The World Bank Economic Review, 39(3), 2025, 632–662 https://doi.org10.1093/wber/lhae037 Article Firms and Labor in Times of Violence: Evidence from Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 the Mexican Drug War Hale Utar Abstract This paper examines the impact of violence resulting from drug trafficking on manufacturing firms in an emerg- ing economy. By utilizing comprehensive longitudinal data spanning all of Mexico from 2005 to 2010, and em- ploying an instrumental variable strategy that leverages plausibly exogenous spatiotemporal variations in the homicide rate during the outbreak of drug violence, the analysis reveals a significant negative effect of violence on plant output, employment, product scope, and capacity utilization. The negative effect on employment is entirely driven by blue-collar employment and concentrated among low-wage, female-intensive firms. Further, consistent with a violent-environment-induced blue-collar labor-supply shock, the results show positive effects on blue-collar wages and negative effects on white-collar wages at the firm level. Output resilience to violence is also shown to be lower among labor-intensive, domestically selling and sourcing, less diversified firms. These findings show the rise of drug violence has a significant negative effect on development of domestic industrial capability in Mexico and shed light on the characteristics of the most affected firms and the channels through which they are affected. JEL classification: L25, L60, O12, O14, O18, O19, R11, O54, F14 Keywords: violent environment, manufacturing firms, economic development, low-wage female workers, real- location 1. Introduction Organized crime, often centering on drug trafficking, claims as many lives globally as all other armed con- flicts combined, a stark reality highlighted in the UNODC (2019) report. This phenomenon is particularly pronounced where the high demand for drugs from wealthier nations meets with weaker institutions of the Hale Utar is the Sidney Meyer Professor of International Economics at Grinnell College, Grinnell, USA; her email address is utarhale@grinnell.edu. The author thanks Luis B. Torres Ruiz and Gabriel Arturo Romero Velasco for help with the micro- data; David Shirk, Octavio Rodriguez Ferreira, Laura Calderón, and Juan Camilo Castillo for sharing their data; Rafael Dix- Carneiro, Kerem Cos ¸ ar, Thomas Dohmen, David Dorn, Christian Dustmann, Jonathan Eaton, Marcela Eslava, Teresa Fort, Emilio Gutierrez, Beata Javorcik, Ruixue Jia, Krisztina Kis-Katos, Jann Lay, Sebastian Sotelo, John McLaren, Josef Zweimüller, and audiences at the Aarhus-Kiel Workshop, the CESIfo, CBRT Policy Evaluation Workshop, ECINEQ 2021 Meeting, IZA workshop on “Labor Markets and Innovation During Times of War and Reconstruction,” Warwick University “Political Violence, and Labor Market Outcomes Workshop,” TIGN-Bogotá, CAED-Ann Arbor, G”ottingen, Hamburg-GIGA, IZA (Bonn), University College London, the World Bank, Johns Hopkins SAIS, University of Virginia, SHUFE (Shanghai), CUHK (Hong Kong), and UNSW (Sydney) for helpful comments. Victor Eduardo Zapata Garcia and Elainia Gupta provided valuable research assistance. A supplementary online appendix is available with this article at The World Bank Economic Review website. C The Author(s) 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited. The World Bank Economic Review 633 developing world, often resulting in violence that lacks clearly defined actors and boundaries. The result- ing violence may further hinder economic development by distorting incentive mechanisms and affecting the participation decisions of economic agents or by obstructing the efficient allocation of resources. For example, in the Mexican city Ciudad Juárez, 283 homicides were reported per 100,000 inhabitants in 2010. In neighboring El Paso, Texas, the number was just 0.8 homicides per 100,000. The distance be- tween the two cities is only a few miles, but the levels of violence are orders of magnitude apart. Aside Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 from the direct consequences of violence on the people involved, how does a violent and conflict-afflicted environment matter for firms and industrial development? Shortly after 2007 there was a marked increase in drug-related violence in Mexico. The number of intentional homicides surged by nearly 200 percent from 2007 to 2010, a rise attributed to unforeseen and unintended consequences of a significant shift in the government’s approach to drug enforcement (Dell 2015). Further, the rise in violence was exacerbated by an externally induced rise in cocaine prices during the same period (Castillo, Mejia, and Restrepo 2020).1 The scale of violence escalated to the point where, by 2010, the drug-trade-related violence in Mexico, commonly referred to as the Mexican Drug War, surpassed the conflicts in Iraq and Afghanistan, becoming one of the most violent conflicts of the 21st century. Remarkably, civilians in Juárez faced a higher risk of being killed than those in Baghdad, Iraq (Mora 2009).2 Despite the harrowing nature of this period and its significance, scant attention has been given to assessing the economic repercussions at the firm level. It is important to identify characteristics of firms that make them vulnerable to violent conflicts, and the channels through which they are affected, to inform targeted industrial policies to minimize the negative impact of these episodes. This paper addresses this gap by investigating the effects of violent conflicts on manufacturing firms in an emerging economy using the escalation in drug-trade-related violence in Mexico. I leverage longitudinal plant-level data covering all of Mexico for the period 2005–2010 and utilize the spatiotemporal variation in the homicide rate due to the Mexican Drug War. Focusing on the spatial vari- ation within a country enables me to isolate the impact of violence from the macroeconomic environment and institutional settings that typically co-vary with violence across countries. To derive causal effects of a violent and conflict-afflicted local environment on manufacturing firms, the empirical strategy exploits within-establishment variation over time and across local labor markets as the longitudinal data allow me to control for observable and unobservable differences between firms and local labor markets that may confound the estimates using plant fixed effects. The period of analysis is characterized by substantial variation in violence over time and among metropolitan areas across the country (see fig. 1). