Policy Research Working Paper 10222 Rebel Governance and Development The Persistent Effects of Guerrillas in El Salvador Antonella Bandiera Lelys Dinarte-Diaz Juan Miguel Jimenez Sandra V. Rozo Maria Micaela Sviatschi Development Economics Development Research Group November 2022 Policy Research Working Paper 10222 Abstract How does territorial control by non-state actors affect long- outcomes about 30 years after the guerrillas first controlled term development? This paper investigates the economic, them, relative to adjacent areas that were then controlled by social, and political consequences of temporary territo- the formal state. The results suggest that informal commu- rial control by guerrillas during the Salvadoran Civil War. nity institutions in guerrilla-controlled areas led to enduring During this period, the guerrillas displaced state authorities land fragmentation and disengagement with the govern- and promoted the creation of self-governing institutions ment. The paper argues that when non-state actors develop that were highly representative of local values and openly governance institutions as an alternative to the state, this distrusted the state and elites. Using a spatial regression can lead to negative development effects through lasting discontinuity design, the paper shows that areas once norms of distrust of out-groups. under guerrilla control have experienced worse economic This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at ldinartediaz@worldbank.org and srozo@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Rebel Governance and Development: The Persistent Effects of Guerrillas in El Salvador * Antonella Bandiera† Lelys Dinarte-Diaz‡ Juan Miguel Jimenez§ Sandra V. Rozo¶ Maria Micaela Sviatschi|| Keywords: Armed non-state actors, economic development, El Salvador JEL Classification: O10, N3 * We are very grateful for discussions with and comments from Kathleen Beegle, Eli Berman, Leah Boustan, Carine Clert, Joana Corzo, Janet Currie, Claudio Ferraz, Thomas Fujiwara, Guy Grossman, Allan Hsiao, Seema Jayachandran, Ilyana Kuziemko, Horacio Larreguy, Moritz Lenel, Bentley Macleod, Ricardo Marten, Christopher Mays Johnson, Ed- uardo Montero, Eduardo Morales, Diana Moreira, Melanie Morten, Nathan Nunn, Berk Ozler, Mauricio Romero, Jake Shapiro, Juan Vargas, Leonard Wantchekon, Austin Wright, and seminar participants at the 2021 RIDGE Virtual Forum, 2022 BREAD, CIDE-PEV, ITAM, NBER Political Economy, NBER National Security, Princeton University, Universidad de Los Andes, Universidad El Rosario, and University of Geneva. We also thank Carlos Schmidt-Padilla for generously providing access to some datasets. This work has been supported by the Research Support Budget at the World Bank. The authors have no conflicts of interest to report. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. All errors are due only to the authors. † ITAM. email: antonella.bandiera@itam.mx ‡ Development Research Group, The World Bank, CESifo. email: ldinartediaz@worldbank.org § University of British Columbia. email: juamiji@gmail.com ¶ Development Research Group, The World Bank, IZA, CEGA, CESR-USC. email: sandrarozo@worldbank.org || Princeton University, NBER, BREAD, CESifo. email: msviatschi@princeton.edu “Mayors, judges, security posts, everything disappears, (...), practically the whole state disappears, and the state was us.” (FMLN Military Commander, March 2022) I INTRODUCTION Civil wars are common and persistent. At least 100 countries have experienced episodes of in- ¨ ternal armed conflict since 1946 (Pettersson and Oberg, 2020). Seminal literature has documented the large negative effects of these wars on economic development (Blattman and Miguel, 2010), showing that such conflicts directly depress economic growth because violence depletes factors of production. Yet, war undermines economic growth indirectly as well. Armed non-state actors may also affect long-term development during civil conflicts by seizing territory and extracting resources, expropriating land, mobilizing local populations, and imposing their own governance and economic structures to shape regions under their control (e.g., Wood, 2008; Arjona, Kasfir and Mampilly, 2015; Arjona, 2016; Stewart, 2018; Breslawski, 2021; Grasse, Sexton and Wright, 2021; anchez De La Sierra, 2020). Although territorial control is a key rebel strategy,1 little is known S´ about its development implications or whether effects persist once these groups relinquish control. We are the first to explore this mechanism. It is not obvious how territorial control by non-state armed actors affects long-term development. Rebels can promote security for residents, protect civilians from external armed actors, establish economic and judicial institutions, and provide basic public goods. Nonetheless, they may also displace private economic activity and induce norms of distrust of the state with negative devel- opment consequences that can last for decades. In this paper, we focus on the long-term development impacts of territorial control by the Farabundo on Nacional, FMLN) in El Sal- ı National Liberation Front (Frente Farabundo Marti para Liberaci´ Mart´ vador. The FMLN was an armed organization formed in October 1980 that united the five largest leftist guerrilla organizations in El Salvador.2 Starting in 1985, the FMLN (herein FMLN, rebels, or guerrillas) established full territorial, economic, social, and political control in multiple areas, effectively replacing the Salvadoran state. Territorial control by non-state actors is associated with several factors that could affect long-term 1 Territorial control is a key aspect of irregular warfare. For example, in Colombia, the FARC (Fuerzas Armadas Revolucionarias de Colombia) controlled many remote areas before the Peace Agreement was signed in 2016, much as Peru’s Shining Path (Sendero Luminoso) controlled the Andes Valley in the 1980s. 2 These included Fuerzas Populares de Liberaci´on Farabundo Mart´ ı, Ejercito Revolucionario del Pueblo, Resistencia Nacional, Partido Comunista Salvadoreno ˜ , and Partido Revolucionario de los Trabajadores Centroamericanos. 2 development, such as the provision of basic public goods and the establishment of new institu- tions to gain support from local communities. The Salvadoran context suggests one particularly relevant feature: the creation of alternative governing authorities that relied on community initia- tives and promoted distrust of the state (Pearce, 1986; Binford, 1997; Wood, 2003, 2008). In areas controlled by the FMLN, guerrillas eliminated local state authorities and promoted community- based organizations to address health, education, and economic problems. These peasant-run participatory institutions were autonomous and ideologically founded on community values, al- truistic solidarity, and distrust of the state and elites. These organizations flourished as an al- ternative to state institutions, distancing local communities from national politicians and further eroding trust in the state. Moreover, guerrillas also substituted subsistence crops for large-scale agricultural production to foster economic self-sufficiency among peasants. In contrast, during the same period, nearby areas remained under state control without changes in governance.3 We examine the effects of the FMLN’s temporary territorial control by comparing areas around and near the boundaries of its zones between 1985 and 1992, as documented in the United Na- tions map used during peace talks between the Salvadoran government and the FMLN. Figure 1 illustrates these areas and boundaries. The Salvadoran government and the FMLN jointly ap- proved the map and submitted it to the United Nations during the peace talks from 1990 to 1992. We assess the effects of FMLN territorial control on development 20 years later by examining cur- rent changes in economic activity (proxied by night light luminosity), human capital (measured as years of education), and a wealth index that approximates average household living standards. Our empirical strategy employs a spatial regression discontinuity design and uses geospatial data on night light luminosity for 2013 and census tract data on education and wealth for 2007. More- over, to disentangle mechanisms, we combine information from multiple sources on measures of social capital, land markets, and trust. A review of the validity of the empirical design shows that all geospatial and economic variables observed prior to the Civil War vary smoothly around the boundaries of rebel-controlled areas. In particular, the use of geocoded data from multiple sources—including covariates that proxy state capacity, violence, demographics, agricultural production, and land concentration—confirm 3 The only counterinsurgency strategy promoted by the Salvadoran government was the CONARA (Commission for the Restoration of Areas) in 1983. This initiative was implemented in two departments (San Vicente and Usulut´ an) but did not target guerrilla territory. It was similar to the Strategic Hamlet Program in South Vietnam and aimed to halt the influence of communism. However, the plan failed to produce the expected results and was quickly discontinued. 3 there were no differences in these dimensions before guerrillas seized control. The only significant difference between the areas around the boundary is a small discontinuity (approximately 17 m) in altitude. This is consistent with qualitative evidence and findings from our interviews with former combatants that show the FMLN’s territorial boundaries were primarily defined by war strategies ˜ and thus independent of preexisting economic conditions (Castaneda, 2016).4 In particular, rebel territories included strategic locations that offered a topographic advantage against the enemy. Figure 1. Guerrilla-Controlled Areas Source: Castaneda ˜ (2016). Notes: This map shows the location of guerrilla-controlled areas. It was submitted to the United Nations for the Cha- pultepec Peace Accords and approved jointly by the Salvadoran Government and the FMLN as part of the cease-fire negotiation process from 1990 to 1992. Results reveal that FMLN control in the mid-1980s had large and persistent negative effects on development outcomes in the long run. After almost 20 years, areas inside FMLN-controlled ter- ritories had less night light luminosity, lower human capital, and worse wealth outcomes relative to lands just outside these areas. The magnitudes of the estimated effects are important. By 2013, areas once controlled by the guerrillas experienced nearly 18.6 percent lower night light luminos- ity than places never under rebel control (approximately 5.2 percent lower GDP than in areas the guerrillas did not control).5 We also show that individuals who live today in areas previously 4 In our analysis, we show that this difference in altitude is not correlated with any economic outcomes at baseline, which confirms that these locations were chosen solely for strategic advantage. 5 De Groot et al. (2021) estimate that, absent conflicts around the world between 1960 and 2007, global GDP would have risen by 15.7 percent. Hence, an effect of five percent is sizable. 4 controlled by guerrillas have a wealth index 0.121 standard deviations (sd) lower than individu- als living in nearby areas, as well as 0.28 fewer years of education. Moreover, we show that the effects on education are concentrated among individuals who had not finished their education by the time guerrillas gained territorial control; in addition, these effects are not driven by selective migration and they persist over time. What explains these enduring negative effects? The entire region has experienced the same formal institutions since the war ended, and guerrillas no longer govern any areas. We hypothesize, how- ever, that the informal norms developed through the participatory institutions promoted by the FMLN between 1985 and 1992—combined with its view that citizens should guarantee their needs independent of elites and the state (Pearce, 1986)—induced persistent changes in relationships with the state and formal economic structures. Participatory institutions can influence local liv- ing standards through effects on attitudes such as “political disenchantment” with government.6 Local cooperation could foster social capital (Bauer et al., 2016) but could also breed resentment towards the state and elites if it developed in order to avoid dependence on those out-groups. At the same time, it could reinforce the economy of subsistence implemented during the guerrilla period if citizens from these areas still distrusted the state and mentally associated investors and large landowners with the repressive agriculture elites of the 1950s. In line with these arguments, our quantitative results show that individuals living in areas once controlled by the FMLN trust the state less and are less likely to engage with politicians. Trust has a central role in the effective functioning of state institutions (Banfield, 1967; Almond and Verba, 2015; Coleman, 1990). On the one hand, less trust may affect the supply of public goods if politicians have less information about local populations. For instance, citizens may fail to com- municate their needs effectively if they do not believe politicians will respond; this disengage- ment consequently prevents the government from providing public goods effectively (Jablonski and Seim, 2022; Buntaine, Nielson and Skaggs, 2021). On the other hand, less trust in state institu- tions can decrease demand for public services since citizens do not view the government as legiti- mate or capable (Mishler and Rose, 2001; Alsan and Wanamaker, 2018; Lowes and Montero, 2021; Martinez-Bravo and Stegmann, 2022). Our results support this channel because we document that residents of areas that experienced guerrilla control today report lower access to/utilization of public services despite the fact that we also observe: (i) more public investment in infrastruc- 6 The term was first used by Moodie (2011) as “democratic disenchantment.” 5 ture in these same areas; (ii) a larger number of schools per capita and better road density; and (iii) no differences in state presence such as hospitals, state buildings, and police stations. Distrust of the state and elites could also explain the endurance of economic structures imple- mented during guerrilla control. Our results largely show that inside guerrilla-held territory, com- mercial plots that were historically owned by elites are smaller and have lower productivity today. Moreover, we find suggestive evidence that agro-based industry did not develop there. In partic- ular, we find only a small share of individuals working in services and industry in former guerrilla areas relative to the control group. These results provide evidence that the subsistence agriculture implemented during guerrilla control continues to this day, potentially affecting long-term eco- nomic outcomes. Yet, there is no evidence that rebel-controlled areas altered income equality. Our focus group results suggest that regardless of whether large agricultural landowners would want to invest in former guerrilla areas today, residents there would be less willing to let them because they distrust such people as a result of their pre-War experience with agricultural elites. We rule out alternative mechanisms. First, an increase in violence during and after the conflict does not explain our results. They hold when we exclude areas close to the rebel border, sug- gesting that violence at the border was not the main driver, and to using different bandwidths.7 Furthermore, there was no increase in deaths, battles, and victims from 1980 to 1992 in guerrilla- controlled areas relative to nearby areas outside rebel control. Second, these results do not seem to stem from selective migration of individuals from guerrilla-controlled areas. Third, the results are not driven by differences in land tenure in former FMLN areas relative to nearby areas. Fourth, it is unlikely that the effects emerge from forced child recruitment into guerrilla groups. In fact, qualitative evidence suggests the Salvadoran Army extensively recruited children by force but the guerrillas did not.8 Finally, we demonstrate that the results do not arise from differential patterns of public or private investment in the post-conflict period between treated and control areas. For example, we find no differences in the supply and quality of public education and road connec- tivity, and no changes in the number of businesses in former FMLN areas relative to nearby areas. This paper has multiple policy implications. First, it offers evidence that the creation of alterna- 7 We rule out other dynamics associated with a border, such as a higher incidence of landmines along the boundary, because El Salvador ran a very successful program that cleared all mines by 1994. 8 It is estimated that of 60,000 Salvadoran Army combatants, about 48,000 (or 80 percent) were under 18 years of age, while only 2,000 of the 9,000 FMLN members were under 18 (or 20 percent) (Courtney, 2010). Moreover, a survey of child soldiers by UNICEF at the end of the war showed that while 91.7 percent of FMLN recruits had joined voluntarily, close to 53 percent of underage Salvadoran Army soldiers were forcibly recruited (Courtney, 2010). 6 tive self-governing institutions by non-state actors can hinder development by changing social norms. In particular, we show that distrust of the state can lead to disengagement with gov- ernance and associated elites, affect access to and utilization of public services, and reinforce a subsistence economy. Second, the results provide some evidence of a mismatch between histori- cal social norms and the best action given the current environment (Nunn, 2022). While distrust of the state may have been optimal behavior during the period of guerrilla control in these areas (given historic repression by the state and elites), such distrust afterward may have been subop- timal because the state had changed and those elites were gone (Boyce, 1995). Finally, our results provide the first empirical evidence on the consequences of living at the margin of the state. Scott (2010) highlights how hill societies in Southeast Asia prefer to live at the margin of the state to avoid slavery and taxes. In our case, we provide some evidence that “opting out” from the state could have negative effects on development. Overall, this paper demonstrates that historic territorial control by non-state actors and their es- tablishment of local governance can partly explain long-term development paths. Particularly in Latin America, local governance by non-state actors has featured prominently in several com- munities since colonization. Our findings are consistent with seminal studies that show the role of historical institutions in long-term development (e.g., Acemoglu, Johnson and Robinson, 2001; Acemoglu and Robinson, 2012; Dell, 2010; Dell, Lane and Querubin, 2018; Acemoglu et al., 2019; Nunn, 2020; Dell and Olken, 2020; Lowes and Montero, 2021). This paper not only documents persistence but also sheds light on specific mechanisms. Our evidence closely relates to Dell, Lane and Querubin (2018), who show how village governance in Vietnam increased social capital and development by crowding-in cooperation with government. We complement this work by sub- stantiating how local rebel governance (developed in parallel and as an alternative to the state) can hinder long-term development by reducing economic production and cooperation with for- mal government due to enduring norms of distrust. This paper also provides new insights on the development consequences of conflict (e.g., Collier, 2008; Blattman and Miguel, 2010; Bauer et al., 2016; Leon, 2012; Fergusson, Ib´ ˜ and Riano, 2020; anez ˜ and Valencia Caicedo, 2020). We complement previous work by showing that the economic Riano legacies of war and their effects on social norms are not only by-products of violence or the de- struction of factors of production but also by-products of institutions left by rebels. Furthermore, we offer evidence that these effects can be unequally distributed because they arise from changes 7 in economic and institutional structures, not solely from higher levels of violence. This distinction is necessary to understand the lasting effects of conflict.9 If destruction of physical capital or the temporary reduction of human capital due to violence could explain all the effects, the negative development impacts could be mitigated in the short to medium term (Miguel and Roland, 2011). However, if effects on development emerge from structural changes in the economy and social norms, they will be more persistent and difficult to change. In addition, we contribute to a growing literature on rebel governance by considering the effects on development in areas that have experienced uncontested control by insurgents. Scholars have recently shown that non-state actors can govern the political, economic, and social lives of resi- dents in an orderly fashion and establish institutions that regulate civilian behaviors (Arjona, 2016; anchez De La Sierra, 2020; Grasse, Sexton and Breslawski, 2021; Loyle et al., 2021; Stewart, 2018; S´ Wright, 2021; Liu, 2022). These studies show that armed groups must first win over local pop- ulations. This phenomenon occurred extensively in El Salvador (Wood, 2003), where rebels had an incentive to create systems of governance (Arjona, Kasfir and Mampilly, 2015). Most previous work focuses on factors that produce rebel governance, so little is known about how it affects de- velopment outcomes or whether any effects remain after rebels relinquish control. The design of effective post-conflict policies depends on understanding the impacts of conflict in specific con- texts and identifying the mechanisms behind those effects. Finally, the mechanisms we study link to recent evidence that highlights how land fragmentation can change economic development (Foster and Rosenzweig, 2022; De Janvry et al., 2015). We offer novel proof that historical factors are key to understanding the role of land fragmentation in economic growth. In particular, we show that land is more fragmented after rebel occupation and governance. Moreover, it was challenging to consolidate plots in former guerrilla territories because distrust of out-groups was rampant and hampered the benefits of scale. II HISTORICAL BACKGROUND II.A The origin of Salvadoran guerrillas: The FMLN The leading causes of the civil war and the motivation for the FMLN both originated in the coun- try’s long history of authoritarian rule, political exclusion, and economic inequality. As early as the 1930s, most agricultural land was owned by a small group of coffee-plantation owners who 9 While there is agreement on the negative impacts of conflict in the short run, there is no consensus on long-term ˜ and Valencia Caicedo, 2020). effects (Riano 8 met their demand for workers through a mostly unfree labor force that lived in harsh conditions. Economic modernization that occurred after the Second World War led to the expansion of com- mercial crops but did little to diversify the elites who controlled crop cultivation and exports as well as the incipient financial and manufacturing sectors (Colindres, 1976; Sevilla, 1985). Years of military rule helped forge an alliance between these elites and politicians that was based on the maintenance of class structures and the exclusion of peasants and workers from the political system (Wood, 2003). As a result, rural labor unions remained illegal, labor practices continued to be coercive, and land was unequally distributed. Although such political exclusion has characterized most of El Salvador’s history, the military regime allowed some level of political competition in the 1960s. However, this process came to a halt in 1972 after the mayor of San Salvador, Jos´ ´ Duarte—a popular opposition e Napoleon candidate—was allowed to run for president. Duarte won but the military quickly overruled the results. This decision sparked protests and mass mobilization in urban and rural areas that met with brutal repression. Peasants responded with outrage to the assassinations of rural leaders, students, teachers, and peers during these years (Wood, 2003). Security forces machine-gunned several marches and state resources flowed to paramilitary organizations and “death squads” as well. By 1980, more than one thousand people each month were killed for political reasons. By the mid-1970s, several guerrilla groups were operating in the San Salvador area. By the late 1970s, five major guerrilla organizations were recruiting supporters among students and work- ers in cities and among peasants in rural areas. Confronted with the growth of the guerrilla movement, divisions within the oligarchic alliance began to deepen; in October 1979, a group of reformist military officers overthrew the president and installed a new junta. Yet, instead of changing strategy, these new leaders tightened the repression of guerrilla groups (Wood, 2003). As El Salvador spiraled towards civil war, the five biggest guerrilla groups founded the FMLN in November 1980. In January 1981, they launched their first major operation, usually known as the “final offensive.” Although this failed to unseat the government, it consolidated the FMLN as the major fighting force against the Salvadoran state and provoked a change in strategy as the guerrillas retreated to rural areas to regroup and prepare for a longer fight. During the first years of the war, the FMLN forced landlords (who had begun to flee the country- side in the late 1970s) and the government out of rural areas. At the peak of the war in 1984, the FMLN had an estimated 8,000 to 15,000 combatants (Williams, 1998), ran operations in 30 percent 9 of the country (70 municipalities out of 262), and controlled 80 percent of all strategic territory (FMLN, 1984). Intense and indiscriminate state violence in disputed areas after the war’s onset caused the insurgent ranks to grow and motivated many previously apolitical peasants to fight for the rebels.10 As Wood (2003, p.18) shows in her extensive work on collective action during the Salvadoran Civil War, participation in the insurgency was mostly voluntary and explained as “an act of defiance of long-resented authorities and a repudiation of perceived injustices (particularly the brutal and arbitrary violence by security forces).” This indiscriminate state violence against civilians also drew the attention of human rights ac- tivists abroad and highlighted the role of the United States in training and arming the Salvadoran military. As a result, U.S. government officials withdrew their financial and military help and persuaded Salvadoran military leaders to curb their violence against civilians in late 1983 (Wood, 2003). Consequently, levels of violence dipped in 1984; this was also related to the election of a civilian president—former San Salvador mayor Jos´ ´ Duarte—and to a change in FMLN e Napoleon strategy based on the establishment of zonas liberadas (liberated zones) in the countryside. Most analysts argue that by 1985, the war had reached a stalemate, and the FMLN’s hold on its areas was stable and undisputed. By 1989, the FMLN was strong enough to plan and launch a massive offensive in several urban areas. This led to the formal negotiations in 1990 under U.N. mediation that ultimately ended the war on January 16, 1992. II.B Boundaries of FMLN territorial control The treatment of interest is full territorial control by insurgents between 1985 and 1992. The boundaries that define assignment to treatment are shown in Figure 1.11 Areas inside these bound- aries were under guerrilla control, while areas outside were either controlled by the Salvadoran Armed Forces or disputed by both parties. Evidence suggests that military and geographic consid- erations, such as protection offered by mountains and hills (as opposed to economic differences at ´ the boundaries of interest) explain the formation of these areas of control (Alvarez, 2011). Indeed, as shown below, the rebels did not select areas based on preexisting economic conditions. As one FMLN commander (1984, p. 2) wrote in his memoir: “The domain of most of the strategic elevations and the northern mountain range gives the FMLN a total topographical advantage over the army.” 10 Violence during El Salvador’s civil war was lopsided: state agents were responsible for 85 percent of deaths, most of which were civilians (Green and Ball, 2019). 11 As mentioned in the introduction, this map was used in the peace accord meetings between the Salvadoran govern- ment and the FMLN from April 1990 to January 1992. It is typically viewed as recognition by the state of the magnitude of the insurgent territorial presence (Ch´avez, 2011). 10 Initially, the Salvadoran state entirely controlled the regions under analysis. In 1981, the guerrillas conducted a countrywide offensive against 12 main military bases in a bid to promote an insur- rection (MINED, 2009). Although this failed, it prompted a change in military strategy and thus the group’s geographic dispersion with the goal of establishing a presence on all fronts through the aforementioned liberated zones. The first of these was organized as early as 1982 (Castaneda, ˜ 2016). Liberated zones are a key guerrilla warfare tactic and comprise areas where insurgents can generate support by providing basic public goods and establishing their own institutions. The concept dates back at least to Mao Zedong’s military strategy in which base areas were a winning tactic against a conventional army. They consisted of local strongholds in (preferably) mountain- ous areas where insurgents could elicit popular support by creating systems of governance (Mao, 1966). As this idea evolved, the strategic location of these zones in mountainous areas remained important and it has been copied by non-state armed actors ranging from communist guerrillas in Guatemala in the 1980s (Moran, 1985) to ethnic armed organizations in Burma in 2021.12 Importantly, historical evidence and FMLN documents suggest that after 1984, the boundaries of FMLN-controlled areas were extremely stable for at least two reasons. First, by 1984, the FMLN controlled approximately 80 percent of the militarily strategic territory (FMLN, 1984). Second, by that same year, more than 80 percent of the Salvadoran Army’s offensive capacity was in perma- nent use. Therefore, they could not reconquer areas under FMLN occupation but instead had to strengthen the defense of areas the state still controlled (FMLN, 1984). Hence by 1985, the conflict ˜ had effectively entered a virtual stalemate (Castaneda, 2016). As a result, we focus on guerrilla- controlled areas that were stable between 1985 and 1992. Figure 1 shows the three strips of the country where the FMLN had full control by 1985: the northern, central, and coastal areas.13 II.C Rebel governance in FMLN-controlled areas Our study region experienced significant cultural, social, and economic transformations during FMLN control. As mentioned in Section II.B, the creation of systems of governance to win over local populations is a fundamental part of rebel strategies that can leave profound changes in their wake. As Wood (2003) emphasizes, rebel control in El Salvador left a legacy of new values, norms, economic practices, and beliefs that contrasted sharply with the prevailing culture before the war 12 Several cases of self-governing institutions can be found in varying locations throughout history, including the Democratic Republic of the Congo, South Sudan, Liberia, Indonesia, Bolivia, Colombia, El Salvador, and Guatemala. See Arjona, Kasfir and Mampilly (2015) for an analysis of some of these case studies. 13 The absence of an FMLN presence in the western region is usually attributed to the legacies of the massacres of indigenous peasants by state and paramilitary forces in the 1930s (Lauria-Santiago and Gould, 2008). 11 and with those in areas under state control. Before the FMLN arrived in rural areas, coercive methods helped maintain an acquiescent peas- ant culture. Several scholars have noted that as a result of the unequal distribution of land and coercive labor practices, attitudes of fatalism, self-deprecation, and individualism were com- mon among peasants in El Salvador during most of the twentieth century (Mart´ ´ 1990; ın-Baro, Wood, 2003). This culture of individualism and fatalism can be traced to peasant dependency on landowners, which undermined links with their peers and bred resentment towards political and economic elites (Scott, 1972; Mart´ ´ 1990). ın-Baro, The guerrillas transformed many of these patterns when they consolidated their control in 1985. Three major changes took place: the promotion of civil society and self-governance initiatives to address basic needs, the fragmentation of landholding, and new patterns of production. Rebel groups eliminated the state’s local and judicial administration in the areas they controlled ın Alvarez, 2010), creating a pressing need for new institutions to administer the life of (Mart´ civilians there. Thus, the FMLN promoted the formation of semi-autonomous local councils, the Poderes Populares Locales (PPL) first and the “dual powers” (Poderes de Doble Cara) later, to substi- tute for formal state authorities (FMLN, 1984; Pearce, 1986; Binford, 1997). These new governing structures administered and organized the local population; their main purpose was to procure public goods and resolve issues affecting the community (Pearce, 1986). Although they existed in various forms, all fostered democratic activity by residents. Peasants participated in their own government and largely viewed these local powers as legitimate (Pearce, 1986). Some groups had popular assemblies, and sometimes they also held elections for positions. These organizations addressed issues ranging from water provision to the establishment of community legal codes (Pearce, 1986; FMLN, 1983). The guerrilla-controlled areas witnessed the emergence of diverse and plentiful civil-society in- ´ stitutions to organize peasants and handle pressing development issues (Alvarez, 2013; Velado, 1993). The FMLN supported these community-based groups and viewed them as a way to orga- nize the population independently of both the state and the guerrillas (FMLN, 1984). These organizations flourished, not least because of their autonomous nature. With the change of government strategy from indiscriminate violence (due to the U.S. constraints) and the consolida- tion of areas controlled by the FMLN in 1985, the autonomy of these institutions meant residents 12 did not fear being labeled as insurgents by the government, and they were also protected by the FMLN (Binford, 1997). Despite overwhelming peasant support for the FMLN, neutrality was pos- sible and common inside FMLN areas because the guerrillas were extremely restrained in their use of violence and promoted autonomy as a policy (Wood, 2008). According to various schol- ars, two major legacies of rebel governance and the promotion of civil-society organizations were the consolidation of close-knit communities and the creation of alternative political and economic models of development (Binford, 1997). Crop production also changed in FMLN areas. Large landowners had begun to flee in the late 1970s as social unrest grew. In occupying their areas, the guerrillas targeted export crops for sab- otage, which further eroded the surviving haciendas and massive agricultural production. Since many food products were unavailable during the war, the FMLN promoted subsistence farming and supported the expropriation of large and abandoned landholdings (Wood, 2008). It permitted peasants to occupy land regardless of whether or not they participated in the insurgency (Wood, 2003), which divided large properties into smaller plots (Wood, 2008). These new models of pro- duction and labor contracts led to the “peasantization” of formerly commercial agriculture and the fragmentation of rural markets. As such, most large entrepreneurs and large-scale agricultural concerns were eliminated and replaced by cooperatives and individual peasant farmers (Wood, 2010; Binford, 1997). By the end of the war, rebel control had reshaped the social and economic landscape. These areas now featured close-knit communities with access to small plots of land. Residents there self- organized to provide public goods and profoundly distrusted and resented repressive political and economic elites. II.D Post-conflict context After the government and the FMLN jointly approved the Chapultepec Peace Accords on January 16, 1992, the Salvadoran Civil War ended. Scholars often refer to these peace agreements as the most successful in the post-Cold War period. Why? The cease-fire held; the FMLN became a legal political party; military, judicial, and electoral institutions were reformed; an Office of Human Rights Council was established; a Truth Commission was formed; and a limited agrarian reform was enacted (Moodie, 2011). The Catholic Church and the United Nations were the mediators of the peace talks that culmi- 13 nated in a final agreement regarding five main areas (United Nations, 1992). First, the armed forces were modified and the FMLN was demobilized.14 Second, the National Civil Police (NCP) replaced the National Guard.15 Third, there were modifications to the judicial system and the defense of human rights.16 Fourth, the electoral system was modified to create the Supreme Elec- toral Tribunal, the highest administrative and jurisdictional authority on elections. At the political level, the measures sought to guarantee FMLN leaders and their members the full exercise of their civil and political rights within a framework of absolute legality. Finally, measures were imposed in both the economic and social fields. The main ones included land distribution to landless peas- ants and ex-combatants from both the Salvadoran military and guerrilla groups. Moreover, the agreement established that land tenure inside territories not controlled by the state was to be hon- ored and land titles were to be granted to peasants working there at the moment. Overall, the proposed policy changes—with the exception of the land tenure measure—were implemented at the national level. During this period, the Salvadoran private sector boomed and the economy moved away from a concentration of power among 14 elite families to open to international markets (Boyce, 1995). This was reflected in the transformation of the economy from a primarily agricultural model of coffee, sugar, and cotton exploitation towards more diversified growth in commerce, agricultural export businesses, industry, and financial services. For instance, while agricultural exports represented approximately 25 percent of Salvadoran GDP in the 1970s, agriculture’s share fell to less than five percent (Rettberg, 2007) towards the end of the century. Salvadoran businesses blossomed and between 1990 and 1995, the economy grew at an average rate of 6.2 percent, much of it nurtured by growing domestic investment rates in commerce, financial services, and industry. One of the winners was the private sector, which received stability, a friendly investment climate at home, and economic rules that enabled it to compete in a new international macroeconomic environment (Rettberg, 2007). Unfortunately, however, economic prosperity was just as unevenly distributed during this period as in the past (Moodie, 2011). 14 According to the agreement, the armed forces’ sole objective would be to defend the sovereignty of the State while remaining apolitical and respecting human rights. 15 The NCP replaced the old security forces with a civil and democratic doctrine, quotas were established for the new personnel in which demobilized elements of the FMLN and former National Police would participate, and a National Academy of Public Security was created to train the agents of the NCP with an emphasis on respect for Human Rights. 16 Measures included the creation of the Judicial Training School to train judges and magistrates to adjust to the country’s new reality, a reform of the National Council of the Judiciary (which appoints and evaluates judges) to give it greater independence, and a reform of the election process and terms of the magistrates of the Supreme Court of Justice. 14 Although the FMLN established itself as a political party in the presidential election of March 1994, it lost by a wide margin. The two major candidates were Ruben Zamora of Democratic Con- ´ Sol of ARENA, a right-wing party. Voter fears that a vergence (the FMLN) and Armando Calderon left-wing victory might lead to a new wave of political violence were reinforced by ARENA’s cam- paign ads that featured images of wartime destruction (Wantchekon, 1999). These led to its victory and subsequent pessimism about the prospects for democratic consolidation in El Salvador. The FMLN finally secured a presidential victory with the election of Mauricio Funes in 2009. However, during Funes’s presidential term, several corruption scandals erupted and no substantial policy changes were implemented relative to ARENA. III DATA This section describes the primary sources of data used in the study. Appendix A presents a detailed account of the database construction and Appendix B presents summary statistics of all variables employed in the analysis. III.A Guerrilla-controlled territories To analyze the role of guerrilla territorial control in long-term development, we geocoded the map that depicts FMLN-controlled areas (see Figure 1). It shows areas the FMLN controlled during the conflict, areas the state controlled, and disputed ones. As Figure 2 illustrates, no boundaries of the guerrilla-controlled territories coincide with the administrative departments and municipal boundaries of El Salvador today. Therefore, our estimated treatment effects are unlikely to be con- taminated by the influence of a compound treatment comprised of guerrilla control and changes in administrative boundaries. Similarly, Table 1 rules out that we are detecting the effect of the land reform and guerrilla control simultaneously. In sum, the evidence presented so far implies that our identification strategy credibly isolates the effects of guerrilla control from other potential confounders. 15 Figure 2. Guerrilla-Controlled Territories and Administrative Boundaries (a) Departmental Boundaries Under Guerrilla Control (b) Municipal Boundaries Under Guerrilla Control Notes: The figure presents the areas under guerrilla control in red and shows that these areas do not coincide with the administrative departments and municipal boundaries of El Salvador today. III.B Geospatial variables We use geospatial data to test the validity of the local continuity assumption around the bound- aries of guerrilla-controlled areas. Elevation was obtained from NASA’s Shuttle Radar Topogra- phy Mission (SRTM). Information on surface water bodies comes from the MERIT Hydro dataset. Agro-climatic yield rasters with a spatial resolution of five arc minutes (nine km) come from the Global Agro-Ecological Zones (GAEZ) project. For all yields, we are using the 30-year average beginning in 1961. Figure C.1 maps guerrilla-controlled territories, altitude, and main rivers in El Salvador. It il- 16 lustrates that the rebels located disproportionately in high altitudes as part of their war strategy (FMLN, 1984) and that rivers often marked the boundaries of their territories. III.C Development outcomes The long-term development impacts of guerrilla territorial control are measured using 2013 night light luminosity (as a proxy for local economic activity) and 2007 population and household cen- sus data. Night light luminosity. Data on night light luminosity comes from the Defense Meteorological Satellite Program Operational Linescan System. This data was obtained from the US National Oceanic and Atmospheric Administration (NOAA) web page. It has a resolution of 30 arc seconds × 30 arc seconds (i.e., approximately 1 km × 1 km) and spans 1992 to 2013. The main results use data for 2013 as it is the last year of data available from the Operational Linescan System (OLS) flown by the U.S. Air Force Defense Meteorological Satellite Program (DMSP). To study the durability of effects, we also used individual years between 1992 and 2013. 2007 Population and Household Census. The General Directorate of Statistics and Censuses (Di- on General de Estad´ recci´ ısticas y Censos, DIGESTYC) provided anonymous microdata from the 2007 census for this study. The data includes the socioeconomic characteristics of all households and individuals, including but not limited to labor market outcomes; educational attainment; material ownership (e.g., having a car, a TV, etc.); use of public services (electricity, sewerage, and others); migration; and other characteristics of all dwellings in El Salvador. 2007 Census Cartography. DIGESTYC also provided maps of the tracts for the 2007 census. Each tract represents a small area with a fixed geographic perimeter. In 2007, the average tract in our sample included 131 households and 473 individuals. Small tract units facilitate accurate identifi- cation of guerrilla territorial control, which is approximated using the geographic coordinates of the tract centroids. In sum, we explore the effects of FMLN control via night light luminosity,17 human capital (mea- sured as years of education and literacy rates), and a wealth index (constructed as suggested by the Demographic and Health Surveys program).18 The wealth index is the first factor from the 17 The challenge of night light luminosity data is the significant fraction of observations that take the value of zero and also the existence of extreme values in the right tail of the distribution (Michalopoulos and Papaioannou, 2013; Pinkovskiy and Sala-i Martin, 2016). To account for this concern, the outcome is transformed using the inverse hyper- bolic sine transformation, which can be interpreted as a logarithmic dependent variable (Pence, 2006). 18 Step-by-step instructions for constructing the index are available at: https://dhsprogram.com/topics/ 17 principal component analysis of a household’s cumulative living standard. The estimates use the average index of all households in each census tract. The index includes household characteristics such as asset ownership (e.g., bicycles and television); materials used for housing construction; types of water access; and sanitation facilities. IV EMPIRICAL STRATEGY IV.A Spatial regression discontinuity design We estimate the long-term development impacts of rebel territorial control between 1985 and 1992 using a spatial regression discontinuity design around the boundaries illustrated in Figure 3. The specification is: 400 i ys = β1 Ts + β2 f (ds ) + β3 Ts × f (ds ) + αs + εs (1) i=1 where ys represents the contemporaneous economic and social development outcomes of interest observed at the census tract unit s. Ts is a treatment indicator equal to one if the tract intersects a guerrilla-controlled zone. ds is the normalized perpendicular distance from each tract’s centroid to the guerrilla-controlled boundary.19 f (ds ) is a polynomial function of the distance to the bound- ary which, interacted with Ts , controls for smoothness in the geographic location at each side of the boundary. Finally, since we want to compare treatment and control census tracts that are geo- i splits the boundary in four km segments and equals one if graphically proximate, the indicator αs census tract s is closest to segment i, and zero otherwise. We include 400 fixed effects for the min- imum distance from the centroid of each tract to each of 400 segments of the guerrilla-controlled boundary.20 Standard errors are adjusted for heteroskedasticity. As a robustness check, we also estimate Conley standard errors to account for spatial correlation in the data (Conley, 1999). The baseline results use a local linear polynomial of the normalized distance and limit the sample to tracts within the distance suggested by the optimal bandwidth algorithm of Calonico, Cattaneo and Titiunik (2014) when using night light luminosity as an outcome (which represents approxi- mately 2.26 km). We also present the results under a variety of different bandwidths to check the wealth-index/Wealth-Index-Construction.cfm 19 As a result of the distance normalization, tracts touching the guerrilla-controlled boundary get the value of zero in their distance variable and tracts outside the guerrilla-controlled area get a negative value, contrary to tracts inside. 20 The choice of 400 breaks is to account for enough spatial variation without compromising the variation we are exploiting. 18 robustness of the main findings given the classic trade-off between bias and power. Figure 3. Census Tracts and Boundaries Employed in the Empirical Analysis Notes: The figure shows the actual census tracts used in the analysis in purple. The selected tracts are within approx- imately two km of the guerrilla-controlled boundary (see Figure 1), which is the optimal bandwidth when using the algorithm of Calonico, Cattaneo and Titiunik (2014). IV.B Validation of the local continuity assumption This subsection shows that before the FMLN began to control territories in 1985, areas around the boundary had similar geographic and socioeconomic characteristics. Geography and socioeconomic development before guerrillas controlled the territory To ensure census tracts outside the boundary are an appropriate counterfactual for guerrilla-held ones, we first tested for preexisting differences in geographic and socioeconomic characteristics an and the before the start of guerrilla control. Much of the boundary follows the Sierra of Metap´ Sierra Madre formations (Figure C.1). For this purpose, we estimated equation (1) to test for discontinuities related to geographic char- acteristics (e.g., elevation, slope, and access to waterways) and some socioeconomic characteristics (e.g., road and railway density in 1980 and crop agro-climatic yields from 1961 to 1979). Table 1 19 shows that 33 out of 35 baseline covariates are statistically similar across the boundary. The only exceptions are altitude and sugarcane yields. The statistical significance of the difference in sugarcane yields between controlled and other areas could be driven by the difference in al- titude because that variable is included in the mechanical estimation of past potential sugarcane yields.21 A potential concern with the difference in sugarcane yields is that these can come from land concentration. However, as we show in Table 1, there are no differences in the probabil- ity of being part of the 1980 Land Reform across the boundary, which is a good proxy for land concentration in the 1980s and the strength of the elites. Moreover, the difference in mean altitude is very small (17.13 m from a dependent mean of 502.7) and aligns with the observation that the guerrillas occupied higher territories as a military strategy. Yet, we show that the difference in altitude is not correlated with variables that proxy state capacity or economic development. As a robustness exercise, we include altitude as a control variable and find that all variables are statistically similar across the boundary (see Figures D.1 - D.2), confirming the validity of the local continuity assumption for all outcomes.22 Importantly, results also show that before the guerrillas controlled these territories, the areas were similarly likely to be subject to the national land reform. This initiative redistributed large ha- ciendas to peasants in 1980 in an attempt to palliate increasing levels of distrust of the state and mobilization by the peasantry. Therefore, this result of no differences in land reform across the boundaries shows that the FMLN did not establish its control in places with higher land inequal- ity or differentials in elite strength. Moreover, we find no differences in the number of violent events or massacres across the boundary, yielding further evidence that guerrillas did not estab- lish territorial control according to levels of distrust of the state that are generally associated with historical state and elite repression. Furthermore, we tested the robustness of the validity of the local continuity assumption to dif- ferent choices of bandwidth distance around the cutoff. The results confirm the robustness of the local continuity assumption to the bandwidth choice (see Figures D.5 - D.7).23 21 According to the model documentation of GAEZ v.3 project, altitude and terrain variables are used in the first stage of the agro-climatic analysis. 22 The specification that controls for altitude is not used to report main estimates as it may result in biased coefficients. The estimate that can be identified when adjusting for imbalanced covariates in RD designs is a weighted average of the treatment effects where the weights depend on the conditional distribution of the imbalanced covariate on the treatment, which is not our estimate of interest. See Calonico et al. (2019) for a discussion. 23 An additional assumption is that there should be no selective sorting across the boundary. We discuss this assump- tion in depth later. 20 Table 1. Smooth Condition Test Variable (Year) Coefficient SE Dependent Mean Obs Geographic Characteristics (Before 1980) Altitude (1980) 17.13*** 5.679 502.7 3,652 Slope (1980) 0.352 0.222 7.160 3,652 Ruggedness (1980) 0.440 0.321 10.28 3,652 Hydrography (1980) 0.0261 0.0246 0.230 3,652 Roads and Railway (1980) 0.0198 0.0284 0.370 3,652 Has a City or Village (1945) 0.0136 0.0225 0.100 3,652 Distance to City or Village (1945) -0.0534 0.0464 1.000 3,652 Infrastructure Characteristics (Before 1980) Distance to Communications (1945) 0.0639 0.0501 0.900 3,652 Communications Density (1945) -0.0535 0.0602 0.430 3,652 Part of Land Reform (1980) -0.0135 0.0163 0.0600 3,652 Inside a Wide Cultivated Area (1980) -0.0193 0.0132 0.790 3,652 Had a Parish (1979) -0.00395 0.00565 0.0100 3,652 Distance to Parish (1979) 0.0528 0.0698 3.420 3,652 Distance to School (1980) 0.0782 0.0791 12.11 3,652 Population Demographics (Before 1980) Total Population (1980) 3.010 4.085 162 3,652 Population Density (1980) -171.0 110.2 2165 3,639 Years of Education (1980) -0.160 0.113 4.410 3,639 Birth Rate (1980) -0.0445 0.0757 0.190 3,639 In-migration Share (1980) -0.0110 0.00827 0.140 3,639 Out-migration Share (1980) -0.000182 0.00103 0.0100 3,639 Inside a Highly Populated Area (1980) -0.0160 0.0116 0.800 3,639 Agro-Climatic Potential Yield (1961-1979) Aggregate Yield Index (1961-1979) 0.0269 0.0165 0.0900 3,652 Bean Potential Yield (1961-1979) 0.00514 0.00385 4.080 3,639 Coffee Potential Yield (1961-1979) 0.00385 0.00426 1.690 3,639 Cotton Potential Yield (1961-1979) 0.000764 0.000516 0.710 3,639 Maize Potential Yield (1961-1979) -0.00918 0.0116 9.850 3,639 Wet Rice Potential Yield (1961-1979) 0.0134 0.0115 8.790 3,639 Sugarcane Potential Yield (1961-1979) 0.0460** 0.0196 6.500 3,639 Crops’ High Suitability (1961-1990) Bean High Suitability (1961-1990) -0.0149 0.0105 0.930 3,652 Coffee High Suitability (1961-1990) -0.0145 0.0123 0.150 3,652 Maize High Suitability (1961-1990) 0.00174 0.00511 0.990 3,652 Sugarcane High Suitability (1961-1990) -0.0148 0.0125 0.180 3,652 Conflict (Before 1981) and Incarcerations (1980-1985) Number of War Events (1981) 0.00660 0.0894 0.0410 3,652 Number of War Victims (1981) -0.258 0.4900 0.213 3,652 Number of Incarcerations (1980-1985) 0.00780 0.00740 0.0210 3,681 Notes: The table presents the results of estimating equation (1) for a variety of geographic characteristics, roads and infrastructure availability, demographic characteristics, agro-climatic potential yields, indicators for crop suitability, and for outcomes related to conflict before the guerrillas’ settlement. The information was gathered from diverse sources (See Appendix A for more details). Crops were selected according to their relevance for domestic consumption and exports. The unit of observation is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla- controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates were weighted using a triangular kernel. The dependent mean corresponds to the mean outside the territories of guerrilla control but within the area of analysis. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 21 V MAIN RESULTS V.A Night light luminosity, wealth, and human capital Table 2 presents formal estimates of equation (1) for the main outcomes of interest. All estimates suggest strong and negative impacts of guerrilla territorial control on development outcomes. First, the results show that locations within former guerrilla territories had lower night light lu- minosity in 2013, relative to places outside these areas. The effects are sizable. Approximately 20 years after the end of the Civil War—and about 30 years after guerrillas first controlled these areas—lands that were once under FMLN rule experienced nearly 18.6 percent lower night light luminosity than places with no guerrilla control (see Column 1). These results are robust to differ- ent transformations of the dependent variable (see Table D.1). Considering that a one percentage point (pp) change in luminosity corresponds to a 0.28 pp change in GDP (Henderson, Storeygard and Weil, 2012), areas that had been under guerrilla control had approximately 5.2 percent lower GDP (18.6×0.28 = 5.2) than areas that had not. Second, we document that areas once controlled by the guerrillas are less wealthy and have lower human capital almost two decades after the end of the Civil War. Column 2 of Table 2 shows that areas controlled by guerrillas had a wealth index 0.121 sd lower than other areas. Consistent with these negative effects on wealth, Column 3 shows that residents of areas close to the border but still under guerrilla control had 0.28 less years of education by 2007. In Table D.2, we present the analysis by cohorts that were exposed to guerrillas versus cohorts that had already finished their education when guerrillas arrived in the areas where they lived. The effects are driven by individuals who were school-age during the war, whereas individuals across the boundary who finished their education before 1980 had similar years of education.24 The graphic representation of these effects is in Figure D.9, which depicts a decline in all the outcomes inside former guerrilla areas. The discontinuity is especially strong for night light lumi- nosity. All in all, the estimates present negative and sizable impacts of guerrilla territorial control on long-term development outcomes. 24 In Column 4 in Table D.1 in the Appendix, we also study literacy rates. These were constructed as the number of individuals 18 years or older who can read, divided by the total number of individuals older than 18 years. We find individuals in former FMLN areas had 2.1 percent lower literacy rates than people living outside these areas. This corresponds to a 2.6 percent drop relative to the average literacy rate in 2007. 22 Table 2. Effects of Guerrilla Territorial Control on Night Light Luminosity, Wealth, and Human Capital Night Light Arcsine Wealth Index Years of Education (2013) (2007) (2007) (1) (2) (3) Guerrilla control -0.186*** -0.121*** -0.279** (0.0247) (0.0355) (0.109) Observations 3,652 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 Dependent mean 3.536 -0.0160 6.573 Notes: The table presents the results of estimating equation 1 for the main outcomes. Column 1 shows the effect of whether a census tract was under guerrilla control on the arcsine of night light luminosity from NOAA. Column 2 uses the standardized score of household wealth as dependent variable in the same estimation. Column 3 shows as dependent variable years of education of the population older than 18 years. The unit of observation in all columns is the census tract. Information from Columns 2 and 3 was obtained from the Population Census of 2007. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with an indicator of whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Figure 4 presents effects on night light luminosity for all years of data from 1992 to 2013 to test whether effects endure over the years. The results suggest that not only did negative effects persist since 1992, but also the magnitudes barely changed over the years. Overall, these results confirm our quantitative findings that guerrilla control produced a lasting negative effect on long-term development outcomes. Section VI explores mechanisms to explain this persistence. V.B Robustness checks We used three approaches to test the robustness of our results: (i) the estimation of Conley stan- dard errors and the use of alternative RD and Ordinary Least Squares (OLS) specifications, (ii) a placebo test that uses difference in altitude to define artificial boundaries, and (iii) a restriction in population sorting across boundaries. V.B.1 Alternative specifications First, to account for spatial correlation in our data, we estimate Conley standard errors following Conley (1999). As we show in Table D.3, the statistical significance of the estimated effects remains the same. Next, to see whether the results originate in specific regression discontinuity specifica- 23 tions, we conducted a number of robustness checks (see Appendix D). One potential concern is that the results are valid only for the selected bandwidth. Figure D.10 illustrates that the effects of FMLN territorial control on the main outcomes are robust to different choice of bandwidths between 0.1 and 4 km. Second, in Tables D.4- D.6 in the Appendix, the main results are presented using alternative RD polynomials (constant, linear, and quadratic); using additional bandwidth options; and varying the kernel choice. Lastly, we estimate our main model using the OLS estima- tion approach (see Table D.7). Overall, the results are robust to all these alternative specifications. Figure 4. Effects of Guerrilla Control on the Arcsine of Night Light Luminosity Over Time 0 -.05 Effect on Night Light (Arcsine) -.1 -.15 -.2 -.25 92 93 94 95 96 97 98 99 00 01 02 03 04 05 06 07 08 09 10 11 12 13 Year Notes: This figure shows the coefficients obtained from the estimation of equation 1 for each year between 1992 and 2013. The gray coloring illustrates 95 percent confidence intervals. The estimates shown include up to 400 fixed effects. The figure illustrates the coefficients of each yearly estimation from 1992 to 2013. Overall, the effect of guerrilla control on night light luminosity is negative and stable over time. V.B.2 The use of altitude to define borders One relevant concern regarding our empirical strategy is that since FMLN territories were defined using altitude as the main geographic feature for the borders, the results may reflect some socioe- conomic characteristic associated with higher-altitude areas rather than rebel control. As shown above, there are no statistical differences in variables that measure economic productivity and state capacity at baseline in areas later controlled by the FMLN. Nevertheless, we conducted a placebo exercise by selecting pairs of neighboring census tracts in 24 areas that were never under guerrilla control but which have the same difference in altitude as tracts inside FMLN areas (Table 1). The intuition here is that if negative effects on development outcomes stemmed from significant altitude differences, there would be similar effects on out- comes in areas with the same altitude differences that were not under FMLN control. Results are in Table D.8 in the Appendix. The effects on development are positive and smaller in magnitude than the estimated effects for FMLN control. Moreover, we repeat the same exercise with tracts outside guerrilla areas that have larger altitude differences. Even in this extreme case (that com- prises a small percentage of tracts in our sample), the effects are small. Finally, Table D.9 shows estimates of the main effects when we restrict the sample to census segments without a sudden change in altitude relative to their immediate neighbors. Results do not change. These findings provide evidence that the main effects are not the by-product of higher altitudes but rather the consequence of guerrilla control. V.B.3 Population sorting One potential concern is that individuals in FMLN areas may have moved to nearby areas (our control group) by the time the boundaries formed. We address this concern by evaluating the effects for individuals who never moved (or “stayers.”) Table D.10 shows that results are of similar magnitude and significance as for the whole sample, suggesting that in-sample migration may not be a concern. Figure 5 presents more evidence that the effects do not arise from out-migration from FMLN territories. The figures illustrate the estimates of equation (1) on education outcomes observed at the individual level for the subsample of “stayers.” As shown in Figure 5, effects remain negative and statistically significant. These results align with qualitative evidence that shows the guerrillas provided key defense func- tions for peasants in their areas (Pearce, 1986), suggesting we should not expect out-sorting to areas controlled by the Salvadoran state. Finally, we explore whether recent and selective migration at the time of the boundary could explain differences in economic development across the boundary, but we find no evidence of it. In particular, we trimmed the sample in two ways: first, we omitted the 10.4 percent of the control-group sample with the highest education and wealth, as in-migration to nearby control areas is 10.4 percent. Second, we omitted the 3.3 percent of the guerrilla sample with the lowest education and wealth, as in-migration to guerrilla areas is 3.3 percent. The estimates based on the trimmed samples remain similar (see Table 3). Moreover, we take advantage of the fact that 25 the census contains information on the year individuals arrived in each location to account for in- sample migration in 1980 and 1985 in Columns 3–4 and 5–6. Results do not change. In addition, rates of migration across the boundaries are very low (less than one percent). Figure 5. Effects of Guerrilla Control on Education Outcomes of the Nonmoving Population Only (a) Wealth Index (b) Average Years of Education .5 .2 0 Coefficient magnitude Coefficient magnitude 0 -.5 -.2 -1 -1.5 -.4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) (c) Literacy Rate .02 0 Coefficient magnitude -.02 -.04 -.06 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Notes: The figure illustrates the results for each outcome variable obtained from the estimation of equation 1 using the “stayers” subsample. The gray coloring illustrates 95 percent confidence intervals. Overall, we find that the effects of guerrilla control on the three outcomes are consistent under a wide range of bandwidths (0.1 to 4 km). 26 Table 3. Accounting for Selective In-migration Trimming using the All-Time In-migration Rate 1980 In-migration Rate 1985 In-migration Rate Wealth Index Wealth Index Wealth Index Years of Education Years of Education Years of Education (1) (2) (3) (4) (5) (6) Guerrilla control -0.101*** -0.260** -0.121*** -0.277** -0.121*** -0.274** (0.0353) (0.107) (0.0358) (0.109) (0.0358) (0.109) Observations 3,630 3,637 3,630 3,637 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 2.266 Dependent mean -0.0330 6.538 -0.0280 6.570 -0.0280 6.565 Notes: The results follow the specification of equation 1 for the Wealth Index and Years of Education outcomes. How- ever, we trim the dependent variables by using different in-migration rates. In Columns 1 and 2, we use the all-time in-migration rate to trim the 10.4 percent most educated and wealthy people and the 3.3 percent least educated and wealthy from the treated and control groups’ respective distributions. In Columns 3 and 4, we use the in-migration rate from 1975 to 1980 to trim the 0.4 percent most educated and wealthy people and the 0.6 percent least educated and wealthy from the control group’s respective distributions. In Columns 3 and 4, we use the in-migration rate from 1979 to 1985 to trim the 0.7 percent most educated and wealthy people and the 0.8 percent least educated and wealthy from the control group’s respective distributions. The unit of observation in all columns is the census tract. Information from all columns was obtained from the Population Census of 2007. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. V.C External validity In Appendix D, we conducted two analyses to rule out that the results are specific to our RD sample. First, we show that at baseline the RD sample is similar in characteristics to the rest of the country (see Table D.11). Second, we show how the main results change once we move outside the two km bandwidth. Figure D.11 shows that results are pretty homogeneous across space. VI EXPLORING POTENTIAL MECHANISMS Why would the FMLN’s influence endure so many years after its territorial control ended? As discussed above, one explanation concerns the reshaping of economic structures and local gov- ernance that led to the demise of the former commercial agriculture system and disengagement with the state.25 In particular, reliance on plot-based agriculture and lack of public engagement with the state could have created a negative feedback loop that depressed living standards in the long run. We hypothesize that many of these changes continued through the present day due to lasting distrust of the state and elites. We find that even though former guerrilla areas today have 25 While most agricultural elites from the 1950s lost prominence after the conflict, distrust of the state and new elites or landowners could persist due to historical factors even if the new entrants provided better labor conditions. 27 more state presence, access to and utilization of public services is still lower than in the control group. We present similar findings for the agricultural sector. Even though the land can still pro- duce commercial crops that could exploit increasing returns to scale, we find that commercially exploited plots are smaller and a larger share of individuals work in subsistence agriculture in former guerrilla areas. We also find lower productivity and some evidence that the agro-based industry sector is less developed in former guerrilla areas. This pattern could also be explained by distrust of agricultural elites associated with the state. In interviews, we learned that even though commercially oriented farmers would like to invest in these areas, residents are reluctant to let them due to high distrust of out-groups. This is true even when the new entrants have no ties to elites from the 1970s. In this section, we also rule out alternative mechanisms such as an increase in violence, selective migration, lower public and private investment, and child recruitment. VI.A Transformation of social norms: lower trust in and engagement with the state and asso- ciated groups Both FMLN documents and scholarly work suggest that the organization of the rural population was a key strategy rebels employed against the state (FMLN, 1983, 1984; Binford, 1997; Pearce, 1986). In guerrilla areas, the FMLN’s social base set up participatory government to replace the municipal administration. As noted above, they eliminated state and judicial authorities and es- tablished community-based organizations—first the PPL and later the “dual powers” plus a vari- ety of others—to represent peasants and address key development issues. (Binford, 1997; FMLN, 1984). In particular, guerrillas implemented communal models of agricultural production with a focus on subsistence crops. Self-governance can promote the formation of social capital but also reduce trust in and engage- ment with the state and related groups. First, if self-governance presents an alternative to absent state institutions, it may reduce engagement with the state and the government even when the state regains control. Moreover, such disengagement can reduce demand for state services and create informational problems that prevent the efficient provision of public goods. Second, dis- engagement can lead to lasting norms of distrust of the state and associated groups such as large agricultural producers, preventing investment and reinforcing the subsistence economy through the present day. We study the validity of this mechanism by examining contemporary attitudes towards the state, public goods provision, and agricultural outcomes. Table 4 presents the estimates of equation (1) using data from the Latin American Public Opin- 28 ion Project (LAPOP) in 2004–2016 for outcomes related to trust and engagement.26 We used the data to construct four indicators of political attitudes and behaviors including: political participa- tion, engagement with politicians, nondemocratic engagement, and trust in institutions (see more details in Appendix A). Although individuals living in former FMLN areas are not less likely to participate in politics or to engage violently with politics (Columns 1 and 3), they exhibit less engagement with politicians and less trust in institutions (Columns 2 and 4). We also find evidence of more trust towards members of the community in these areas, providing further evidence on how former guerrilla governance may have reinforced social capital within the community and distrust in the state.27 Consistent with the lack of trust in politicians and the state, Table D.12 shows that residents of former FMLN areas were more likely to cast blank votes in the 2014 presidential elections and the 2015 municipal elections. This effect is small given that on average only a low percentage of people turn in blank votes (one percent at each polling station). We find that, if anything, residents of former guerrilla areas were less likely to vote for the left. However, these votes did not seem to benefit the right. For example, in the 2015 municipal elections, the share of blank votes increased at the expense of both the left and the right.28 26 LAPOP conducts surveys of public opinion throughout the Western Hemisphere, including North, Central, and South America and the Caribbean. LAPOP’s core project is the AmericasBarometer, a rigorous comparative survey of political and social attitudes and demographic and economic characteristics. 27 As we show in Table D.13, results are robust when we use the simple sum of questions related to each outcome instead of the inverse covariance index as in Table 4. 28 During the (post-conflict) 1989–2009 period, ARENA, the main right-wing political party in El Salvador, won all the presidential elections. After 20 years came the victory of the FMLN candidate Mauricio Funes (2009–2014), which raised expectations among FMLN supporters of an improvement in living conditions. However, several corruption scandals erupted during Funes’s presidential term. This could explain the differences in support for the FMLN presidential candidate in 2014 (Salvador Sanchez Ceren) in former guerrilla areas that we observe in the data. 29 Table 4. Effects of Guerrilla Territorial Control on Attitudes towards the Government Inverse Covariance Index (ICW) Political Engagement with Non-Democratic Trust in Distrust of Members of Participation Politicians Engagement Institutions the Community (Share) (2004-2016) (2004-2016) (2004-2016) (2004-2016) (2004-2016) (1) (2) (3) (4) (5) Guerrilla control 0.166 -0.663* -0.180 -0.734** -0.161** (0.218) (0.349) (0.370) (0.335) (0.0738) Observations 270 275 199 273 295 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 Dependent mean 0.000 0.000 0.000 0.000 0.118 Note: The table presents the results of estimating equation 1 for our outcomes related to political discontent and mis- trust. Column 1 shows the political participation scope that includes questions regarding whether the citizen votes, attends protests, and attends government meetings. Column 2 reports the engagement with politicians scope that measures the extent to which citizens contact state authorities and/or bureaucracies to solve issues and attend gov- ernment/political meetings. Column 3 shows the nondemocratic engagement scope that measures the extent to which citizens approve the use of alternative or violent means to engage in politics. Column 4 reports the trust in institutions item that measures the extent to which citizens trust different types of Salvadoran institutions, including the police, the powers of state, and local government. The table reports the inverse covariance-weighted average index as dependent variables. Column 5 reports the share of individuals who report believing that the members of their community are not trustworthy at all. The unit of observation in all columns is the census tract. The information was obtained from the Latin American Public Opinion Project (LAPOP) survey. Controls not shown include a linear polynomial of the dis- tance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. We also examine attitudes towards out-groups such as agricultural elites by conducting focus groups in these communities with key actors in the agricultural and private sector. One common pattern is that individuals from former guerrilla areas are more likely to distrust external actors. Many farmers in these areas associate external actors and agricultural investment with the largely repressive elites of the 1950s. As one farmer stated: “There is a lot of prejudice towards the private sector and large landowners and they are not welcomed in the area since they are not part of the community and want to change the way the community works, which is based on popular arrangements.” Peasants living in the community also supported this view. One said: “Here we all know each other and rely a lot on family networks for production.” However, as one private investor noted: “This model of family or communal agriculture did not work since these areas are poorer today.” Overall, distrust of out-groups still exists in these areas even though the self-governing institutions 30 are gone. In the next subsections, we analyze how these differences in social norms affect the durability of changes implemented during guerrilla control, such as the absence of the state and the subsistence economy. VI.A.1 Transformation of local governance: public goods provision As we noted earlier, less political engagement and less trust in institutions may complicate the provision of public goods by the state and affect demand for state-provided services. We explore the validity of these arguments in Table 5 by examining different contemporaneous outcomes that measure demand and supply of state services. First, we measure state efforts to provide public goods in these areas by analyzing the effects on public investment. Second, we evaluate the effects on effective state service supply using the total number of schools, hospitals, and state buildings per 100k inhabitants, plus road density, as outcomes of interest. Third, to measure perceived access to and utilization of public goods by citizens in former guerilla areas, we estimate equation (1) using rates of access/usage of sewerage service, potable water, electricity, and garbage collection service. These rates were estimated as the number of households with access to each public service relative to the total number of households in each census tract. (see Appendix A for details on the constructions of these measures). The estimates yield three key results. First, Column 1 indicates there is more public investment (measured as any government expenditures in social projects related to infrastructure in sectors such as electricity, water and sewerage, and education) inside former guerrilla areas. Moreover, there are more schools per 100k inhabitants and greater road density in such areas than others (see Columns 3 and 4).29 This result is consistent with qualitative evidence highlighting an increase in school investments in the post-conflict period. However, as shown previously, the larger number of schools inside former rebel territories does not translate into better educational outcomes.30 29 In Figure D.12, we also assess the yearly number of primary schools where the national exam is administered. Similarly, we find a larger number of such schools since 1999 in former guerrilla areas. 30 In Table D.14, we also show there are no significant differences in the distance of each segment to the closest ıa) and incarcerations between treated and control areas, demonstrating that the lack of local police station (comisar´ development in former guerrilla areas is not due to lack of state capacity or enforcement. 31 Table 5. Effects of Guerrilla Control on Public Goods Provision Panel A: Supply of State Services and Public Goods Public Investment Hospitals per 100k Schools per 100k Road Density Public Buildings (1995-2015) Population (2015) Population (2007) (2014) per 100k (2020) (1) (2) (3) (4) (5) Guerrilla control 0.127** -2.938 27.76*** 0.246* -7.243 (0.0614) (4.626) (10.07) (0.128) (139.1) Observations 1,068 3,668 3,668 3,681 3,275 Dependent mean 0.320 15.21 96.61 1.200 1930 Panel B: Demand for State Services 2007 - Share of Households that report having Sewerage Garbage Water Electricity Daily Water Frequency (6) (7) (8) (9) (10) Guerrilla control -0.0255 -0.0523*** -0.0392** -0.0290*** 0.000661 (0.0180) (0.0185) (0.0193) (0.00862) (0.0196) Observations 3,668 3,668 3,668 3,668 3,622 Dependent mean 0.400 0.510 0.780 0.910 0.740 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 Note: The table presents the results of estimating equation 1 for our outcomes related to public goods provision. Panel A presents results related to the supply-side provision of public goods and services. Panel B shows outcomes related to household demand for public goods and services. Column 1 shows whether the canton has received public investment for any social project (FISDL), mostly related to construction or infrastructure updates. Columns 2, 3, and 5 report the number of hospitals, schools, and public buildings per 100k inhabitants, respectively. Column 4 shows the road density in each census tract, which is measured as the length of all roads in the unit divided by their area. Columns 6–9 report the treatment effect on the share of households with any of the marked services within each census tract. Column 10 shows whether the household receives water daily or not. Information in the latter columns comes from the Population Census of 2007. Information in Columns 2 and 4 comes from Google maps. The unit of observation in Column 1 is the canton, but for the rest of the columns it is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects that represent the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Second, in terms of demand, citizens report less access to and utilization of public services in former guerrilla areas relative to other areas (Columns 6–9). Importantly, we see less demand for the exact services in which we observe an increase in public investment. Third, we also examine how access to and utilization of public services is affected by the quality of these services. Column 10 shows no differences in reported daily water frequency across the boundaries. In Table D.15, we also analyze whether the quality of education, measured by the education of teachers and the number of teachers, is lower in former guerrilla areas than in others. This might explain why there are worse education outcomes despite the larger presence of physical schools inside treated areas. However, we do not find evidence of significant differences in either of these variables; this 32 suggests the quality of education was not different across treatment and control areas. The effects of FMLN control in public goods provision are robust to different bandwidths, as shown in Figure D.13. The graphic representation of these effects is also illustrated in Figure D.14. All in all, areas with past guerrilla presence have higher investments in infrastructure. At the same time, residents of these areas also report less access to and utilization of public services relative to people in control areas. Low levels of institutional trust and political engagement may partly explain these effects: if citizens do not trust the state, they will demand fewer public goods or believe they lack adequate access. VI.A.2 Transformation of economic structures: land fragmentation, agricultural productivity, and inequality As part of their self-governance initiatives, the FMLN promoted land access for formerly land- less peasants. The group invited peasants in its areas to occupy and cultivate properties formerly owned by large landowners. In the absence of state authorities, many peasants organized coop- eratives and occupied tens of thousands of hectares of land, eventually claiming these properties under the terms of the peace agreement (Wood, 2010). By the war’s end, new patterns of land use and tenure had formed. This transfer of agrarian property rights eliminated large extension crops for export and converted the land into small plots. The peasants’ newfound land access and their wariness of political and economic elites may have prevented the scale-up of farming activities due to their refusal to sell to large landowners or allow them to enter the market. This restricted improvements in their material well-being and increased their attachment to these lands, highlighting an important eco- nomic mechanism that links the FMLN with lower development through greater fragmentation of agricultural land and lower productivity. We explore this hypothesis by studying the effects of guerrilla control on agricultural productivity and land fragmentation today. Land fragmentation Given that guerrilla-controlled areas experienced a redistribution of commercial plots to peasants from 1985 to 1992, we start by analyzing the amount of land fragmentation today in former FMLN areas. First, we consider the plot size in those areas relative to places nearby. Table 6 shows that commercial plots are much smaller inside former guerrilla areas. This is true for every type of plot we examined, whether owned or rented (see Columns 1 and 2, respectively). In addition, 33 Column 3 shows that the plot size for cultivation is also smaller for commercial producers (Panel A). In contrast, we find no differences for subsistence crops (Panel B); this offers further evidence that the land fragmentation of the Civil War period still exists today. This result is confirmed in Column 4, which shows that the share of land owned by commercial farmers is much smaller within former FMLN areas. In Table D.16 in the Appendix, we also look at the Simpson index to measure land fragmentation. Consistent with these results, we find more land fragmentation in former FMLN areas relative to nearby locations. Agricultural productivity Next, we analyze whether changes in economic production during the period of guerrilla control led to changes in agricultural productivity today. Table 7 presents the results of the spatial RD analysis for the total extension of land cropped (Panel A), the share of the land harvested (Panel B), and the actual crop yield in 2005 (Panel C). Consistent with qualitative evidence, we find that the production of export crops such as sugarcane dropped significantly. Moreover, the measures of productivity are much lower in later years in former guerrilla areas. The estimates confirm that actual crop yield was lower for all crops within controlled areas. In the Appendix, we explore differences in the occupations of employed individuals currently living in former guerrilla areas relative to other areas. Unsurprisingly (and in line with previous results), we find that individuals in these areas work disproportionately in agriculture (specifically subsistence agriculture) but less in other occupations known to create more value added. These include, for example, agro-based industry and services (See Table D.17 and Figure D.15). These occupational differences between treated and control areas are maintained even as we increase the bandwidth around the discontinuity from two to 18 km, suggesting these differences are not affected by the location or creation of urban centers close to the discontinuity (See Figure D.16). 34 Table 6. Effects of Guerrilla Control on the Size of Plots Panel A: Size of Plots by Producers Focused on Commercial Activity (Ha) Own Area Total Area Cultivated Area Share of Owned Area (2007) (2007) (2007) (2007) (1) (2) (3) (4) Guerrilla control -1.100** -1.255** -0.543** -0.0402* (0.538) (0.541) (0.231) (0.0233) Observations 2,021 2,003 2,017 1,838 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 5.135 6.799 2.529 0.352 Panel B: Size of Plots by Producers Focused on Subsistence Activities (Ha) Guerrilla control 0.00451 0.0202 0.0133 0.0183 (0.0146) (0.0160) (0.0124) (0.0276) Observations 2,309 2,298 2,292 1,677 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 0.228 0.656 0.596 0.678 Note: The table presents the results of estimating equation 1 for our outcomes related to the size of land used by producers for agricultural activity. Panel A shows the outcomes for the average plot managed by producers focused on commercial activities. Panel B does the same, but for the average plot managed by producers focused on subsistence activities. Column 1 uses as dependent variable the size of the land the producer owns. Column 2 uses the size of the total land the producer manages, which could also include rented land. Column 3 uses the area cultivated by the producer. Column 4 uses the share of the total area managed by the producer that the producer owns. Information in all columns comes from the Agricultural National Census of 2007. The unit of observation in all columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 35 Table 7. Effects of Guerrilla Control on Agricultural Productivity Panel A: Crop Production in 2005 (1,000 Tons) Subsistence crops Cash crops Bean Maize Coffee Sugarcane (1) (2) (3) (4) Guerrilla control -0.00167 -0.0110 -0.00540 -1.829*** (0.00161) (0.0324) (0.00789) (0.529) Observations 3,652 3,652 3,652 3,652 Dependent mean 0.102 1.914 0.458 15.460 Panel B: Share of Harvest in 2005 (Ha) Guerrilla control -0.0112*** -0.0310*** -0.0202*** -0.00357 (0.00356) (0.0109) (0.00753) (0.00230) Observations 3,651 3,651 3,651 3,651 Dependent mean 0.0360 0.113 0.0780 0.0260 Panel C: Actual Crop Yield in 2005 (Tons/Ha) Guerrilla control -0.00471*** -0.0161*** -0.00622** -1.078*** (0.00126) (0.00586) (0.00242) (0.241) Observations 3,566 3,550 3,649 3,649 Dependent mean 0.400 2.254 0.835 61.220 Bandwidth (Km) 2.266 2.266 2.266 2.266 Notes: The table presents the results of estimating equation (1) for outcomes related to agriculture. Panel A shows results using as dependent variable each crop’s production in 1,000 tons. Panel B uses as dependent variable the share of harvested land of each crop from the total area of each census tract. Panel C uses the actual yield of each crop, which is measured as the total production over the total of cultivated land for each crop. The unit of observation in all columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Inequality and cooperatives As shown in the previous section, guerrilla-controlled areas were more likely to experience land fragmentation and remain divided today. Yet, even though these areas are poorer, there may be less income inequality. We explore this notion by evaluating reported income from household surveys and the wealth index from the demographic census. Table D.18 presents results for different measures of inequality. We find that while individuals have less income in former guerrilla areas (Columns 1 and 2), there is no evidence of lower in- 36 equality in terms of wealth assets or income. This result is robust to different measures.31 Further- more, we also find no difference in whether farmers belong to a cooperative in a former FMLN territory relative to areas outside (see Table D.20). This result also aligns with the evidence in Table 1 showing no baseline differences in the 1980 state-implemented land reform across the bound- aries that transformed many large plots into cooperatives. Although the FMLN redistributed land to peasants, most of its agricultural programs targeted subsistence crops. They did not teach peas- ants how to grow and commercialize export crops. VI.B Ruling out differences in land tenure We previously posited that enduring distrust of elites and the state in former guerrilla areas pre- vented the entrance of large landowners and private investment in general, isolating these com- munities from improvements in productivity. Another possibility is that individuals in former guerrilla areas remained attached to the land since they did not have the right to sell it. This is unlikely since a major aspect of the Peace Agreements was to respect the land tenure that formed during guerrilla territorial control.32 Indeed, using the Agriculture Census, we do not observe differences between treatment and control areas in land ownership or land property rights.33 VI.C Ruling out migration This section explores potential differences in migration patterns in former FMLN areas compared to those outside the boundary. The guerrillas promoted changes that might have induced different patterns of worker selection. For example, high-ability workers could have migrated from these areas due to fear of expropriation of their income (out-migration). But there could have been ad- verse selection of workers if guerrilla-held areas attracted less productive peasants or individuals with more egalitarian preferences into the areas (in-migration). We explore these migration patterns empirically in Table 8 using data from the 2007 census. Columns 1–5 examine impacts on international migration. In particular, we estimate equation (1) for the share of international emigrants during the period of FMLN territorial control and af- terward, the number of years since the international emigrant left the household, and the share of households receiving remittances. Unfortunately, the 2007 census does not include questions 31 We also checked the robustness of these results using the Wealth Index from the 2007 Census at the census tract level (Table D.19). We do not find any difference across the boundary when using this measure to assess inequality. 32 As mentioned in Section II.D, the agreement established that individuals occupying land in “conflict zones” could keep it. The state was in charge of administering land to anyone who claimed ownership; it gave current occupants a land title and 30 years to pay for the land. 33 Results are available upon request. 37 related to internal migration. However, international migration is significant in El Salvador. The results suggest that residents of former guerrilla areas were not more likely to migrate abroad or to receive remittances than those in nearby locations, and that—if anything—migration abroad seems more recent. The coefficients are also negative, indicating that people were less likely to migrate internationally. These results provide evidence that former guerrilla areas did not face more “brain drain,” a result consistent with the idea that elites did not live in these locations and left their operations there when their workers stayed. Moreover, it accords with the idea that peasants supported the guerrilla movement and wanted to stay in these locations. We examine in-migration outcomes in Columns 6–9 using data from the 2007 census. To evaluate if there was more migration into rebel areas, we estimate equation (1) for the share of individuals who always lived in the same location, the share of individuals who lived in the same location as their mothers, in-migration during the Civil War period, and years since arrival. Each variable is defined in Appendix A. There is no evidence of large differences in migration patterns for areas under guerrilla control. Moreover, the coefficient estimates are positive, suggesting areas under guerrilla control were less likely to have more in-migration. Importantly, the magnitude of the estimated coefficients is small and close to zero for all these outcomes. The sign of the coefficients is consistent with the idea that communities in previous guerrilla areas are closed to individuals from outside. To further examine if there was more migration from FMLN areas by highly educated people, we examine the same outcomes in Columns 6–9 of Table D.21 using the sample of individuals who had finished at least high school by the time the conflict started. The magnitude of all the coeffi- cients in Table D.21 is close to zero and not significant, implying that migration of highly educated individuals may not underpin the effects. Moreover, the sign of the coefficients in Columns 6–8 highlights that, if anything, more in-migration of highly selected individuals occurred. 38 Table 8. Effects of Guerrilla Control on Migration Outcomes International Migrants Always Lived in Same Location People who Arrived Years since During Control At any Time Years since Households that Received Received Remittance from Same Location as the Mother During Control Arrival (Share) (Share) Departure Remittances (Share) War Migrant (Share) (Share) (Share) (Share) (1) (2) (3) (4) (5) (6) (7) (8) (9) Guerrilla control -0.00219 -0.00221 -0.341 -0.00674 -0.00194 0.00788 0.00648 -0.00452 -0.218 (0.00171) (0.00498) (0.27700) (0.00427) (0.00126) (0.00956) (0.00978) (0.00321) (0.41100) Observations 3,637 3,637 3,396 3,637 3,637 3,637 3,637 3,637 3,524 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 2.266 2.266 2.266 2.266 Dependent mean 0.0230 0.112 7.416 0.103 0.0140 0.766 0.730 0.0620 16.470 Note: The table presents the results of equation 1 for our outcomes related to migration. Columns 1–5 focus on outcomes for international migrants. Columns 6–9 focus on internal in-migration flows. All information was obtained from the Population Census of 2007. The unit of observation in all columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla- controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Finally, we also looked at temporal migration for work by analyzing whether individuals work in a different census tract or municipality from where they live across the boundaries (Table D.22). We do not find that residents of former guerrilla areas are more likely to work outside their com- munity, providing further evidence that out-migration may not drive the results. It is puzzling that these people do not migrate today given better economic opportunities nearby. It is possible that they prefer not to leave their village due to strong social ties, more “rootedness,” and because they do not trust outsiders. This idea aligns with our results in Table 4 (Column 5), where we show that individuals in former guerrilla areas are more likely to trust members of their community than they trust residents of nearby areas. Evidence has shown individuals can remain attached to their land when property rights are not defined or depend on land use, and plots are isolated from large markets (Albertus, Espinoza and Fort, 2020; De Janvry et al., 2015). Two elements suggest this was not the case in El Salvador. As we noted earlier, after the 1992 peace agreements, the state recognized property rights in guerrilla territory. Consequently, occupants of those lands could sell their plots and migrate if they wished. We also show that our results do not vary according to the distance of each individual to the road network or a main city (see Table D.23). VI.D Ruling out conflict and violence persistence This section explores whether the negative effects of guerrilla control stemmed mainly from higher conflict or the persistence of violence, which may have been more intense in areas close to the 39 boundary where control was contested. As such, conflict or violence may be the source of the negative effects in the development outcomes. We test this mechanism in several ways. First, we estimate equation (1), controlling for the seg- ments of disputed areas where the Salvadoran government and the guerrillas usually fought. Sec- ond, we use a doughnut-hole approach to exclude all observations within 80 m from the boundary of guerrilla-held territories.34 Results for the main outcomes of interest from these exercises are in Table 9. In general, the coef- ficients are negative, statistically significant, and similar in size. This suggests conflict is not the main factor behind the negative effects. We obtain similar results using outcomes of war crimes, including number of deaths, disappearances, and other conflict-related crimes as reported by the Truth Commission. Results in Table D.24 once again support the idea that areas under guerrilla control did not experience disproportionately higher crimes relative to other areas. The negative coefficient associated with the war crime estimates suggests that former guerrilla areas experi- enced less crimes, leading to lower-bound estimates of our main outcomes. Finally, we appraise the role of guerrilla control in contemporaneous measures of crime to judge whether the historical presence of guerrillas prevented the development of criminal actors such as gangs. The social capital in former guerrilla areas may have done so (Sviatschi, 2020); tightly knit communities with strong social ties can better prevent crime because they raise detection probabilities and attach shame to criminal behaviors (Buonanno, Montolio and Vanin, 2009). If social capital lasts, we expect fewer crimes linked to non-state armed actors, which are pervasive in El Salvador. But if our results stemmed from violence during or after guerrilla control of these areas, we should expect more violence today. To test these hypotheses, we considered homicide rates during 2017 using police data, and victimization rates from 2004 to 2016 in LAPOP surveys. Table D.25 in the Appendix presents the results. Consistent with the finding that violence during the conflict was not greater in guerrilla-controlled areas, the results largely suggest no differences in homicide rates between areas once under and outside FMLN control.35 If anything, the esti- mates are negative, which suggests that the documented differences in long-term development did not arise from increases in conflict or violence. Additionally, there is evidence that residents of 34 We excluded all segments within an 80 m distance to shut off almost every segment close to the boundary and inside the guerrilla zone with an immediate neighbor outside it. 35 Figures D.17 and D.18 confirm this result for different bandwidths. 40 areas once under guerrilla control are less likely to be victims of violent crime or extortion related to gang activity. This aligns with enduring norms of cooperation and higher levels of social capi- tal as well as with qualitative evidence gathered from interviews with locals and former guerrilla commanders, who repeatedly expressed thoughts such as: “The fact that the maras (gangs) are barely present in these areas reflects that the self-organization of the people worked.” (Joaqu´ ın Villalobos, FMLN Military Commander, interview conducted on March 23, 2022). VI.E Other mechanisms Other potential mechanisms such as child recruitment, disproportionate improvements in control areas, or changes in the supply and quality of education could underlie our results. Peace agreement and post-conflict policies— As we noted in the background section, all reforms of the peace agreement (e.g., the reform of the Judicial System) were implemented at the national level and may not prompt our effects. For example, although the agreement created a national police force, we find no differences in enforcement or number of state institutions across areas. As we have shown, the results do not arise from differences in post-land redistribution that was also part of the agreement. Finally, the post-conflict political environment does not seem to explain the results. For example, the ARENA government elected right after the agreement did not lower public investment to punish guerrilla areas. Although we cannot assess the voting patterns of these areas immediately after the war, evidence from the 2014 and 2015 elections do not show that these areas favored a specific party. If anything, there was a small and negative effect on the vote share for the leftist party, as shown in Table D.12. We find that effects on development were still negative when the FMLN won elections, and when former guerrilla areas received more investment (not less) related to infrastructure reconstruction efforts such as roads and schools. 41 Table 9. Effects of Guerrilla Territorial Control on Main Outcomes, Controlling for Conflict Panel A: Separating Disputed areas from Government controlled areas Night Light Arcsine Wealth Score Years of Education (2013) (2007) (2007) (1) (2) (3) Guerrilla control -0.127*** -0.170*** -0.438** (0.0314) (0.0587) (0.188) Disputed area 0.0851* -0.0670 -0.230 (0.0473) (0.0678) (0.207) Observations 3,652 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 Dependent mean 3.536 -0.0160 6.573 Panel B: Doughnut Hole Analysis (400 m) Guerrilla control -0.164*** -0.123** -0.308** (0.0520) (0.0541) (0.153) Observations 1,564 1,555 1,555 Bandwidth (Km) 0.981 0.981 0.981 Dependent mean 3.245 -0.197 5.845 Note: The table presents results for the main outcomes but under different specifications that help discard the hypothesis that effects were driven by conflict. Panel A shows results when separating the control group between government- controlled areas and areas disputed by guerrillas. Notice that in Panel A, the omitted category concerns segments under pure governmental dominance. Panel B shows results using a doughnut-hole methodology with a hole of 400 m. The unit of observation in all columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Recruitment— Coercive recruitment has figured prominently in work that ties civil conflict to lower levels of education. Early military experience is a bad substitute for education and labor market experience, and child soldiers lose key formative years of schooling (Blattman and Miguel, 2010). Nonetheless, coercive recruitment is unlikely to be the force behind our results. First, child soldiers were not prominent in the FMLN. Estimates suggest that of the 9,000–12,000 FMLN members, only 2,000 (about 20 percent) were under age 18, while the percentage of underage combatants in 42 the Salvadoran Army was 80 percent (48,000 of 60,000 combatants) (Courtney, 2010). Likewise, most historical studies conclude that FMLN recruitment was mostly voluntary. A study UNICEF conducted shows that while 91.7 percent of FMLN recruits had joined voluntarily, close to 53 percent of underage Salvadoran Army soldiers were forcibly recruited (Courtney, 2010). Post-conflict investment in guerrilla areas— In terms of private investment, we do not find qualita- tive evidence that large agricultural producers or firms are less willing to invest in these locations. What’s more, several counterarguments suggest that lower private investment does not underpin our main results. First, control group areas are extremely close (just 2 km away); thus, it is difficult to argue why private investment would target nearby control areas but not guerrilla areas, since the guerrillas were gone by the post-war period. Second, it is highly unlikely that contemporane- ous investors know of the boundary with former guerrilla areas in the 1980s. As mentioned above, these boundaries do not overlap with the country’s administrative division or its road network. In other words, they are not salient. Evidence from our focus groups supports this argument since few interviewees knew the exact boundaries of these areas. (If anything, they knew which munic- ipalities were affected by guerrillas but not the exact boundaries). Third, there are no differences between treatment and control areas in the number of new businesses such as restaurants, malls, and markets across the boundaries.36 Finally, we do not find evidence that fear of expropriation in former guerrilla areas explains the results. We observe that residents of these areas are as likely as those in the control group to believe that it is acceptable to invade private property or engage in antidemocratic behavior (See Table D.26). Spillovers in non-guerrilla areas during territorial control and in the post-war period— We also rule out that effects stem from counter-insurgency in nearby control areas during the period of guerrilla control. It’s possible that nearby control areas benefited from the lack of development or the agri- cultural focus in guerrilla areas. In this case, effects would be concentrated close to the boundary or just in the rd sample. First, evidence from the doughnut-hole analysis shows this is not so: ef- fects are robust to excluding observations close to the boundary. Second, if we increase the sample beyond the bandwidth to 17 km, we see homogeneous effects on development.37 36 We checked for this possibility by digitizing the number of commercial establishments 2 km around the boundary of past guerrilla presence using Google Maps. We could not distinguish any significant differences between treated and control areas in the number of commercial establishments. The estimates are available upon request. 37 One potential concern (as with many rd designs) is that we observe cross-sectional differences today in develop- ment. Due to the lack of panel data, we cannot disentangle how much the effects originate in improvements in the control group and deterioration in former guerrilla areas over time versus just improvements in the control group and no changes in former guerrilla areas. Hence, we analyze heterogeneous effects based on distance to a main road, dis- 43 VII DISCUSSION This paper explores the long-term development impacts of guerrilla territorial control in El Sal- vador. We employ a spatial regression discontinuity that compares night light luminosity, wealth, and human capital in areas close to the boundary of FMLN control. Our results suggest that guerrilla control had sizable negative and enduring consequences for night light luminosity, wealth, and human capital. These effects are not completely accounted for by sorting from treated areas, out-migration, higher conflict intensity, child recruitment by rebels, lower public and private investment, or differences in education quality. We believe our main results arise from the guerrilla-directed transformation of local governance structures and social capital, which led to lasting changes in economic structures and relations with the state. In fact, areas once occupied by guerrillas today feature higher political discontent and institutional distrust, smaller land holdings, and lower agricultural productivity. Moreover, residents of these areas have largely abandoned agricultural activities related to commercial exploitation in favor of subsistence ones. These findings offer key insights for other countries that experience internal armed conflict or post-conflict development but which do not have quality data to conduct similar analyses. In particular, these results suggest that areas under rebel governance may embark on a negative development path that is likely to continue. We believe that an increase in public investment itself is not enough to guarantee recovery in such areas, and that trust in the state and private sector must resume to foster productive growth in the long term. This paper also provides evidence of a mismatch between historic cultural norms (such as trust) in the state and actions that may best suit the current environment. Before the Salvadoran con- flict and the guerrilla seizure of territory, the state and associated elites had repressed the rural peasantry for more than 100 years. This created the distrust later reinforced in guerrilla areas. It persists today long after the disappearance of those actors and in spite of the fact that commu- nities might benefit from engaging with the new government. Although extreme distrust of the state may have been an optimal response during the conflict, this may not be true today. tance to a city, and population density. If effects emerged only via improvements in control areas but no changes in the treated group, we would expect a mitigation of the negative effects in better-connected regions. Table D.23 shows this is not so. These results imply that even areas that were more developed before guerrilla control are equally affected by their historical presence today. 44 References Acemoglu, Daron and James A Robinson. 2012. Why nations fail: The origins of power, prosperity, and poverty. Currency. Acemoglu, Daron, Simon Johnson and James A Robinson. 2001. “The colonial origins of compar- ative development: An empirical investigation.” American economic review 91(5):1369–1401. Acemoglu, Daron, Suresh Naidu, Pascual Restrepo and James A Robinson. 2019. “Democracy does cause growth.” Journal of political economy 127(1):47–100. Albertus, Michael, Mauricio Espinoza and Ricardo Fort. 2020. “Land reform and human capital development: Evidence from Peru.” Journal of Development Economics 147:102540. Almond, Gabriel Abraham and Sidney Verba. 2015. The civic culture: Political attitudes and democracy in five nations. Princeton university press. Alsan, Marcella and Marianne Wanamaker. 2018. “Tuskegee and the health of black men.” The quarterly journal of economics 133(1):407–455. ´ Alvarez, ın. 2011. “De guerrilla a partido pol´ Alberto Mart´ ı para ıtico: el Frente Farabundo Mart´ ´ Nacional (FMLN).” Historia y pol´ la Liberacion ıtica: Ideas, procesos y movimientos sociales (25):207–233. ´ Alvarez, ın. 2013. Alberto Mart´ “Sociedad civil y movimientos sociales en El Salvador de postguerra.” Historia Actual Online (32):59–71. Arjona, Ana. 2016. Rebelocracy. Cambridge University Press. Arjona, Ana, Nelson Kasfir and Zachariah Mampilly. 2015. Rebel governance in civil war. Cam- bridge University Press. Banfield, Edward C. 1967. “The moral basis of a backward society.”. a, Joseph Henrich, Edward Miguel and Tamar Bauer, Michal, Christopher Blattman, Julie Chytilov´ Mitts. 2016. “Can war foster cooperation?” Journal of Economic Perspectives 30(3):249–74. Binford, Leigh. 1997. “Grassroots development in conflict zones of northeastern El Salvador.” Latin American Perspectives 24(2):56–79. 45 Blattman, Christopher and Edward Miguel. 2010. “Civil war.” Journal of Economic literature 48(1):3–57. Boyce, James K. 1995. “Adjustment toward peace: An introduction.” World development 23(12):2067–2077. Breslawski, Jori. 2021. “The Social terrain of rebel held territory.” Journal of Conflict Resolution 65(2-3):453–479. Buntaine, Mark T, Daniel L Nielson and Jacob T Skaggs. 2021. “Escaping the disengagement dilemma: Two field experiments on motivating citizens to report on public services.” British Journal of Political Science 51(2):685–705. Buonanno, Paolo, Daniel Montolio and Paolo Vanin. 2009. “Does social capital reduce crime?” The journal of law and economics 52(1):145–170. Calonico, Sebastian, Matias D Cattaneo, Max H Farrell and Rocio Titiunik. 2019. “Regression discontinuity designs using covariates.” Review of Economics and Statistics 101(3):442–451. Calonico, Sebastian, Matias D Cattaneo and Rocio Titiunik. 2014. “Robust nonparametric confi- dence intervals for regression-discontinuity designs.” Econometrica 82(6):2295–2326. ˜ Castaneda, Eduardo Sancho. 2016. “El control social y territorial de una fuerza insurgente. El caso de la Resistencia Nacional (RN) durante el conflicto armado.” Revista Polic´ ıa y Seguridad ´ Publica pp. 191–226. avez, Joaquın. 2011. Ch´ “Revolutionary Power, Divided State.” Mapping Latin America: A Cartographic Reader pp. 250–253. Coleman, James. 1990. “Social capital. Foundations of social theory.” The Cambridge: Belknap Press of Harvard University . Colindres, E. 1976. “La tenencia de la tierra en El Salvador.;[Land tenure in El Salvador].” Estudios Centroamericanos (El Salvador)..(Set 31(335-336):463–472. Collier, Paul. 2008. The bottom billion: Why the poorest countries are failing and what can be done about it. Oxford University Press, USA. Conley, T.G. 1999. “GMM estimation with cross sectional dependence.” Journal of Econometrics 46 92(1):1–45. URL: https://www.sciencedirect.com/science/article/pii/S0304407698000840 Courtney, Jocelyn. 2010. “The Civil War That Was Fought by Children: Understanding the Role of Child Combatants in El Salvador’s Civil War, 1980-1992.” Journal of Military History 74(2). ¨ De Groot, Olaf J, Carlos Bozzoli, Anousheh Alamir and Tilman Bruck. 2021. “The global economic burden of violent conflict.” Journal of Peace Research . De Janvry, Alain, Kyle Emerick, Marco Gonzalez-Navarro and Elisabeth Sadoulet. 2015. “Delink- ing land rights from land use: Certification and migration in Mexico.” American Economic Review 105(10):3125–49. Dell, Melissa. 2010. “The persistent effects of Peru’s mining mita.” Econometrica 78(6):1863–1903. Dell, Melissa and Benjamin A Olken. 2020. “The development effects of the extractive colonial economy: The dutch cultivation system in java.” The Review of Economic Studies 87(1):164– 203. Dell, Melissa, Nathan Lane and Pablo Querubin. 2018. “The historical state, local collective action, and economic development in Vietnam.” Econometrica 86(6):2083–2121. Fergusson, Leopoldo, Ana Mar´ ıa Ib´ ˜ and Juan Felipe Riano. 2020. “Conflict, educational at- anez tainment, and structural transformation: La violencia in colombia.” Economic Development and Cultural Change 69(1):335–371. FMLN. 1983. “Sobre el Desarrollo del FMLN.” Comandancia General del FMLN . FMLN. 1984. “Situacion Revolucionaria y Escalada Intervencionista en la Guerra Salvadorena.”. Text digitized by Centro de Estudios Marxistas ”Sarbelio Navarrete” (CEM) and Servicio Infor- mativo Ecumenico y Popular (SIEP). Foster, Andrew D and Mark R Rosenzweig. 2022. “Are There Too Many Farms in the World? Labor Market Transaction Costs, Machine Capacities, and Optimal Farm Size.” Journal of Political Economy 130(3):636–680. Grasse, Donald, Renard Sexton and Austin Wright. 2021. “The Logic and Impacts of Rebel Public Services Provision: Evidence from Taliban Courts in Afghanistan.”. 47 Green, Amelia Hoover and Patrick Ball. 2019. “Civilian killings and disappearances during civil war in El Salvador (1980–1992).” Demographic Research 41:781–814. Henderson, J Vernon, Adam Storeygard and David N Weil. 2012. “Measuring economic growth from outer space.” American economic review 102(2):994–1028. Hijmans, Robert J, Susan E Cameron, Juan L Parra, Peter G Jones and Andy Jarvis. 2005. “Very high resolution interpolated climate surfaces for global land areas.” International Journal of Climatology: A Journal of the Royal Meteorological Society 25(15):1965–1978. Jablonski, Ryan S and Brigitte Seim. 2022. “What politicians don’t know can hurt you: The effects of information on politicians’ spending decisions.” Available at SSRN . Lauria-Santiago, Aldo A and Jeffrey L Gould. 2008. To Rise in Darkness. Duke University Press. Leon, Gianmarco. 2012. “Civil conflict and human capital accumulation the long-term effects of ´ Journal of Human Resources 47(4):991–1022. political violence in peru.” Liu, Shelley X. 2022. “Control, Coercion, and Cooptation: How Rebels Govern after Winning Civil War.” World Politics 74(1):37–76. Lowes, Sara and Eduardo Montero. 2021. “The legacy of colonial medicine in Central Africa.” American Economic Review 111(4):1284–1314. Loyle, Cyanne E, Kathleen Gallagher Cunningham, Reyko Huang and Danielle F Jung. 2021. “New Directions in Rebel Governance Research.” Perspectives on Politics pp. 1–13. Mao, Zedong. 1966. Selected military writings of Mao Tse-tung. Foreign Languages Press. ın Alvarez, Alberto. 2010. Mart´ From Revolutionary War to Democratic Revolution: The ı National Liberation Front (FMLN) in El Salvador. Germany. Farabundo Mart´ Mart´ ´ Ignacio. 1990. “La violencia en Centroam´ ın-Baro, ´ psicosocial.” Salud, erica: una vision Problema y Debate (Buenos Aires) 2:53–66. Martinez-Bravo, Monica and Andreas Stegmann. 2022. “In vaccines we trust? The effects of the CIA’s vaccine ruse on immunization in Pakistan.” Journal of the European Economic Association 20(1):150–186. Michalopoulos, Stelios and Elias Papaioannou. 2013. “Pre-colonial ethnic institutions and contem- porary African development.” Econometrica 81(1):113–152. 48 erard Roland. 2011. “The long-run impact of bombing Vietnam.” Journal Miguel, Edward and G´ of Development Economics 96(1):1–15. URL: https://www.sciencedirect.com/science/article/pii/S0304387810000817 MINED. 2009. “Historia de El Salvador.” Ministerio de Educacion de El Salvador . Mishler, William and Richard Rose. 2001. “What are the origins of political trust? Testing in- stitutional and cultural theories in post-communist societies.” Comparative political studies 34(1):30–62. Moodie, Ellen. 2011. El Salvador in the aftermath of peace. In El Salvador in the Aftermath of Peace. University of Pennsylvania Press. Moran, Rolando. 1985. Un trabajo de masas para ganar la guerra. In Pueblos en Armas, ed. Marta Harnecker. Punto Final. Nunn, Nathan. 2020. “The Historical Roots of Economic Development.” Science 367(6485):eaaz9986. Nunn, Nathan. 2022. “On the Dynamics of Human Behavior: The Past, Present, and Future of Culture, Conflict, and Cooperation.” American Economic Review Papers and Proceedings . Pearce, Jenny. 1986. “Promised Land.” Peasant Rebellion in Chalatenango, El Salvador . Pence, Karen M. 2006. “The role of wealth transformations: An application to estimating the effect of tax incentives on saving.” Contributions in Economic Analysis & Policy 5(1). Pettersson, Ther´ ¨ ese and Magnus Oberg. 2020. “Organized violence, 1989–2019.” Journal of peace research 57(4):597–613. Pinkovskiy, Maxim and Xavier Sala-i Martin. 2016. “Lights, camera. . . income! Illuminating the national accounts-household surveys debate.” The Quarterly Journal of Economics 131(2):579– 631. Rettberg, Angelika. 2007. “The private sector and peace in El Salvador, Guatemala, and Colom- bia.” Journal of Latin American Studies 39(3):463–494. ˜ Juan Felipe and Felipe Valencia Caicedo. 2020. “Collateral damage: The legacy of the secret Riano, war in Laos.”. 49 Riley, Shawn J, Stephen D DeGloria and Robert Elliot. 1999. “Index that quantifies topographic heterogeneity.” intermountain Journal of sciences 5(1-4):23–27. Ritter, Paul. 1987. “A vector-based slope and aspect generation algorithm.” Photogrammetric Engineering and Remote Sensing 53(8):1109–1111. S´ ´ 2020. “On the origins of the state: Stationary bandits and taxation in anchez De La Sierra, Raul. eastern congo.” Journal of Political Economy 128(1):000–000. Scott, James C. 1972. “Patron-client politics and political change in Southeast Asia.” American political science review 66(1):91–113. Scott, James C. 2010. The art of not being governed: An anarchist history of upland Southeast Asia. Nus Press. ´ ´ economica Sevilla, Manuel. 1985. La concentracion en El Salvador. Vol. 3 Instituto de Investiga- ´ ciones economicas y sociales. Stewart, Megan A. 2018. “Civil war as state-making: Strategic governance in civil war.” International Organization 72(1):205–226. ıa Micaela. 2020. Spreading Gangs: Exporting US Criminal Capital to El Salvador. Sviatschi, Mar´ Technical report Mimeo. United Nations. 1992. Cahpultepec Peace Agreements. Technical report Mimeo. Velado, Oscar Morales. 1993. “Organizaciones no gubernamentales y sociedad civil en El Sal- vador.” Realidad: Revista de Ciencias Sociales y Humanidades (31):83–105. Wantchekon, Leonard. 1999. “Strategic voting in conditions of political instability: The 1994 elec- tions in El Salvador.” Comparative Political Studies 32(7):810–834. Williams, Michael C. 1998. “Keeping the Peace: Multidimensional UN Operations in Cambodia and El Salvador.”