAT is a dataset based on business social-media activity that intends to measure business activity after the occurrence of exogenous shocks, such as natural disasters or pan- demics. This database was developed within Data for Good at Meta and is based on the work of Eyreet al.(2020a), which aims to nowcast business recovery following emergencies byutilising online posting activity as a key indicator. The authors’ main assumption is that a sufficiently strong external shock can influence the aggregate posting behaviour of Facebook business pages, which, in turn, can serve as a proxy for business performance during disruptive events. ¹²The short-term impacts are measured using the daily data, the medium-term impacts are defined by the monthly data, and the long-term impacts are measured using the yearly data. ¹³Additional satellite-based data used in extensions and robustness checks are described in AppendixA. 10 Eyre et al.(2020a) compare their methodology with other measurements of economic activ- ity based on business surveys, mobile phone information and time series of satellite imagery, concluding that their methodology renders reasonable similar estimates of the recovery pe- riod after the occurrence of natural disasters. Lam et al.(2022) generally adopt this methodology to produce Facebook’s BAT. The ag- gregate BAT is produced at the subnational level (level 2 of the Global Administrative Areas- GADM) and by industries (called business verticals¹⁴). The main metric of the BAT is what the authors call “activity quantile”. This metric is the result of comparing the business daily post count with the daily posting frequency during the baseline period, where the baseline period is 90 days before any specific date. When its value is around 0.5, it signals a normal level of activity or, as the authors call it, the “pre-crisis-like behaviour”. Thus, economic disruptions cause the activity quantile to deviate from the central value of 0.5: values below 0.5 indicate economic distress, while values above 0.5 signify economic expansion. Notice thatLam et al. (2022) adopt a fixed-cohort approach, where the sample of Facebook pages is chosen at a spe- cific date (for instance, the shock date) and remains the same in the post-crisis period. There- fore, regional full recovery effectively means that the full sample of business pages return to their “normal” posting activity.¹⁵ In this paper, we leverage daily and monthly BAT data produced by Facebook in the con- text of the Covid-19 pandemic, which covers the period from March 2020 to November 2022. However, in our empirical analysis (Section5) we restrict our sample to the period from July 2020 to November 2022 to avoid conflating the effects of social distancing measures imple- mented by public and private entities with the adverse impacts of violent events.¹⁶The qual- ity filters applied by Facebook mean that we have good-quality data for 22arrondissements out of 42, including the country’s capital.¹⁷The descriptive statistics of the activity quantile by business vertical are shown in Table3. Apart from the category “All” –which includes all industries–, the business verticals with more weight on our sample are “Public Good”, “Profes- sional Services” and “Business & Utility Services”. On the other hand, “Grocery & Convenience ¹⁴The authors call business verticals to their grouping of businesses into different industries based on the page admin self-reported business type. AppendixBprovides more details on this dataset, the list of business verticals and their corresponding description. ¹⁵Importantly, as pointed by the authors, the real-time nature of the activity quantile makes the adoption of a dynamic-cohort approach unfeasible. In a dynamic approach, the sample of business pages varies representing firms exiting and entering the markets. However,Lam et al.(2022) argue that in the short run it is not possible to determine whether a business that has stopped posting does so because it has exited the market or it is just a pause as a consequence of the disruption of an external shock. ¹⁶We select July 2020 as the starting point of our sample because the Oxford Stringency Index (OSI)—which measures the intensity of social distancing policies during the pandemic—shows a sharp decline in that month, indicating the relaxation of such measures. ¹⁷The list of these 22arrondissementsis provided in AppendixB. Notably, they accumulate 80.4% of Haiti’s population in 2020. 11 Stores”, “Lifestyle Services” and “Manufacturing” only report 33 observations each. The dis- crepancy in the number of observations arises from the exclusion of data points associated with fewer than 10 business pages, in accordance with privacy protection protocols.¹⁸Inter- estingly, over the period the average activity quantile of “Public Good” is slightly above 0.50, indicating that the spike in social conflict in Haiti has not disrupted the normal activity of this sector. This is not the case for “Travel”, “Retail” and “Home Services”, which are among the sec- tors (with a reasonable number of observations) that, on average, have deviated (downwards) more from the normal posting behaviour over the period. Table 3: Summary Statistics of Activity Quantile by Business Vertical Business Vertical Observations Mean Std. Dev. Min Max All 726 .43 .15 .011 .85 Business & Utility Services 231 .46 .11 .12 .81 Grocery & Convenience Stores 33 .36 .11 .13 .63 Home Services 198 .41 .13 .11 .75 Lifestyle Services 33 .35 .17 .06 .64 Local Events 66 .30 .12 .10 .58 Manufacturing 33 .49 .12 .30 .80 Professional Services 264 .40 .13 .10 .76 Public Good 297 .53 .15 .20 .91 Restaurants 132 .46 .15 .08 .86 Retail 165 .40 .18 .08 .95 Travel 165 .36 .13 .07 .67 Notes: This table presents the activity quantile by business vertical for the period March 2020 – November 2022. The table includes 22arrondissementsfor which BAT data is available.Source:Facebook’s BAT, authors’ own calculations. The time series analysis (Figure3) shows that the aggregate activity quantile has progres- sively deviated downwards from the value of 0.5, coinciding with the deteriorating social and politicalenvironment(seeFigure1), reachingitslowestvalueinthelastquarterof2022. Figure 4reveals that the fallout is not limited to the political and economic capital, rather it has dis- rupted the economic activity in otherarrondissements, as shown by the progressively lighter blue shading over time. To check how representative the BAT data is, we compare Facebook’s network coverage with population counts across arrondissements and find a strong match, far from significant subnational biases (see AppendixBfor visual depiction). While Face- book’s BAT provide valuable insights into online business activity and industry-specific dy- namics, they capture only few dimensions of economic performance (marketing and sales). To complement this, we incorporate NASA’s Black Marble night-time lights data, which offer a broader, geospatial perspective on economic activity by measuring light emissions as a proxy ¹⁸Due to the small number of observations, we exclude these business verticals from our industry level anal- ysis. 12 for infrastructure use and energy consumption. This combination allows us to analyse eco- nomic trends from both digital and physical lenses, enhancing the robustness of our findings. Figure 3: Time Series Evolution of Activity Quantile (Business Vertical “All”) by month-year Notes: This figure presents the time series evolution of the average activity quantile (business vertical “All”) for the period March 2020 – November 2022. The figure includes onlyarrondissementsfor which data is available. Source:Facebook’s BAT, authors’ own calculations. Figure 4: Heat Map of the Spatial Distribution of the Activity Quantile (Business Vertical “All”) byArrondissement (a) 2020 (b) 2021 (c) 2022 Notes: This figure presents heat maps with the geographical distribution (arrondissement) of the annual average activity quantile (business vertical “All”) for the period March 2020 –- November 2022.arrondisse- mentswith unavailable data are left blank.Source:Facebook’s BAT, authors’ own calculations. 3.2 NASA’s Black Marble night-time lights Since the influential paper ofHenderson et al.(2012) and seminal contributions made byDoll et al.(2006),Sutton et al.(2007) andGhosh et al.(2009), NTL have become a well-established proxy for subnational economic activity, particularly in contexts where conventional subna- tional economic figures are not available. The economic rationale of using NTL as a proxy of 13 economic activity is that they are strongly correlated with infrastructure, urbanization, and energy consumption. Importantly, electricity, which is one of the main producers of artificial lighting, is an economic “normal good”, where its consumption increases as the available in- come rises. Geographically speaking, this means that as regions develop and residential and commercial infrastructure spreads, we would expect an increase in the production of artificial light and radiance. Thus, NTL can signal varying regional levels of economic development. In scenarios where data collection and production is not feasible and information and commu- nication technologies have not penetrated, NTL provides a proxy for economic activity with wide coverage over time and across geographies. In this paper we make use of NASA’s Black Marble nighttime lights (BM-NTL). These ra- diance data are based on the Visible Infrared Imaging Radiometer Suite (VIIRS) of the Suomi National Polar-orbiting Partnership (SNPP) satellite.