. Wood, Elisabeth. 2003. Insurgent collective action and civil war in El Salvador. Cambridge Uni- versity Press. Wood, Elisabeth. 2010. “Agrarian Reform, Land Occupation, and the Transition to Democracy in El Salvador.” Distributive Justice Transitions pp. 141–76. 50 Wood, Elisabeth Jean. 2008. “The social processes of civil war: The wartime transformation of social networks.” Annu. Rev. Polit. Sci. 11:539–561. 51 Contents (Appendix) A Data Sources and Variable Definitions 53 A.A Guerrilla territories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 A.B Geospatial variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 A.C Population and Household Census of 2007 (PHC) . . . . . . . . . . . . . . . . . . . . . . . 55 A.D Presidential election results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 B Descriptive Statistics 58 C Maps 59 D Robustness Tests 60 D.A Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60 D.B Main Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 D.C Public Goods Provision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 E Qualitative Study 95 E.A Sample definition and recruitment of participants . . . . . . . . . . . . . . . . . . . . 95 E.B Instruments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 E.C Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 E.D Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 52 A Data Sources and Variable Definitions A.A Guerrilla territories ˜ • Territories under control by 1991: Following Castaneda (2016), this study uses the maps that document FMLN-held areas as submitted to the United Nations and approved by the dif- ferent political parties in El Salvador during the cease-fire process. Since the map originally had an image format, we used ArcMap to digitize it by hand and convert it to a shapefile format. Thus, this is the only part of the spatial analysis that is not coded. A.B Geospatial variables • Night light luminosity: Data on night light luminosity comes from the Defense Meteo- rological Satellite Program Operational Linescan System (DMSP-OLS). This data was ob- tained from the US National Oceanic and Atmospheric Administration (NOAA) at https: //ngdc.noaa.gov/eog/download.html. This data has a resolution of 30 arc seconds (1 km2 ) and spans 1992 to 2013. The challenge with night light luminosity data is the signifi- cant fraction of observations that take the value of zero and the existence of extreme values in the right tail of the distribution (Michalopoulos and Papaioannou, 2013; Pinkovskiy and Sala-i Martin, 2016). To account for this potential concern, we adjust the outcome of interest using the logarithm and the inverse hyperbolic sine transformation.38 • Elevation: Elevation was obtained from the Google Earth Engine Data Catalog and is avail- able at https://developers.google.com/earth-engine/datasets/catalog/USGS_ SRTMGL1_003. This data provides elevation information in meters at the 3 arc-seconds spa- tial resolution (90 mts2 ). The digital elevation model (DEM) was created based on the images of the Shuttle Radar Topography Mission (SRTM) of NASA. In this study, we calculated the average elevation for each census tract. • Slope: this study uses the terrain() function in R to compute the slope from the elevation data accordingly with Ritter (1987).39 The algorithm uses four neighboring pixels to compute each pixel’s slope in degrees. Thus, higher values represent steeper terrain. Our study uses the average of the slope at the census tract level. 38 2 The inverse hyperbolic sine transformation is defined as log (yi +(yi +1)1/2 ) and can be interpreted as a logarithmic dependent variable (Pence, 2006). 39 Documentation of the R tool can be found at https://www.rdocumentation.org/packages/raster/ versions/3.4-10/topics/terrain 53 • Ruggedness: This study implements the terrain ruggedness index of Riley, DeGloria and El- liot (1999) using the tri() function in R.40 The algorithm uses five neighboring pixels to calcu- late each pixel’s index from the elevation data. Our study uses the average of the ruggedness index at the census tract level. • Rivers and lakes: Information on surface water bodies comes from the Google Earth Engine Data Catalog and is available at https://developers.google.com/earth-engine/ datasets/catalog/MERIT_Hydro_v1_0_1. The data comes from the MERIT Hydro dataset with a 3 arc-seconds spatial resolution (90 mts2 ). Our variables take the value of one if a river or lake passes by a census tract. • Precipitation: Precipitation was obtained from the Global Climate Database created by Hij- mans et al. (2005) that is available at http://www.worldclim.org/. This data provides a historic time series of rainfall in millimeters from 1960 to 2018 at the 2.5 minutes spatial resolution (21 km2 ) with a monthly periodicity. This study standardizes the series from 1960 to 1979 and calculates the standardized average of rainfall for each census tract from 1975 to 1979. • Temperature: Maximum temperature was obtained at the Global Climate Database created by Hijmans et al. (2005) and is available at http://www.worldclim.org/. This data provides a historic time series of temperature in Celsius from 1960 to 2018 with a monthly periodicity at the 2.5 minutes spatial resolution (21 km2 ) with a monthly periodicity. This study standardizes the series from 1960 to 1979 and calculates the standardized average of temperature for each census tract from 1975 to 1979. • Historical crop yield: Agro-climatic yield rasters were obtained from the Global Agro- Ecological Zones version 3.0 (GAEZ v 3.0) project and are available at https://www.gaez. iiasa.ac.at. The data has a spatial resolution of 5 arc-minutes (9 km2 ) and a yearly peri- odicity. We used the 30-year average starting in 1961 of the most relevant crops in terms of consumption and exports for 1990 (i.e., coffee, cotton, rice, beans, and sugarcane). • Roads and railways in 1980: the map outlining the road and railway network in 1980 for El Salvador was obtained from the United States Library of Congress and is available at https://www.loc.gov/resource/g4840.ct000627/. This map was made by the 40 Documentation of the R tool can be found at https://www.rdocumentation.org/packages/spatialEco/ versions/1.3-7/topics/tri 54 Central Intelligence Agency. Since the map originally had an image format, we used Ar- cMap to digitize it by hand and convert it to a shapefile format. Our variable takes the value of one if a census tract contains part of a road or railway. • Distance to the capital: We calculated the Euclidean distance in kilometers from the centroid of each census tract to San Salvador, the capital city of El Salvador. • Distance to the coast: We calculated the Euclidean distance in kilometers from the centroid of each census tract to the nearest coast. • Distance to departmental boundaries: We calculated the Euclidean distance in kilometers from the centroid of each census tract to the nearest departmental boundary. A.C Population and Household Census of 2007 (PHC) The PHC of 2007 is available at http://www.censos.gob.sv/censo/Default.aspx. • Census cartography: DIGESTYC also provided maps of the 12,435 census tracts (segmentos censales) in the 2007 census. Each census tract represents a small area with a fixed geographic perimeter. On average, they have an area of 1.7 km2 , a perimeter of 5.5 km, 131 households, and 473 individuals. • Wealth score: we built a wealth score that represents the living conditions of each household using household characteristics and asset ownership such as the type of roof, access to water, television, etc. To construct the score, we used a principal component analysis following the steps recommended by the Demographic and Health Surveys program (DHS), which can be consulted at https://dhsprogram.com/topics/wealth-index/Wealth-Index-Construction. cfm. We calculate the average of this measure for each census tract. • Years of education: The PHC asks each individual the total number of years of education in single years. However, our variable only takes into account individuals older than 18 years since most of this population already finished secondary school. We calculate the average of this variable for each census tract. • Literacy rate: The PHC asks each individual if they can read and write. Thus, our literacy rate variable is the number of individuals older than 18 years who can read in each tract over the total population in the same age range in the same tract. • Public good provision rates: The PHC asks each household if they have water access, sew- 55 erage, electricity, and garbage services. Our rates are calculated as the total number of house- holds who report having the service in each tract over the total households in the same tract. • Total number of hospitals: The Ministry of Health of El Salvador provided us with the location of all hospitals in El Salvador in 2015. The variable we use is the total number of hospitals in each census tract. • Total number of schools: The Ministry of Health of El Salvador provided us with the loca- tion of all schools in El Salvador in 2007. The variable we use is the total number of schools in each census tract. • Economically active population: Our variable is calculated at the segment level and is the sum of all people 16 years or older who are working or in search of work in the census tract over the people in the same age range in the same tract. • Working population: Our variable is calculated as the total individuals who worked last week at least one hour, no matter the occupation, in a given census tract. This variable is normalized by the total population aged 16 years or older. • Salaried population: Our variable is calculated as the total individuals in a given census tract who worked last week and received any sort of compensation for it. This variable is normalized by the total population aged 16 years or older. • Public workers: Our variable is calculated as the total individuals in a given census tract who worked last week in the public sector. This variable is normalized by the total popula- tion aged 16 years or older. • Independent workers: Our variable is calculated as the total individuals in a given census tract who worked last week as independent workers. This variable is normalized by the total population aged 16 years or older. • Total of employers: Our variable is calculated as the total individuals in a given census tract who employed at least one person for his or her own business. This variable is normalized by the total population aged 16 years or older. • Weekly worked hours: These are the average hours the working population worked last week in a given census tract. 56 • International migrants: This is the total number of people who are reported by their house- holds to be outside El Salvador in 2007 for each census tract. • International migrants in the war period: This is the total number of people who left El Sal- vador between 1979 and 1990 and are reported by their households to be outside El Salvador in 2007 for each census tract. • Remittances rate: This is the share of households in a given census tract that report receiving monetary help from a member outside El Salvador in 2007. • In-migration during the war period: This is the total number of individuals who reported in 2007 that they arrived in a given census tract between 1979 and 1990. • Moving population: This is calculated as the number of people in a given census tract who reported in 2007 any relocation in their entire life. • Moving population share: This is calculated as the moving population in each census tract over the total population in the same tract. A.D Presidential election results All data related to elections was provided by the Tribunal Supremo Electoral of El Salvador, which included the list of results and coordinates for each polling station. • Left voting share: This is calculated as the total votes for the FMLN party over the total valid votes for each polling station in El Salvador. • Right voting share: This is calculated as the total votes for the ARENA party over the total valid votes for each polling station in El Salvador. • Blank voting share: This is calculated as the total blank votes over the total valid votes for each polling station in El Salvador. • Turnout share: This is calculated as the total valid votes over the total number of people registered to vote in each polling station in El Salvador. 57 B Descriptive Statistics Table B.1. Summary Statistics of the Variables Used in the Estimation Mean SD Min Max Obs Panel A: Cease-fire map of 1991 Segment under guerrilla control 0.167 0.373 0.000 1.000 12,435 Distance to nearest controlled area -8.647 12.243 -67.010 17.955 12,435 Panel B: Geographic characteristics Night light density (2013) 25.560 23.436 0.000 62.258 12,432 Arcsine(Night light) 3.181 1.471 0.000 4.825 12,432 Log(Night light) 2.419 1.609 -12.084 4.131 12,432 Night light (Weighted by surface area) 25.560 23.436 0.000 62.258 12,432 Altitude (DEM) 496.403 287.500 0.000 2,185.623 12,433 Slope 7.458 5.153 0.000 30.127 12,432 Ruggedness 10.916 8.274 0.000 184.795 12,432 Hydrography 0.282 0.450 0.000 1.000 12,435 Coffee Yield (1961-1990) 1.606 0.203 0.339 2.982 12,427 Cotton Yield (1961-1990) 0.742 0.086 0.000 1.006 12,427 Dry Rice Yield (1961-1990) 5.231 0.514 0.000 5.615 12,427 Wet Rice Yield (1961-1990) 8.805 0.905 0.000 9.381 12,427 Bean Yield (1961-1990) 4.097 0.180 2.674 4.470 12,427 Sugarcane Yield (1961-1990) 6.257 1.075 0.000 8.884 12,427 Monthly Mean Rainfall (1975-1979) -0.067 -0.032 12,301 Monthly Minimum Temperature (1975-1979) 0.121 0.035 0.068 0.218 12,301 Monthly Maximum Temperature (1975-1979) 0.207 0.055 0.119 0.413 12,301 Roads and Railway (1980) 0.376 0.484 0.000 1.000 12,435 Distance to Coast 32.183 18.688 0.000 101.246 12,435 Distance to Capital 50.165 40.581 0.000 175.078 12,435 Panel C: Socioeconomic characteristics (2007 census) Wealth Index -0.168 0.890 -2.336 1.723 12,393 Sewerage Service Rate 0.352 0.428 0.000 1.000 12,406 Water Access Rate 0.744 0.323 0.000 1.000 12,406 Electricity Rate 0.865 0.186 0.000 1.000 12,406 Garbage Rate 0.439 0.441 0.000 1.000 12,406 Hospitals per 100k Population 15.637 62.482 0.000 787.402 12,406 Schools per 100k Population 112.544 219.692 0.000 11,111.111 12,406 Total Population 463.011 137.741 2.000 3,462.000 12,406 Female Head Rate 0.342 0.095 0.000 1.000 12,406 Gender Rate 0.474 0.032 0.316 1.000 12,406 Average Age 27.508 3.510 14.600 52.143 12,406 Fertility Rate 0.663 0.060 0.000 1.000 12,405 Years of Education 6.098 2.759 0.000 15.272 12,406 Literacy Rate 0.787 0.139 0.000 1.000 12,406 Attended School Rate 0.769 0.140 0.000 1.000 12,406 International Migrants 22.310 21.781 1.000 182.000 11,725 Total War Migrants 4.387 6.522 0.000 103.000 11,725 Migrants’ Gender Rate 0.630 0.197 0.000 1.000 11,708 Remittances Rate 0.105 0.091 0.000 1.000 12,406 In-migration at War Period 0.055 0.062 0.000 0.575 12,406 Moving Population 364.771 141.797 0.000 3,440.000 12,406 Moving Population Share 0.784 0.185 0.000 1.000 12,406 Economically Active Population 0.528 0.142 0.000 1.000 12,406 Working Population 0.477 0.135 0.000 1.000 12,406 Salaried Population 0.439 0.150 0.000 1.000 12,406 Weekly Worked Hours 44.581 6.045 8.111 85.139 12,403 Public Worker 0.038 0.036 0.000 0.333 12,406 Private Worker 0.214 0.122 0.000 1.000 12,406 Employer 0.015 0.024 0.000 0.364 12,406 Independent Worker 0.118 0.074 0.000 0.623 12,406 Notes: Summary statistics of most raw variables used in the analysis. 58 C Maps Figure C.1. Mapping of Altitude, Main Rivers, and Guerrilla-Controlled Territories Notes: The figure maps the guerrilla-controlled areas, main rivers, and the variation in altitude for El Salvador. The latter is at a resolution of three arc-seconds and based on the DEM model of NASA’s SRTM. 59 D Robustness Tests D.A Empirical Strategy Figure D.1. Plots of Smoothness around the Discontinuity (a) Altitude (b) Slope of Altitude (c) Terrain Ruggedness 20 1 .5 10 0 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 -.5 -10 -1 -1 -20 -1.5 -30 -2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (d) Hydrographic Bodies (e) Roads-Railway Infrastructure (f) Had a City (1945) .1 .1 .1 .05 .05 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 0 -.05 -.05 -.1 -.1 -.1 -.15 -.15 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (g) Distance to City (1945) (h) Distance to Comms (1945) (i) Comms Density (1945) .4 .4 .4 Estimate magnitude Estimate magnitude Estimate magnitude .2 .2 .2 0 0 0 -.2 -.2 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 60 Figure D.2. Plots of Smoothness around the Discontinuity (cont’d) (a) Part of Land Reform (1980) (b) Inside Cultivated Area (1980) (c) Had a Parish (1979) .05 .04 .05 Estimate magnitude Estimate magnitude Estimate magnitude 0 .02 0 -.05 -.05 0 -.1 -.02 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (d) Distance to Parish (1980) (e) Distance to School (1980) (f) Total Population .8 .8 30 .6 .6 20 Estimate magnitude Estimate magnitude Estimate magnitude .4 .4 10 .2 .2 0 0 0 -10 -.2 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (g) Population Density (h) Years of Education (i) Birth Rate 600 1 .2 400 .1 Estimate magnitude Estimate magnitude Estimate magnitude .5 200 0 0 0 -.1 -200 -400 -.5 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 61 Figure D.3. Plots of Smoothness around the Discontinuity (cont’d) (a) In-migration Share (b) Out-migration Share (c) Inside Highly Populated Area (1980) .04 .01 .1 .02 .05 .005 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 0 -.02 -.05 -.04 -.005 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (d) Aggregate Yield Index (1961–79) (e) Bean Agro-climatic Yield (f) Coffee Agro-climatic Yield .05 .01 .01 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 0 -.01 -.05 -.01 -.02 -.02 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (g) Cotton Agro-climatic Yield (h) Maize Agro-climatic Yield (i) Wet Rice Agro-climatic Yield .05 .04 .002 .02 .001 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 0 -.02 -.001 -.05 -.002 -.04 -.003 -.06 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 62 Figure D.4. Plots of Smoothness around the Discontinuity (cont’d) (a) Bean High Suitability (b) Coffee High Suitability (c) Maize High Suitability .04 .05 .02 .02 .01 Estimate magnitude Estimate magnitude Estimate magnitude 0 0 0 -.05 -.02 -.01 -.1 -.04 -.02 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (d) Sugarcane High Suitability (e) War Events (f) War Victims .05 1 8 6 0 Estimate magnitude Estimate magnitude Estimate magnitude .5 4 -.05 2 0 -.1 0 -2 -.15 -.5 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (g) Incarcerations .04 Estimate magnitude 0 -.02 .02 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 63 Figure D.5. Smooth Condition Test Under Different Bandwidths (a) Altitude (b) Slope of Altitude (c) Terrain Ruggedness 5 40 3 4 20 Coefficient magnitude Coefficient magnitude Coefficient magnitude 2 3 0 2 1 -20 1 0 0 -40 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 496.4 Mean of Outcome: 7.46 Mean of Outcome: 10.92 (d) Hydrographic Bodies (e) Roads-Railway Infrastructure (f) Had a City (1945) .4 .3 .2 .2 Coefficient magnitude Coefficient magnitude Coefficient magnitude .1 .2 .1 0 0 0 -.1 -.1 -.2 -.2 -.2 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .28 Mean of Outcome: .38 Mean of Outcome: .09 (g) Distance to City (1945) (h) Distance to Communications (1945) (i) Communications Density (1945) .4 .4 .2 .2 .2 0 Coefficient magnitude Coefficient magnitude Coefficient magnitude 0 0 -.2 -.2 -.2 -.4 -.4 -.6 -.4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 1.21 Mean of Outcome: 1.24 Mean of Outcome: .33 Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 64 Figure D.6. Smooth Condition Test Under Different Bandwidths (cont’d) (a) Part of Land Reform (1980) (b) Inside Cultivated Area (1980) (c) Had a Parish (1979) .08 .1 .2 .06 .1 0 Coefficient magnitude Coefficient magnitude Coefficient magnitude .04 -.1 0 .02 -.2 -.1 0 -.02 -.3 -.2 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .1 Mean of Outcome: .67 Mean of Outcome: .01 (d) Distance to Parish (1980) (e) Distance to School (1980) (f) Total Population 40 .5 1 20 Coefficient magnitude Coefficient magnitude Coefficient magnitude 0 .5 0 -.5 0 -20 -40 -.5 -1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 4.13 Mean of Outcome: 20.35 Mean of Outcome: 160.59 (g) Population Density (h) Years of Education (i) Birth Rate 500 .5 .5 0 0 0 Coefficient magnitude Coefficient magnitude Coefficient magnitude -500 -.5 -.5 -1000 -1 -1 -1500 -1.5 -1.5 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 1870.92 Mean of Outcome: 4.01 Mean of Outcome: .17 Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 65 Figure D.7. Smooth Condition Test Under Different Bandwidths (cont’d) (a) In-migration Share (b) Out-migration Share (c) Inside Highly Populated Area (1980) .1 .1 .005 .05 Coefficient magnitude Coefficient magnitude Coefficient magnitude .05 0 0 -.005 0 -.05 -.05 -.01 -.1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .14 Mean of Outcome: .01 Mean of Outcome: .71 (d) Aggregate Yield Index (1961–79) (e) Bean Agro-climatic Yield (f) Coffee Agro-climatic Yield .06 .04 .2 .04 .02 .1 Coefficient magnitude Coefficient magnitude Coefficient magnitude .02 0 0 -.02 0 -.1 -.02 -.04 -.04 -.06 -.2 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 0 Mean of Outcome: 4.06 Mean of Outcome: 1.67 (g) Cotton Agro-climatic Yield (h) Maize Agro-climatic Yield (i) Wet Rice Agro-climatic Yield .01 .15 .3 .1 .2 Coefficient magnitude Coefficient magnitude Coefficient magnitude .005 .05 .1 0 0 0 -.05 -.1 -.1 -.005 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .71 Mean of Outcome: 9.94 Mean of Outcome: 8.630000000000001 Notes: The results follow the specification of equation (1). The estimates shown include 400 break fixed effects. 66 Figure D.8. Smooth Condition Test Under Different Bandwidths (cont’d) (a) Bean High Suitability (b) Coffee High Suitability (c) Maize High Suitability .15 .1 .1 .1 .05 Coefficient magnitude Coefficient magnitude Coefficient magnitude .05 .05 0 0 0 -.05 -.05 -.05 -.1 -.1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .92 Mean of Outcome: .13 Mean of Outcome: .98 (d) Sugarcane High Suitability (e) War Events (f) War Victims .1 .6 3 .4 2 .05 Coefficient magnitude Coefficient magnitude Coefficient magnitude .2 1 0 0 0 -.05 -1 -.2 -.1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .17 Mean of Outcome: .02 Mean of Outcome: .09 (g) Incarcerations .06 .04 Coefficient magnitude .02 0 -.02 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Mean of Outcome: .08 Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 67 D.B Main Outcomes Figure D.9. Effects of Guerrilla Control on Main Outcomes (a) Arcsine(Night Light) (b) Wealth Index .2 .2 .1 Estimate magnitude Estimate magnitude .1 0 0 -.1 -.1 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (c) Years of Education .6 .4 Estimate magnitude 0 .2 -.2 -.4 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 68 Figure D.10. Effects of Guerrilla Control on Main Outcomes under Different Bandwidths (a) Arcsine (Night Light) (b) Wealth Index 0 .2 -.1 Coefficient magnitude Coefficient magnitude 0 -.2 -.3 -.2 -.4 -.5 -.4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Mean of Outcome: 3.18 Mean of Outcome: -.17 (c) Years of Education .5 0 Coefficient magnitude -.5 -1 -1.5 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Mean of Outcome: 6.100000000000001 Notes: The results follow the specification of equation 1. The estimates shown include up to 400 break fixed effects. The figure illustrates the coefficients for 40 individual estimations, one for each of the different bandwidths around the discontinuity. The gray coloring illustrates 95% confidence intervals. 69 Figure D.11. External Validity for Main Outcomes (a) Arcsine (Night Light) (b) Wealth Index .5 4 0 3 Outcome Mean Outcome Mean -.5 2 -1 1 -1.5 -2 0 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 Distance to the Boundary Distance to the Boundary (c) Years of Education 8 6 Outcome Mean 4 2 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 -1 -1 -1 -1 -1 -1 -1 -1 Distance to the Boundary Notes: The figure shows the raw mean of each outcome by bin. Each bin corresponds to the distance to the boundary in kilometers, which ranges from 17 kilometers outside the guerrilla-controlled boundary to 18 kilometers within the boundary. Negative values signal being outside the boundary and positive values mean being inside the boundary. 