¹⁹NASA pre-process and provides radi- ance information that is cloud-free, atmospheric, terrain, vegetation, snow, lunar, and stray light-corrected DNB radiances.²⁰This product, which was released in 2018, adds to the two well-known sources of NTL: 1) the DMSP-OLS nighttime lights, which is a low resolution (1km x 1km) radiance data that covers the period 1992-2013; 2) the high resolution (500m x 500m) radiance data based on the VIIRS which is provided by the Colorado School of Mines (CSM) (available from 2012 onwards in their monthly and annual versions). Importantly, NASA’s Black Marble offers several advantages in comparison with these two sources. First, it offers a higher resolution than the DMSP-OLS NTL and avoids the well-known problem of top-coding (capping the maximum values of radiance or brightness that can be recorded). In addition, VIIRS NTL includes a built-in calibration to guarantee the comparability of data across both time and space. Second, it deals better with distortions related to snowfall and seasonal vege- tation than the CSM’s radiance data, offering a higher radiometer calibration (Iddawela,2023). Furthermore, NASA’s BM-NTL data is constructed based on specialised algorithms to remove stray light, cloud cover, and ephemeral lighting (e.g., wildfires, gas flares). The NASA’s BM-NTL comes in three main products (VNP46 products): VNP46A2, daily ¹⁹The Suomi NPP crosses the equator at approximately 13:30 PM (ascending node) and 1:30 AM (descend- ing node). While capturing radiance at 1:30 AM might reduce the chance of identifying changes in economic activity in less populated/urbanised areas, it has the advantage of minimizing the risk of capturing non-human generated radiance and human activity tend to stabilise which facilitates across-time comparisons (Cao et al., 2022) ²⁰Each of these issues can potentially decrease the quality of the NTL. Clouds, for instance, can make it dif- ficult to detect the human-generated radiance on the Earth’s surface. The atmosphere can capture and absorb light no-generated by human activity. The terrain conditions and the sharp angles this can generate might affect the amount of radiance detected by satellites. Dense vegetation, as clouds, can obstruct the emission of light gen- erated by human activity. Snow, by reflecting moonlight, can make some areas to appear brighter than others, leading to an over-estimation of radiance. Something similar happens with moonlight, depending on the moon’s phases. Finally, sunlight can reflect on the Earth’s surface and this can be captured by satellites, contaminating the artificial light generated by human activity. 14 moonlight and atmosphere corrected NTL; VNP46A3, monthly composites generated from daily atmospherically- and lunar-BRDF-corrected NTL radiance; and VNP46A4, yearly com- posites generated from daily atmospherically- and lunar-BRDF-corrected NTL radiance (see AppendixAfor more details).²¹ Despite the availability of the VNP46A2 daily NTL product, we opted to exclude it from this study. The most basic version of this product (DNB BRDF-Corrected NTL) contains a substan- tial number of zeros and missing values at thearrondissement-day andcommune-day level, rendering the series unsuitable for our purposes.²²To address this limitation, NASA provides an alternative product: the gap-filled daily series of DNB BRDF-corrected nighttime lights, which imputes missing observations to ensure temporal continuity in the data. NASA’s gap filling procedure for this product uses the latest high-quality retrieval available in the previous days (Román,2021).²³While this allows researchers to have workable daily NTL time series, it exponentially increases the auto-regressive nature of the time series. More important for our purposes, it poses the risk of artificially reversing the temporal sequence of events needed to identify possible causal impacts of violence on economic activity. The VNP46A3 product is based on the daily NTL data from VNP46A2. Specifically, all daily observations classified as clear-sky, high-quality data are first selected for inclusion in the con- struction of the monthly composite (this effectively means to remove observations affected by aurora, incorrect snow flag and cloud contamination). As a second step, boxplots metrics and inter-quantile ranges are used to identify and remove outliers. The monthly figures are cal- culated by obtaining the mean values of the observations left after applying the two previous steps. Finally, the monthly radiances with values smaller than 0.5������/������ 2 /&#3627408480;&#3627408479;are reclassified as zero (Wang et al.,2022). In the case of the VNP46A3 product, the gap filling procedure is based on historical data and not on the latest (day-specific) high-quality retrieval (Román, 2021). This procedure is arguably more neutral with respect to the timing of violence events, however, it still poses the risk of obscuring the temporal sequence of events. In this line,Wang et al.(2022) advice against the use of the VNP46A3/A4 composites marked with “gap-filled” quality flags for purposes of quantitative analysis or change detection. For the monthly series (VNP46A3), we use the all-angle composite snow free (without gap- ²¹The radiance units of measure of these products is Watts per Square Meter per Steradian (������/������ 2 /&#3627408480;&#3627408479;), which measures the portion of a sphere covered by the light being observed. ²²For instance, in the case of Haiti between 2018 and 2023 around 30% of the total number of observations in the pairs arrondissement-day are zero or missing values. ²³Importantly, the gap-filling procedure is applied at the cell level before aggregation to thecommuneand arrondissementlevel. This means it not only affectscommunesandarrondissementswith entirely missing and/or zero-valued cells on a given day, but also those with some missing and/or zero-valued cells in the non-gap-filled version. Therefore, when gap-filling is applied at the cell level and values are subsequently aggregated to the communeandarrondissementlevel, the resulting totals differ from those in the non-gap-filled version. 15 filled values), filtering out poor quality composites (where the number of observations used for the composite is less than or equal to three). We chose to use the all-angle composite band (combination of near-nadir and off-nadir angles), to make a reasonable balance between pixel resolution and full coverage of all possible sources of human-made radiance. While near- nadir (satellite’s nadir point – limited area close to the area below the satellite) observations provide a higher resolution and reduce atmospheric interference, off-nadir observations pro- vide an angle that is more suited to detect non-isotropic sources of radiance (radiance sources that radiate energy not uniformly across all angles), at the cost of having less resolution. We then aggregate the radiance values at the level ofcommuneandarrondissementusing the ad- ministrative boundaries of the Global Administrative Areas (GADM). After these manipula- tions, we obtain complete high quality, non-missing/non-zero and non-gap filled series for 45communesand 23arrondissements.²⁴²⁵Importantly, we conduct our baseline analysis (Section5) and robustness checks (Section6) at thecommune-level, thereby expanding the cross-sectional dimension to 45 units (instead of 23arrondissement), which provides suffi- cient variation for reasonable panel estimation. However, in our online appendix we include the results at thearrondissement-level. Table4shows that the average monthly radiance fluctuates around 40.97������/������ 2 /&#3627408480;&#3627408479;, with the maximum radiance registered in Port-au-Price. While the Covid pandemic negatively im- pacted the radiance levels during the first part of 2020, Figure5reveals a clear pattern over the whole period: as the recent episode of social conflict intensified, economic activity—proxied by NTL—declined. Notably, the downturn in productive activities after mid-2018 is consis- tently captured by both Facebook’s BAT data and NASA’s Black Marble NTL. ²⁴We applied an additional cleaning step to the nighttime lights data for November 2023, as the radiance values for that month far exceed the pre-crisis levels observed in early 2018, without any plausible economic justification for such a surge. Indeed, the series shows a sharp spike between October and November 2023, which is fully reversed by December. This makes the radiance level of November 2023 an outlier not detected by NASA’s algorithms. This surge is driven by a sharp drop in the number of zero-valued cells in November 2023 relative to the long-run trend, leading to abnormally high summed values at the arrondissement level . To address this issue, we retrieved pixel-level radiance data for Haiti for October and November 2023, and assumed that any pixel with a zero value in October would also have a zero value in November (this in addition to the zero-valued cells of November), this makes the number of zero values to return to its long-run trend. We then aggregate the data at the arrondissement level and compute the radiance growth rate between October and November, which we use to derive corrected values for November 2023. ²⁵The list of these 45communesand 23arrondissementsis provided in AppendixA. They represent approxi- mately 75% of Haiti’s population in 2020. 