70 Figure D.12. Effects of Guerrilla Control on the Number of Primary Schools per 100k Population Over Time Notes: This figure shows the coefficients obtained from the estimation of equation 1 for each year between 1999 and 2018. The gray coloring illustrates 95% confidence intervals. The estimates shown include up to 400 break fixed effects. The figure illustrates the coefficients of each yearly estimation from 1999 to 2018. Overall, the effect of guerrilla control on the number of primary schools per capita is positive and stable over time. 71 D.C Public Goods Provision Figure D.13. Effects of Guerrilla Control on Public Goods Provision (a) Sewerage Service (b) Water Access (c) Electricity .15 .2 .05 .1 Estimate magnitude Estimate magnitude Estimate magnitude .1 .05 0 0 0 -.05 -.05 -.1 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (d) Garbage Disposal (e) Total Hospitals (f) Total Schools .1 .1 .1 .05 Estimate magnitude Estimate magnitude Estimate magnitude .05 0 0 0 -.1 -.05 -.05 -.2 -.1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (g) Road Density 2 1.5 Estimate magnitude .5 1 0 -.5 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 72 Figure D.14. Effects of Guerrilla Control on Household Conditions under Different Bandwidths (a) Sewerage Service Rate (b) Water Access (c) Electricity Rate .15 .1 .1 .1 .05 Coefficient magnitude Coefficient magnitude Coefficient magnitude 0 .05 0 -.05 0 -.1 -.1 -.05 -.15 -.2 -.1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .35 Mean of Outcome: .74 Mean of Outcome: .86 (d) Garbage Rate (e) Total Hospitals (f) Total Schools .1 .1 .6 0 .4 0 Coefficient magnitude Coefficient magnitude Coefficient magnitude -.1 -.1 .2 -.2 -.2 0 -.3 -.4 -.3 -.2 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Bandwidth (Km) Bandwidth (Km) Mean of Outcome: .44 Mean of Outcome: .07 Mean of Outcome: .47 (g) Road Density .5 0 Coefficient magnitude -.5 -1 -1.5 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Mean of Outcome: 1.72 Notes: The results follow the specification of equation 1. The estimates shown include up to 400 break fixed effects. The figure illustrates the coefficients for 40 individual estimations, one for each of the different bandwidths around the discontinuity. The gray coloring illustrates 95% confidence intervals. 73 Figure D.15. Plot of the Effect of Guerrilla Control on the Share of Workers by Economic Activity (a) Agriculture (b) Industry .15 .05 .1 Estimate magnitude Estimate magnitude 0 0 .05 -.05 -.05 -.1 -.1 -2 -1 0 1 2 -2 -1 0 1 2 Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI (c) Services (d) Agricultural Workers Growing Cereals and Fruits .1 .5 .05 .4 Estimate magnitude Estimate magnitude 0 .3 -.05 .2 -.1 -.15 .1 -2 -1 0 1 2 -2 -1 0 1 2 Distance to border (Km) Distance to border (Km) Mean residual per bin Linear prediction 95% CI Mean residual per bin Linear prediction 95% CI Notes: The results follow the specification of equation (1). The estimates shown include up to 400 break fixed effects. 74 Figure D.16. Share of Workers by Economic Activity and Distance to the Boundary (a) Agriculture (b) Industry .4 .8 .3 .6 .2 .4 .1 .2 0 0 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 -1 (c) Services .8 .6 .4 .2 0 7 6 5 4 3 2 1 0 -9 -8 -7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 18 -1 -1 -1 -1 -1 -1 -1 -1 Notes: The figure shows the raw mean of each outcome by bin. Each bin corresponds to the distance to the boundary in kilometers, which ranges from 17 kilometers outside the guerrilla-controlled boundary to 18 kilometers within the boundary. Negative values signal being outside the boundary and positive values mean being inside the boundary. 75 Figure D.17. Plotting the Effects of Guerrilla Control on Homicide Rates (a) Homicides (2017) .3 .2 Estimate magnitude 0 .1 -.1 -.2 -1 -.8 -.6 -.4 -.2 0 .2 .4 .6 .8 1 Distance to border (Km) Mean residual per bin Linear prediction 95% CI Notes: This figure shows the results obtained from the estimation of equation (1). The estimates shown include up to 400 break fixed effects. There are no effects of guerrilla control on homicide rates in 2017. Figure D.18. Effects of Guerrilla Control on Homicide Rates under Different Bandwidths (a) Homicides (2017) 1.5 1 Coefficient magnitude .5 0 -.5 -1 .1 .2 .3 .4 .5 .6 .7 .8 .9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3 3.1 3.2 3.3 3.4 3.5 3.6 3.7 3.8 3.9 4 Bandwidth (Km) Mean of Outcome: .308 Notes: This figure shows the results obtained from the estimation of equation (1). The figure illustrates the coefficients for 40 individual estimations, one for each of the different bandwidths around the discontinuity. The estimates shown include up to 400 break fixed effects. The gray coloring illustrates 95% confidence intervals. 76 Table D.1. Effects of Guerrilla Territorial Control on Other Transformations of Night Light Luminosity Transformations of Night Light (2013) Literacy Rate Logarithm Level (Raw) Weighted by Pixel Area (2007) (1) (2) (3) (4) Guerrilla control -0.218*** -1.710*** -1.710*** -0.0212*** (0.0294) (0.339) (0.339) (0.00501) Observations 3,652 3,652 3,652 3,637 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 2.810 30.725 30.725 0.810 Note: The table presents the results of equation 1 using different transformations of night light luminosity. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. We use the algorithm of Calonico, Cattaneo and Titiunik (2014) to set the bandwidth and weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Table D.2. Effects of Guerrilla Control on Years of Education by Age Cohort Years of Education In School Not In School Age at War Age at War (1982-92) (1982-92) (1) (2) Guerrilla control -0.346*** -0.160 (0.121) (0.113) Observations 3,635 3,635 Bandwidth (Km) 2.266 2.266 Dependent mean 7.860 4.410 Notes: The table presents the effects of guerrilla control on the years of education by age cohort. Column 1 estimates the effect for the sample of people who during the war period were school age. Column 2 does the same but uses the sample of people who during this period were not school age. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. We use the algorithm of Calonico, Cattaneo and Titiunik (2014) to set the bandwidth and weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 77 Table D.3. Effects of Guerrilla Territorial Control on Main Outcomes Using Conley Standard Errors Panel A: Conley Standard Errors (0.5 Kms) Night Light Arcsine Wealth Index Years of Education (2013) (2007) (2007) (1) (2) (3) Guerrilla control -0.186*** -0.121*** -0.279*** (0.0242) (0.0343) (0.103) Observations 3,652 3,630 3,637 Panel B: Conley Standard Errors (2 Kms) Guerrilla control -0.186*** -0.121** -0.279** (0.0278) (0.0482) (0.129) Observations 3,652 3,630 3,637 Panel C: Conley Standard Errors (4 Kms) Guerrilla control -0.186*** -0.121** -0.279** (0.0344) (0.0566) (0.142) Observations 3,652 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 Note: The table presents the results of equation 1 using Conley standard errors. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. We use the algorithm of Calonico, Cattaneo and Titiunik (2014) to set the bandwidth and weight using a triangular kernel. Conley standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 78 Table D.4. Robustness Analysis for the Night Light Intensity Outcome Night Light Arcsine (2013) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Polynomial of order zero Guerrilla control -0.153*** -0.160*** -0.153*** -0.147*** -0.346*** -0.153*** -0.153*** -0.160*** -0.153*** -0.147*** -0.346*** -0.153*** (0.0278) (0.0278) (0.0277) (0.0295) (0.0220) (0.0278) (0.0278) (0.0278) (0.0277) (0.0295) (0.0220) (0.0278) Observations 1,494 1,344 1,443 1,406 4,946 1,442 1,494 1,344 1,443 1,406 4,946 1,442 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 0.588 0.441 0.538 0.510 3.388 0.535 0.588 0.441 0.538 0.510 3.388 0.535 Dependent mean 3.247 3.201 3.205 3.183 3.666 3.200 3.247 3.201 3.205 3.183 3.666 3.200 Panel B: Polynomial of order one (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Guerrilla control -0.186*** -0.215*** -0.198*** -0.201*** -0.232*** -0.211*** -0.142*** -0.153*** -0.147*** -0.159*** -0.188*** -0.165*** (0.0247) (0.0252) (0.0248) (0.0233) (0.0238) (0.0237) (0.0298) (0.0298) (0.0295) (0.0273) (0.0272) (0.0275) Observations 3,652 3,373 3,619 4,221 4,019 4,092 2,542 2,342 2,514 2,953 2,808 2,851 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 2.266 2.040 2.235 2.750 2.571 2.630 1.414 1.273 1.395 1.717 1.605 1.641 Dependent mean 3.536 3.517 3.537 3.594 3.568 3.578 3.453 3.440 3.452 3.506 3.497 3.498 Panel C: Polynomial of order two (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Guerrilla control -0.205*** -0.252*** -0.220*** -0.231*** -0.239*** -0.235*** -0.140*** -0.147*** -0.146*** -0.225*** -0.234*** -0.235*** (0.0274) (0.0286) (0.0277) (0.0243) (0.0269) (0.0252) (0.0336) (0.0338) (0.0334) (0.0257) (0.0282) (0.0263) Observations 4,851 4,834 4,842 8,244 7,595 8,096 3,232 3,212 3,220 5,962 5,282 5,824 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 3.303 3.284 3.292 7.583 6.353 7.296 1.927 1.916 1.921 4.424 3.707 4.257 Dependent mean 3.665 3.663 3.664 3.802 3.800 3.807 3.496 3.497 3.498 3.712 3.681 3.706 Note: The table presents the robustness of the effects of guerrilla control on night light intensity using different poly- nomial orders. Panel A shows results for a constant polynomial. Panels B and C present the results using a first and second order polynomial, respectively. Estimations across Columns show different bandwidth and kernel types and different bandwidth size. Robust standard errors in parentheses. “mserd” and “msetwo” specify one and two common MSE-optimal bandwidth selectors for the RD treatment effect estimator, respectively. “cerrd” and “certwo” indicate one or two common CER-optimal bandwidth selectors for the RD treatment effect estimator, respectively. The Kernel row indicates the type of kernel used: triangular, uniform, or epanechnikov. Differences in the number of observations are due to the selection of different bandwidths across specifications. *** p<0.01, ** p<0.05, * p<0.1. 79 Table D.5. Robustness Analysis for the Wealth Index Outcome Wealth Index (2007) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Polynomial of order zero Guerrilla control -0.213*** -0.220*** -0.208*** -0.211*** -0.210*** -0.208*** -0.213*** -0.220*** -0.208*** -0.211*** -0.210*** -0.208*** (0.0506) (0.0552) (0.0503) (0.0517) (0.0486) (0.0507) (0.0506) (0.0552) (0.0503) (0.0517) (0.0486) (0.0507) Observations 1,258 1,124 1,221 1,240 1,173 1,216 1,258 1,124 1,221 1,240 1,173 1,216 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 0.391 0.274 0.358 0.374 0.315 0.354 0.391 0.274 0.358 0.374 0.315 0.354 Dependent mean -0.327 -0.317 -0.331 -0.326 -0.361 -0.332 -0.327 -0.317 -0.331 -0.326 -0.361 -0.332 Panel B: Polynomial of order one (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Guerrilla control -0.120*** -0.100*** -0.109*** -0.118*** -0.103*** -0.107*** -0.144*** -0.118** -0.133*** -0.137*** -0.111** -0.127*** (0.0397) (0.0367) (0.0392) (0.0374) (0.0365) (0.0374) (0.0504) (0.0461) (0.0498) (0.0471) (0.0457) (0.0471) Observations 2,987 3,066 2,933 3,298 3,104 3,179 2,088 2,125 2,057 2,289 2,164 2,204 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 1.756 1.809 1.718 2 1.846 1.901 1.096 1.129 1.072 1.248 1.152 1.186 Dependent mean -0.0530 -0.0510 -0.0570 -0.0350 -0.0500 -0.0470 -0.168 -0.169 -0.176 -0.136 -0.153 -0.145 Panel C: Polynomial of order two (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Guerrilla control -0.104** -0.107** -0.101** -0.140*** -0.136*** -0.142*** -0.139** -0.128** -0.126** -0.125*** -0.145*** -0.120*** (0.0436) (0.0428) (0.0424) (0.0337) (0.0340) (0.0339) (0.0561) (0.0542) (0.0540) (0.0404) (0.0403) (0.0406) Observations 4,308 4,218 4,460 7,227 6,909 7,052 2,861 2,801 2,959 5,001 4,740 4,841 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 2.852 2.768 2.978 5.915 5.514 5.685 1.664 1.615 1.738 3.452 3.218 3.318 Dependent mean 0.0460 0.0360 0.0580 0.201 0.194 0.197 -0.0690 -0.0670 -0.0590 0.104 0.0920 0.0950 Note: The table presents the robustness of the effects of guerrilla control on the wealth index using different polynomial orders. Panel A shows results for a constant polynomial. Panels B and C present the results using a first and second order polynomial, respectively. Estimations across Columns show different bandwidth and kernel types and different bandwidth size. “mserd” and “msetwo” specify one and two common MSE-optimal bandwidth selectors for the RD treatment effect estimator, respectively. “cerrd” and “certwo” indicate one or two common CER-optimal bandwidth selectors for the RD treatment effect estimator, respectively. The Kernel row indicates the type of kernel used: triangular, uniform, or epanechnikov. Robust standard errors in parentheses. Differences in the number of observations are due to the selection of different bandwidths across specifications. *** p<0.01, ** p<0.05, * p<0.1. 80 Table D.6. Robustness Analysis for the Years of Education Outcome Years of Education (2007) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Polynomial of order zero Guerrilla control -0.648*** -0.658*** -0.650*** -0.654*** -0.637*** -0.592*** -0.648*** -0.658*** -0.650*** -0.654*** -0.637*** -0.592*** (0.154) (0.170) (0.157) (0.172) (0.140) (0.111) (0.154) (0.170) (0.157) (0.172) (0.140) (0.111) Observations 1,348 1,154 1,289 1,249 1,289 1,669 1,348 1,154 1,289 1,249 1,289 1,669 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 0.458 0.298 0.409 0.378 0.409 0.746 0.458 0.298 0.409 0.378 0.409 0.746 Dependent mean 5.761 5.849 5.834 5.843 5.834 5.867 5.761 5.849 5.834 5.843 5.834 5.867 Panel B: Polynomial of order one Guerrilla control -0.280** -0.197 -0.230** -0.277** -0.145 -0.236** -0.441*** -0.331** -0.409*** -0.433*** -0.361** -0.422*** (0.117) (0.121) (0.114) (0.115) (0.119) (0.117) (0.157) (0.167) (0.154) (0.155) (0.164) (0.159) Observations 3,308 2,755 3,238 3,369 2,808 3,140 2,297 1,950 2,247 2,336 1,987 2,188 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 2.001 1.572 1.942 2.051 1.615 1.868 1.249 0.981 1.212 1.280 1.008 1.166 Dependent mean 6.510 6.358 6.477 6.514 6.399 6.463 6.168 6.015 6.143 6.192 6.030 6.135 Panel C: Polynomial of order two Guerrilla control -0.283** -0.260* -0.229* -0.281*** -0.305*** -0.290*** -0.484** -0.374** -0.466** -0.328** -0.263** -0.285** (0.139) (0.133) (0.139) (0.102) (0.108) (0.103) (0.188) (0.178) (0.189) (0.129) (0.134) (0.130) Observations 4,441 4,357 4,296 7,167 6,274 6,902 2,951 2,892 2,852 4,934 4,265 4,731 Bandwidth type mserd mserd mserd msetwo msetwo msetwo cerrd cerrd cerrd certwo certwo certwo Kernel triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov triangular uniform epanechnikov Bandwidth (Km) 2.956 2.884 2.834 5.815 4.796 5.488 1.725 1.683 1.654 3.394 2.799 3.202 Dependent mean 6.828 6.791 6.776 7.270 7.178 7.269 6.425 6.402 6.398 6.984 6.767 6.949 Note: The table presents the robustness of the effects of guerrilla control on the number of years of education using different polynomial orders. Panel A shows results for a constant polynomial. Panels B and C present the results using a first and second order polynomial, respectively. “mserd” and “msetwo” specify one and two common MSE- optimal bandwidth selectors for the RD treatment effect estimator, respectively. “cerrd” and “certwo” indicate one or two common CER-optimal bandwidth selectors for the RD treatment effect estimator, respectively. The Kernel row indicates the type of kernel used: triangular, uniform, or epanechnikov. Estimations across Columns show different bandwidth and kernel types and different bandwidth size. Robust standard errors in parentheses. Differences in the number of observations are due to the selection of different bandwidths across specifications. *** p<0.01, ** p<0.05, * p<0.1. 81 Table D.7. Effects of Guerrilla Territorial Control on Main Outcomes Using Ordinary Least Squares Night Light Arcsine Wealth Index Years of Education (2013) (2007) (2007) (1) (2) (3) Guerrilla control -0.477*** -0.471*** -1.100*** (0.0257) (0.0221) (0.0607) Observations 12,411 12,370 12,384 Bandwidth (Km) 67.01 67.01 67.01 Dependent mean 3.457 -0.0310 6.505 Note: The table presents the results of equation 1 via Ordinary Least Squares using the whole sample. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. We use triangular kernel weights. Standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 82 Table D.8. Placebo Test for All Pairs of Neighbors Whose Difference in Altitude is between the Following Thresholds Altitude difference between 15 and 20 masl Altitude difference between 20 and 100 masl Altitude Night Light- Arcsine (2013) Altitude Night Light- Arcsine (2013) Any neighbor Any neighbor Both neighbors outside Any neighbor Any neighbor Both neighbors outside pair pair guerrilla area pair pair guerrilla area (1) (2) (3) (4) (5) (6) Difference 17.83*** 0.0206*** 0.0239*** 47.71*** -0.0114*** -0.0172*** (0.0322) (0.00521) (0.00525) (0.201) (0.00384) (0.00430) Neighbor pairs 2,914 2,914 2,515 11,811 11,811 8,742 Wealth Index (2007) Wealth Index (2007) (7) (8) (9) (10) Difference - 0.0149 0.0202** - -0.0456*** -0.0468*** - (0.00921) (0.00980) - (0.00501) (0.00583) Neighbor pairs - 2,910 2,513 - 11,729 8,733 Years of Education (2007) Years of Education (2007) (11) (12) (13) (14) Difference - 0.0818*** 0.0964*** - -0.0540*** -0.0513*** - (0.0307) (0.0336) - (0.0144) (0.0172) Neighbor pairs - 2,911 2,513 - 11,758 8,734 Note: The table presents the placebo test results. The unit of observation in Columns 1 to 3 is the pair of neighboring census tracts conditional on having a difference in altitude between 15 and 20 masl. The unit of observation in Columns 4 and 5 is the pair of neighboring census tracts conditional on having a difference in altitude between 20 and 100 masl. Columns 1, 2, 4, and 5 show the mean difference for all neighbor pairs in the sample. Columns 3 and 6 do the same for pairs in which both neighboring tracts are outside the guerrilla-controlled area. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 83 Table D.9. Main Results Restricting the Sample to Tracts without Sudden Altitude Changes with Respect to Their Neighbors Night Light Arcsine Wealth Index Years of Education (2013) (2007) (2007) (1) (2) (3) Guerrilla control -0.146*** -0.120*** -0.309** (0.0240) (0.0439) (0.137) Observations 2,572 2,561 2,562 Bandwidth (Km) 2.103 2.103 2.103 Dependent mean 3.743 0.118 6.924 Note: The table presents main results without considering segments that have a difference in altitude of more than 100 masl with respect to their neighbors. Column 1 shows the effect of whether a census tract was under guerrilla control on the arcsine of night light luminosity from NOAA. Column 2 does the same but uses as dependent variable a standardized score of household wealth. Column 3 shows as dependent variable years of education of the population older than 18 years. The unit of observation in all columns is the census tract. Information from Columns 2 and 3 was obtained from the Population Census of 2007. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Table D.10. Effects of Guerrilla Territorial Control on Main Outcomes for Individuals Who Have Always Lived in the Same Place Wealth Index Years of Education Literacy Rate (2007) (2007) (2007) (1) (2) (3) Guerrilla control -0.132*** -0.402*** -0.0261*** (0.0356) (0.112) (0.00563) Observations 3,621 3,633 3,633 Bandwidth (Km) 2.266 2.266 2.266 Dependent mean -0.0280 6.785 0.817 Note: The table presents main results for the sample of people who have always lived in the same place. The unit of observation in all Columns is the census tract. The information was obtained from the Latin American Public Opinion Project survey (LAPOP). Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 84 Table D.11. Comparison of Baseline Characteristics Between Census Tracts In and Out the RD-Sample In RD-Sample Out of RD-Sample Baseline Characteristics Mean Obs Mean Obs Geographic Characteristics (Before 1980) Altitude 488.319 3,681 499.802 8,752 Slope 8.624 3,681 6.968 8,751 Ruggedness 12.381 3,681 10.300 8,751 Hydrography 0.320 3,681 0.266 8,754 Infrastructure Characteristics (Before 1980) Roads and Railway 0.401 3,681 0.366 8,754 Had a City/Village 0.128 3,681 0.073 8,754 Distance to City/Village 1.024 3,681 1.285 8,754 Distance to Comms 1.199 3,681 1.257 8,754 Comms Density 0.328 3,681 0.334 8,754 Had Land Reform 0.081 3,681 0.112 8,754 Cultivated Area 0.665 3,681 0.676 8,754 Had a Parish 0.011 3,681 0.011 8,754 Distance to Parish 4.309 3,681 4.055 8,754 Distance to School 16.980 3,681 21.771 8,754 Population Demographics (Before 1980) Total Population 158.233 3,667 161.574 8,735 Population Density 1,418.195 3,666 2,060.920 8,735 Years of Education 3.493 3,666 4.227 8,737 Natality Rate 0.174 3,664 0.175 8,730 In-migration (Share) 0.108 3,636 0.147 8,646 Out-migration (Share) 0.006 3,446 0.008 8,272 Highly Populated Area 0.674 3,681 0.722 8,754 Agro-Climatic Potential Yield (1961-1979) Z-Potential Yield -0.012 3,681 0.005 8,754 Bean Potential Yield 4.056 3,669 4.068 8,632 Coffee Potential Yield 1.678 3,669 1.670 8,632 Cotton Potential Yield 0.709 3,669 0.709 8,632 Maize Potential Yield 9.827 3,669 9.990 8,632 Wet Rice Potential Yield 8.714 3,669 8.591 8,632 Sugarcane Potential Yield 6.408 3,669 6.307 8,632 Crops’ High Suitability (1961-1990) Bean High Suitability 0.858 3,691 0.942 8,736 Coffee High Suitability 0.086 3,691 0.146 8,736 Maize High Suitability 0.980 3,691 0.983 8,736 Sugarcane High Suitability 0.108 3,691 0.194 8,736 Conflict (Before 1981) and Incarcerations (1980-1985) Number of War Events 0.037 3,681 0.018 8,754 Number of War Victims 0.155 3,681 0.056 8,754 Number of Incarcerations 0.018 3,681 0.107 8,754 Note: The table compare the mean and number of observations of outcomes in Table 1 between census tracts in the RD-sample and census tracts outside the sample. 85 Table D.12. Effects of Guerrilla Territorial Control in the Elections of 2014 and 2015 Panel A: 2014 Presidential elections - Guerrillas’ Party won Left Voting Right Voting Blank Voting Turnout Share Share Share Share (1) (2) (3) (4) Guerrilla control -0.0350* 0.0341 0.00387*** 0.0166 (0.0199) (0.0221) (0.00131) (0.0103) Observations 416 416 416 416 Bandwidth (Km) 2.930 2.930 2.930 2.930 Dependent mean 0.483 0.395 0.00700 0.565 Panel B: 2015 Municipal elections Guerrilla control -0.0152 -0.00723 0.00207** 0.0300 (0.0278) (0.0259) (0.000905) (0.0219) Observations 434 434 434 434 Bandwidth (Km) 3.239 3.239 3.239 3.239 Dependent mean 0.411 0.629 0.00700 0.513 Note: The table presents the results of equation 1 for our outcomes related to electoral results. The unit of observation in all Columns is the polling station. Panel A shows the results for the presidential elections of 2014 and panel B does the same for the municipal elections of 2015. The information was obtained from the Salvadoran Electoral Court. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Clustered errors at the Canton level are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 86 Table D.13. Effects of Guerrilla Territorial Control on Political Attitudes Total Sum of Questions per Item/Scope Political Engagement with Non-Democratic Trust in Participation Politicians Engagement Institutions (1) (2) (3) (4) Guerrilla control 1.449 -0.380** 0.181 -4.112*** (1.098) (0.184) (1.183) (1.403) Observations 242 248 172 241 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 12.960 0.383 4.780 11.720 Note: The table presents the results of equation 1 for our outcomes related to political discontent and distrust. Col- umn 1 shows the political participation scope, which includes questions that measure whether the citizen votes, at- tends protests, and attends government meetings. Column 2 reports the engagement with politicians’ scope, which measures the extent to which citizens contact state authorities and/or bureaucracies to solve issues and attend govern- ment/political meetings. Column 3 shows the nondemocratic engagement scope, which measures the extent to which citizens approve the use of alternative or violent means to engage in politics. Column 4 reports the trust in institutions item, which measures the extent to which citizens trust different types of Salvadoran institutions, including the police, the powers of state, and local government. The table uses the simple sum of questions by each item as dependent variables. The unit of observation in all Columns is the census tract. The information was obtained from the Latin American Public Opinion Project survey (LAPOP). Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Table D.14. Effects of Guerilla Territorial Control on Distance to Police Stations and Incarcerations Distance to Incarcerations Police Stations (1992-1999) (1) (2) Guerrilla control 0.0198 0.0193 (0.0614) (0.0137) Observations 3,652 3,681 Bandwidth (Km) 2.266 2.266 Dependent mean 1.850 0.0580 Notes: This table shows the effects of guerrilla control on the distance to the closest local police station (Column 1) and the number of incarcerations per segment between 1992 and 1999 (Column 2). 87 Table D.15. Quality of School Teachers Total Enrollment Total Teachers Certified Teachers Certified Teachers with Teachers with High-School High-School (1) (2) (3) (4) (5) Guerrilla control 9.764 0.519 0.320 0.350 0.452 (35.31) (1.155) (1.123) (0.969) (0.991) Observations 1,522 1,522 1,522 1,522 1,522 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 Dependent mean 386.7 13.42 12.78 11.51 11.88 Notes: This table shows the effects of guerrilla control on school size (Columns 1 and 2) and quality of school teachers (Columns 3–5). Data was obtained from the 2013 teacher census provided by the Ministry of Education. “Total enroll- ment” and “Total teachers” refer to the total number of students and teachers at the school level, respectively. “Certified teachers” refers to teachers who have received a formal accreditation in pedagogy from the Ministry of Education. Table D.16. Simpson’s Index All plots Commercial plots Subsistence plots (1) (2) (3) Guerrilla control 0.0402* 0.0399† 1.28e-05 (0.0217) (0.0267) (0.0267) Observations 2,266 1,913 1,963 Bandwidth (Km) 2.266 2.266 2.266 Dependent mean 0.530 0.420 0.460 Note: The table presents the results of equation 1 for the Simpson’s Index calculated for all plots, commercial plots, and N 2 i ai subsistence plots in the Agrarian Census of 2007. The calculation of the Simpson’s Index is S = 1 − ( N a )2 where ai i i refers to the size of each plot. The unit of observation is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1, † p<0.15. 88 Table D.17. Workers by Economic Activity Share of Workers by Economic Activity Share of Agricultural Workers Agriculture Industry Services Growing Cereals and Fruits (1) (2) (3) (4) Guerrilla control 0.0465*** -0.0261*** -0.0203** 0.0456*** (0.00985) (0.00559) (0.00878) (0.00944) Observations 3,636 3,636 3,636 3,636 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 0.194 0.227 0.579 0.160 Note: The table presents the results of equation 1 for the share of workers in each economic activity. The information was calculated from the Census 2007 and using ISIC v4 to classify each occupation. The unit of observation is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1, † p<0.15. Table D.18. Inequality of Income at the Canton Level Real Per Capita Income Logarithm Level Gini Index Interquartile Range Percentile Range Percentile Range (p75-p25) (p90-p10) (p90-p50) (1) (2) (3) (4) (5) (6) Guerrilla control -0.223*** -39.08** 0.0119 0.0217 1.490 0.0366 (0.0682) (15.25) (0.0163) (0.193) (1.256) (0.284) Observations 542 542 542 542 542 542 Bandwidth (Km) 3.082 3.082 3.082 3.082 3.082 3.082 Dependent mean 5.330 266.8 0.320 2.450 5.240 2.360 Note: The table presents the results of equation 1 for the real per capita income taken from the Household Surveys (2012 to 2018). Each Column represents a different measure of inequality using the real per capita income. Column 4 report the interquartile range, calculated as the difference of the per capita income in percentile 75 minus the per capita income in percentile 25 for each canton. Column 5 shows the percentile range of the difference between percentile 90 and 10 for each canton. Column 6 reports the percentile range of the difference between percentile 90 and 50 for each canton. The unit of observation is at the canton level. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. 89 Table D.19. Inequality of the Wealth Index at the Census Tract Level Wealth Index Gini Index Interquartile Range Percentile Range Percentile Range (p75-p25) (p90-p10) (p90-p50) (1) (2) (3) (4) Guerrilla control -0.00345 66.52 82.59 0.0679 (0.0104) (67.52) (68.28) (0.116) Observations 2,975 2,975 2,975 2,975 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 0.278 3.520 47.23 1.981 Note: The table presents the results of equation 1 for the wealth index constructed from Census of 2007. Each Column represents a different measure of inequality using the real per capita income. Column 2 report the interquartile range, calculated as the difference of the wealth index in percentile 75 minus the wealth index in percentile 25 for each cen- sus tract. Column 3 shows the percentile range of the difference between percentile 90 and 10 for each census tract. Column 4 reports the percentile range of the difference between percentile 90 and 50 for each census tract. The unit of observation is at the census tract level. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. Table D.20. Cooperatives Has a Producer belongs Commercial producer Subsistence producer Producer belongs cooperative to a cooperative belongs to cooperative belongs to cooperative to association (1) (2) (3) (4) (5) Guerrilla control 0.00545 0.00226 0.00879 -0.00301 -0.00480 (0.00354) (0.00960) (0.0158) (0.00429) (0.00579) Observations 929 2,400 2,400 2,400 2,400 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 Dependent mean 0.00900 0.0520 0.0750 0.00800 0.0110 Note: The table presents the results of equation 1 for outcomes related to cooperatives. Information was taken from the Agrarian Census of 2007. The unit of observation is at the census tract level. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guer- rilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The estimates use triangular kernel weights. Robust standard errors in parentheses. 90 Table D.21. Effects of Guerrilla Control on Migration Outcomes for the Highly Educated Population International Migrants Always Lived in Same Location People who Arrived Years since During Control At any time Years since Households who Received Received Remittance from same Location as the Mother During Control Arrival (Share) (Share) departure Remittances (Share) War Migrant (Share) (Share) (Share) (Share) (1) (2) (3) (4) (5) (6) (7) (8) (9) Guerrilla control 0.00151 0.00343 0.226 -0.00573 -0.00112 -0.00376 -0.00713 -0.00491 -0.469 (0.00452) (0.00927) (0.540) (0.00463) (0.00416) (0.0127) (0.0132) (0.00535) (0.531) Observations 3,325 3,325 1,907 3,636 3,325 3,602 3,602 3,602 3,441 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 2.266 2.266 2.266 2.266 Dependent mean 0.0200 0.100 6.220 0.110 0.0100 0.730 0.700 0.0800 17.68 Note: The table presents the results of equation 1 for our outcomes related to migration. Columns 1 to 5 focus on outcomes for international migrants. All information was obtained from the Population Census of 2007. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Table D.22. Share of Individuals who Work in the Same Place as their Residence Work in the Same Place as Residence (Share) (1) Guerrilla control 0.00333 (0.00320) Observations 3,647 Bandwidth (Km) 2.271 Dependent mean 0.987 Note: The table presents the results of equation 1 for individuals who work in the same place as their residence. All in- formation was obtained from the Population Census of 2007. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates use triangular kernel weights. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 91 Table D.23. Heterogeneity by Baseline Distances to Road Network (1980) and Nearest City (1945) Panel A: Heterogeneity by Distance to Road Network in 1980 Arcsine Wealth Index Years of Education (1) (2) (3) Guerrilla control -0.177*** -0.0953** -0.272** (0.0272) (0.0399) (0.125) Control × Distance to Road 0.00503 -0.0267 0.0410 (0.0212) (0.0230) (0.0737) Dependent mean 3.536 -0.0160 6.573 Observations 3,652 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 Panel B: Heterogeneity by Distance to Nearest City in 1945 Arcsine Wealth Index Years of Education (1) (2) (3) Guerrilla control -0.225*** -0.109*** -0.295** (0.0307) (0.0412) (0.116) Control × Distance to City 0.0375** -0.00881 0.0187 (0.0149) (0.0223) (0.0642) Dependent mean 3.536 -0.0160 6.573 Observations 3,652 3,630 3,637 Bandwidth (Km) 2.266 2.266 2.266 Note: The table presents the results from the heterogeneity analysis at baseline for the main outcomes. Panel A shows how the results vary by distance to a road network in 1980. Panel B presents heterogeneity of results by distance to the nearest city in 1945. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 92 Table D.24. Effects of Guerrilla Territorial Control on Crimes during the War Period Total War Events Total War Victims Has a War Event Has War Victims (1) (2) (3) (4) Guerrilla control 0.00660 -0.258 0.00180 0.00322 (0.0894) (0.490) (0.00264) (0.00287) Observations 3,652 3,652 3,652 3,652 Bandwidth (Km) 2.266 2.266 2.266 2.266 Dependent mean 0.0410 0.2130 0.00100 0.00200 Note: The table presents the results of estimating equation 1 for our outcomes related to crimes committed in the war period. Columns 1 and 3 report the total of events related to war and its probability, respectively. A war event can be a massacre, combat, bombing, or any other war event that caused victims from war. Columns 2 and 4 show the total number of victims and the probability of the census tract to have war victims. The unit of observation in all Columns is the canton level. The information was recovered from the registry of victims. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla- controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Table D.25. Effects of Guerrilla Control on Homicide and Victimization Rates Homicides Victim of Any Crime Victim of Gang Extortion (2017) (2004-2016) (2004-2016) (1) (2) (3) Guerrilla control -0.0110 -0.210*** -0.193*** (0.0562) (0.0552) (0.0637) Observations 3,652 94 94 Bandwidth (Km) 2.266 2.266 2.266 Dependent mean 0.314 0.688 0.0420 Note: The table presents the results of equation 1 for our outcomes related to current crime. Column 1 shows the number of homicides reported to police for each census tract in 2017. Column 2 shows the share of people within a census tract who reported being a victim of any type of crime in the LAPOP survey. Column 3 shows the share of people within a census tract who reported being a victim of extortion in the LAPOP survey. The unit of observation in all Columns is the census tract. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla-controlled boundary. We use the algorithm of Calonico, Cattaneo and Titiunik (2014) to set the bandwidth and weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 93 Table D.26. Effects of Guerrilla Territorial Control on Expropriation, Invasion, and Non-Democratic Believes Invading Occupying Overturn Taking Law in Non-Democratic Property Buildings the Government Own Hands Engagement (sum) (1) (2) (3) (4) (5) Guerrilla control -0.110* 0.0616 -0.00269 -0.0734 0.804 (0.0605) (0.0815) (0.0721) (0.136) (1.922) Observations 248 175 248 245 172 Bandwidth (Km) 2.266 2.266 2.266 2.266 2.266 Dependent mean 0.0580 0.109 0.0740 0.245 10.69 Note: The table presents the results of estimating equation 1 for our outcomes related to believes regarding how accept- able it is to engage in certain actions against private property or the government. These are measured in a 1-10 scale, thus, we assume that individuals support these behaviors whenever their agreement level is above 5. Columns 1 to 4 show the share of individuals who think that invading property, occupying buildings, overthrowing the government or taking law in their own hands are acceptable. Column 5 shows the effects of guerilla control on an index comprised of the sum of the raw scores . The information was recovered from the LAPOP Surveys. Controls not shown include a linear polynomial of the distance to the boundary of guerrilla territory, its interaction with whether the tract was under guerrilla control or not, and up to 400 fixed effects representing the closest evenly spaced break in the guerrilla- controlled boundary. The algorithm of Calonico, Cattaneo and Titiunik (2014) was used to set the bandwidth and the estimates weight using a triangular kernel. Robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. 94 E Qualitative Study This appendix provides further information on the methods used in the qualitative component of the study and their main results. E.A Sample definition and recruitment of participants The qualitative study aims to complement the quantitative results by gathering information to understand the dynamics that occurred within the territory controlled by the Salvadoran guerrilla, the stability of the borders, changes in the economic, and the social or political structure caused by the presence of guerrillas in the territory, among other potential mechanisms that can drive the main impacts documented in this study. The target groups were: (i) political-military leaders of the guerrilla, which designed and imple- mented the military strategy and policies with a broad knowledge of the grassroots guerrilla’s social movement; (ii) religious and community leaders with depth knowledge of the armed con- flict; (iii) citizens who lived in the areas controlled by the guerrillas during the civil war; and (iv) former guerrilla members who were prominent in the operational-military area. Given the diversity of these groups, the information was collected using in-depth interviews and focus groups discussions. Groups (i) and (ii) were invited to join individual in-depth interviews and groups (iii) and (iv) were invited to participate in focus group discussions. A total of four focus groups and 8 in-depth interviews were conducted in June 2022. Focus groups were con- ducted in 3 municipalities of El Salvador: two focus groups in Chalatenango and Guazapa (one in each municipality), and two groups in Morazan. These municipalities were selected based on the intensity of guerrilla groups presence during the civil war.41 E.B Instruments Three instruments were developed: (i) for in-depth interviews (for religious or community leaders and political-military leaders); (ii) for focus group discussions of citizens who lived in guerrilla- controlled areas; and (iii) for focus group discussions with former guerrilla members. All three instruments include two components. First, questions related to the economic and social dynamics of guerrilla-controlled areas before and during the war. For example, the questions inquire about the main local economic activity before the arrival of the specific guerrilla group in 41 Since Morazan was a crucial department for the FMLN during the Civil War, two focus groups were conducted there. 95 charge of the area or about the form of government in place during the conflict. Second, questions on participants’ perceptions of changes in social and economic factors after the end of the armed conflict. For example, whether they perceived that the presence of the guerrillas affected the social and community ties in the area now in the present, among other questions. On the other hand, instruments (i) and (ii) also include questions related to the characterization of the geographic space controlled by the guerrillas. For example, in this section the instruments inquire whether and how borders of the controlled territories changed during the conflict, when these borders became more stable, or reasons for guerrilla’s settling in the controlled areas, among others. E.C Approach For the qualitative study, a narrative interviewing technique was used. It consisted of a semi- structured approach to interviewing that uses open-ended questions to allow for more variation in responses. These interviews and focus groups create a natural in-depth discussion that allows to obtain specific details of the different components included in the instruments. The interviews were between 60 to 70 minutes each and the focus group discussions lasted up to 1 hour. A local consultant with expertise on qualitative research and knowledge of the guerrillas’ movement in El Salvador conducted the interviews. She was responsible for recruiting partici- pants who met the eligibility criteria, obtaining their informed consent, and conducting the inter- views and producing their transcripts. For all the interviews, special care was taken to preserve the participants’ anonymity and freedom to consent. Indeed, the strategy for maintaining trust and safety was to be extremely clear to all participants that the purpose of the survey was only academic. Only audio of the conversations was recorded, and no photos or video were allowed. E.D Main results The main messages of the qualitative analysis are summarized below. Establishment of self-governance institutions to promote social capital Our interviews with FMLN commanders show that the consolidation of self-governance institu- tions in controlled areas was a key strategy of the guerrilla. From 1982 onwards, the state dis- appears in its traditional institutional framework. For example, municipal authorities ceased to function, local judges ceased to provide their services, etc. In the words of one the FMLN mil- itary commanders: ‘Mayors, judges, security posts, everything disappears, (...), practically the 96 ın Villalobos, FMLN Military Comman- state disappears, and the state was us [the FMLN]’ (Joaqu´ der, interview conducted on March 23, 2022). As a substitute for power, popular power emerges; that is, power determined by the people. When asked about FMLN-controlled areas, an influ- ential religious leaders that lived in these areas says ‘the project of structural change in control areas was always present. (...). Starting in 1982-1983, these places become controlled territories, the institutions disappear, and the popular powers emerge (...).’ (Religious leader, interview con- ducted on March 25, 2022). In these new institutions, the key principle was the organization of local communities: ‘the individual that lives in a controlled area has a clear consciousness that what prevails in these areas are values. (...) what was consolidated was an idea of social co-responsibility. (Religious leader, interview conducted on March 25, 2022). This strategy was not a by-product of the elimination of state authorities, but rather a deliberate plan to promote the autonomy of peasants from traditional government institutions. The change in military strategy- from a regular to an irregular war- that took place around 1984 was associated with the conviction that the civilian population had their right to live their own lives. Marisol Galindo, an FMLN commander explains: the locals ‘had a right to be on their own land, the right to harvest, to not be treated as armed population,(...), that is, we [the guerrilla] made a clear distinction between guerrilla members and civilian population. (...). We wanted to rescue organizational forms of what today we call the Civil Society (...).’ (Marisol Galindo, FMLN military commander, inter- view conducted January 28, 2022). When the state disappeared, governance was in charge of these informal institutions, like the ‘poder de doble cara’ (or double-faced power), which was the ‘self-governance of civilians, to solve their own needs (...), and it had to be done in confrontation ın Villalobos, FMLN Military Commander, interview conducted on March with the state’ (Joaqu´ 23, 2022). This organization of citizens in the communities made it possible to guarantee social cohesion or the “tejido social.” Our interviews uncovered powerful evidence of the persistence of the social capital generated by these institutions. In several instances, different individuals reflected upon the fact that, although these areas seem to be less developed, they are extremely secure. When the interviewer noted that the zones with guerrilla presence don’t have any gang presence, one of the former combatants said: ‘Yes [they are the most secure], and where judges die of boredom.’ She later added, ‘I relate this to the level of organization that the community achieved. I am going to give you an example; e de las Flores there is a river and thermal waters, and there is a little hotel. If you en San Jos´ go there and say you want to stay there for 10 days, they will ask you, who are you? Who sent 97 you? Once a fugitive gangster (marero) came who believed he could stay. It is impossible. They investigate who sent you, your references.’ (Lorena G, FMLN military commander, interviewed on January 28, 2022). The same point was made in other interviews, where an excombatant said ‘the fact that the maras (gangs) are barely present in these areas reflects that the self-organization ın Villalobos, FMLN Military Commander, interview conducted of the population worked.’ (Joaqu´ on March 23, 2022) The organization of the communities was promoted by local leadership groups, such as the Orga- nization of the Comadres and the Sisters of the Assumption. One of the paradigmatic civil society organizations that developed and still exists today is the Patronato para el Desarrollo de las Comu- nidades de Morazan y el Norte de San Miguel (PADECOMSM). This organization is based on a frame- work of participatory democracy and self-management, with local, zonal, and regional councils that identify problems and devise solutions. The PADECOMSM emerged as a consequence of autonomous space that was granted to civilians in controlled areas. Distrust towards the state Our interviews with locals show that state distrust was pervasive among peasants and lower- income individuals during the civil conflict, and not necessarily circumscribed to guerrilla-controlled areas. This is frequently attributed to the fact that the state was entwined with economic elites, which used highly repressive methods to discipline the workforce. As a result, peasants were usually landless, and endured hard working conditions. When talking about the economic and social conditions in these areas, one military commanders says: ‘In all those areas there were poor peasants and landowners, this was the predominant characteristic, landless peasants and big hacienda owners. (..) an additional issue was that it was problematic [for peasants to work the land, given that rent prices were impossible to afford, I mean, they worked to pay rent and what was left was useless, don’t even think about luxuries like water or electricity, that did not exist, that was a luxury’ (Lorena P, FMLN military commander, interview conducted on January 28). Participants described that, under these conditions, the need to reorganize themselves and create self-governing institutions in controlled areas was urgent, especially to substitute the tradi- tional model where elites and the state coerced labor, and where basic services were lacking. The absence of the state during the territorial control helped to reinforce this view, as the state could not provide any public service or have physical presence during the guerrilla occupation. Indeed, local leaders from the new institutions or international organizations end up providing public ser- 98 vices to the communities, including education and health. As a result, the distrust to the state was more likely to be greater in the controlled territory relative to other non-controlled areas. Migration decisions Participants reported some reasons for not migrating from the controlled areas. The interviews re- flect there was a sense of rootedness in the communities and attachment to their limited economic resources. One guerrilla commander says ‘there were many families, that is why some schools for children emerge [in the controlled zones], because many of these families wanted to stay. (...) What the stories from those years reflect is that there was an important population that did not want to leave’ (Marisol Galindo, FMLN military commander, interview conducted January 28, 2022). Stability of boundaries Ex-guerrilla leaders confirm that the boundaries between the controlled and non-controlled territories were stable after 1984-85. A potential explanation is that around 1984 the guerrilla changed their military strategy. The regular war against the Salvadoran state had reached a stalemate, and the FMLN decides to switch to an irregular strategy, based on the control ın Villalobos, one of the most important FMLN military commanders of liberated zones. Joaqu´ also mentions that the State made a crucial mistake underestimating their capacity and practically left them territory: ‘after they left us our territory, we moved to a superior level of organization ın Villalobos, FMLN Military Commander, interview con- and consolidation of power (...).’ (Joaqu´ ducted on March 23, 2022). All military commanders interviewed agree that after 1984 the bound- aries of the controlled areas were extremely stable, and confirmed that the map we use to identify control areas was the map used and approved by all parties during the peace talks sponsored by the UN. 99