16 [... middle sections omitted for long document ...] Table OA.28: Robustness: Total Real Production (in logs) with control variables (1) (2) (3) (4) Political Events Civil Events Political Fatalities Civil Fatalities Events last year -0.001*** -0.001*** -0.001*** 0.001*** (0.000) (0.000) (0.000) (0.000) Land Surface Temperature 0.024 0.025 0.007 0.028 (0.024) (0.024) (0.024) (0.025) Precipitation 1.859 1.858 1.079 1.987 (1.149) (1.162) (1.236) (1.203) Reach of Health Services 1.380 1.356 1.305 0.915 (2.103) (2.110) (2.009) (2.094) Reach of Education Services-1.715 -1.697 -1.264 -1.153 (2.965) (2.985) (2.779) (3.036) Earthquake 2021 0.001 0.001 -0.001 0.004 (0.009) (0.009) (0.009) (0.009) Floods 2022 -0.027 -0.028 -0.025 -0.030 (0.022) (0.023) (0.023) (0.023) Observations 123 123 123 123 R-squared 0.999 0.999 0.999 0.999 Average Y 8.365 8.365 8.365 8.365 Av. X 15.39 8.862 23.85 9.252 Notes:This table reports coefficients from estimating equation (3) adding control variables (coefficients������ ������are not presented in this table). The dependent variable is the logarithm of real total production in thousands USD������ ������,������inarrondissement������and year&#3627408481;(period 2018 –2022). Column (1) shows the effect of one year lag of political events������ ������,������−������; column (2), civil events; column (3), political fatalities; and column (4), civil fatalities. For a description of the variablesland surface temperature,precipitation,reach of health servicesandreach of education servicessee AppendixA.Earthquake 2021is a binary variable indicating whether an area was affected by the earthquake in 2021, in the affected areas—Baradères, Miragoâne, L’Anse-à-Veau, Jérémie, Anse d’Hainault, Corail, Les Cayes, Aquin, Les Chardonnières, Les Côteaux, or Port- Salut.Floods 2022is a binary variable indicating whether an area was affected by flooding in 2022 in the affected regions—Nord, Nord-Est, Nippes, or Nord-Ouest. All specifications includearrondissementand year fixed effects. Standard errors, clustered at thearrondissement level, are in parentheses. ∗ ������ < 0.10, ∗∗ ������ < 0.05, ∗∗∗ ������ < 0.01. OA-28 Table OA.29: Robustness: Nighttime Lights (in logs) with control variables (1) (2) (3) (4) Political Events Civil Events Political Fatalities Civil Fatalities Events last year -0.001 -0.001 -0.002 -0.001 (0.001) (0.002) (0.002) (0.003) Land Surface Temperature 0.467 0.468 0.442 0.457 (0.699) (0.695) (0.705) (0.689) Precipitation 1.860 1.745 -0.502 0.058 (17.882) (17.913) (17.680) (18.328) Reach of Health Services -45.333 -46.813 -53.992 -53.997 (46.483) (46.788) (48.215) (48.411) Reach of Education Services 42.679 44.211 52.486 52.252 (49.553) (49.999) (52.318) (52.117) Earthquake 2021 0.095 0.099 0.103 0.108 (0.282) (0.283) (0.269) (0.271) Floods 2022 0.125 0.123 0.147 0.144 (0.141) (0.139) (0.142) (0.146) Observations 69 69 69 69 R-squared 0.974 0.974 0.974 0.974 Average Y 2.142 2.142 2.142 2.142 Av. X 25.93 14.97 39.43 15.20 Notes:This table reports coefficients from estimating equation (3) adding control variables (coefficients������ ������are not presented in this table). The dependent variable is the logarithm of nighttime lights������ ������,������inarrondissement������and year&#3627408481;(period 2018 –2023). Column (1) shows the effect of one year lag of political events������ ������,������−������; column (2), civil events; column (3), political fatalities; and column (4), civil fatalities. For a description of the variablesland surface temperature,precipitation,reach of health servicesandreach of education servicessee AppendixA. Earthquake 2021is a binary variable indicating whether an area was affected by the earthquake in 2021, in the affected areas—Baradères, Miragoâne, L’Anse-à-Veau, Jérémie, Anse d’Hainault, Corail, Les Cayes, Aquin, Les Chardonnières, Les Côteaux, or Port-Salut.Floods 2022is a binary variable indicating whether an area was affected by flooding in 2022 in the affected regions—Nord, Nord-Est, Nippes, or Nord-Ouest. All specifications includearrondissementand year fixed effects. Standard errors, clustered at thearrondissementlevel, are in parentheses. ∗ ������ < 0.10, ∗∗ ������ < 0.05, ∗∗∗ ������ < 0.01. OA-29 Table OA.30: Robustness: SE clustered at department level in violence type specification Violence type Political Events Civil Events Political Fatalities Civil Fatalities Panel A: Real Agricultural Production (effect sizes in standard deviations) Events last year-0.001 -0.001 0.000 -0.001 (0.001) (0.001) (0.000) (0.001) Observations 168 168 168 168 R-squared 0.993 0.993 0.993 0.993 Average Y -0.0277 -0.0277 -0.0277 -0.0277 Av. X 13.14 7.494 19.71 7.923 Panel B: Real Textile Production (effect sizes in standard deviations) Events last year -0.003** -0.000 -0.007*** -0.004*** (0.001) (0.002) (0.002) (0.000) Observations 168 168 168 168 R-squared 0.978 0.979 0.967 0.966 Average Y -0.0229 -0.0229 -0.0229 -0.0229 Av. X 13.14 7.494 19.71 7.923 Panel C: Total Real Production (in logs) Events last year -0.001*** -0.001** -0.002*** -0.001*** (0.000) (0.000) (0.000) (0.000) Observations 164 164 164 164 R-squared 0.997 0.997 0.997 0.997 Average Y 8.352 8.352 8.352 8.352 Av. X 13.46 7.677 20.20 8.116 Panel D: Nighttime Lights (in logs) Events last year -0.003*** -0.003 -0.003*** -0.006*** (0.001) (0.003) (0.000) (0.001) Observations 115 115 115 115 R-squared 0.946 0.945 0.948 0.949 Average Y 2.321 2.321 2.321 2.321 Av. X 26.75 14.70 45.15 17.50 Notes:This table presents the coefficient estimates of equation (3), separately for each type of event (coefficients ������ ������are not presented in this table). The dependent variables are real crop production in standard deviations (Panel A), real textile production in standard deviations (Panel B), the log of total real production (Panel C) and the log of nighttime lights (Panel D)������ ������,������inarrondissement������, and year&#3627408481;(period 2018 – 2023 for nighttime lights and 2018 –2022 for production variables). Column (1) shows the effect of one year lag of political events������ ������,������−������; column (2), civil events; column (3), political fatalities; and column (4), civil fatalities. All regressions include arrondissementand year fixed effects. Standard errors are clustered at the department level and are in parenthe- ses. ∗������ < 0.10, ∗∗������ < 0.05, ∗∗∗������ < 0.01. OA-30 Table OA.31: Robustness: Analytic weights in violence type specification Violence type Political Events Civil Events Political Fatalities Civil Fatalities Panel A: Real Agricultural Production (effect sizes in standard deviations) Events last year -0.001*** -0.002*** 0.001** -0.002*** (0.000) (0.000) (0.000) (0.000) Observations 88 88 88 88 R-squared 0.996 0.996 0.996 0.995 Average Y 0.362 0.362 0.362 0.362 Av. X 24.06 13.74 36.41 14.66 Panel B: Real Textile Production (effect sizes in standard deviations) Events last year -0.002*** 0.003*** -0.012*** -0.004*** (0.000) (0.000) (0.004) (0.000) Observations 88 88 88 88 R-squared 0.995 0.995 0.978 0.975 Average Y 0.168 0.168 0.168 0.168 Av. X 24.06 13.74 36.41 14.66 Panel C: Total Real Production (in logs) Events last year -0.001*** -0.001*** -0.002*** -0.002*** (0.000) (0.000) (0.000) (0.000) Observations 88 88 88 88 R-squared 0.999 0.999 0.999 0.999 Average Y 8.820 8.820 8.820 8.820 Av. X 24.06 13.74 36.41 14.66 Panel D: Nighttime Lights (in logs) Events last year -0.003*** -0.005*** -0.002*** -0.001 (0.000) (0.001) (0.000) (0.002) Observations 90 90 90 90 R-squared 0.982 0.980 0.984 0.983 Average Y 2.681 2.681 2.681 2.681 Av. X 33.70 18.60 57.07 22.16 Notes:This table presents the coefficient estimates of equation (3), separately for each type of event (coeffi- cients������ ������are not presented in this table). The dependent variables are real crop production in standard devia- tions (Panel A), real textile production in standard deviations (Panel B), the log of total real production (Panel C) and the log of nighttime lights (Panel D)������ ������,������inarrondissement������, and year&#3627408481;(period 2018 – 2023 for nighttime lights and 2018 –2022 for production variables). Column (1) shows the effect of one year lag of political events ������ ������,������−������; column (2), civil events; column (3), political fatalities; and column (4), civil fatalities. All regressions are weighted using analytic weights based on the population of each arrondissement. This assumes that observa- tions corresponding to larger populations are measured with greater precision (variance is inversely proportional to population size). All regressions includearrondissementand year fixed effects. Standard errors are clustered at thearrondissementlevel and are in parentheses. ∗������ < 0.10, ∗∗������ < 0.05, ∗∗∗������ < 0.01. OA-31 Table OA.32: Robustness: NTL - Wild cluster bootstrap test for individual and joint signifi- cance (1) (2) (3) (4) Political Events Civil Events Political Fatalities Civil Fatalities Events last year -0.003 -0.003 -0.003 -0.006 WB joint test:������-value 0.166 0.667 0.138 0.172 Observations 115 115 115 115 R-squared 0.492 0.487 0.516 0.521 Average Y 2.321 2.321 2.321 2.321 Av. X 26.75 14.70 45.15 17.50 Notes:This table reports coefficients from estimating equation (3) adding control variables (coefficients������ ������are not presented in this table). The dependent variable is the logarithm of nighttime lights������ ������,������inarrondissement������and year&#3627408481;(period 2018 –2023). Column (1) shows the effect of one year lag of political events������ ������,������−������; column (2), civil events; column (3), political fatalities; and column (4), civil fatalities. All specifications includearrondissementand year fixed effects. Standard errors, clustered at thearrondissementlevel, are in parentheses. ∗ ������ < 0.10, ∗∗ ������ < 0.05, ∗∗∗ ������ < 0.01. OA-32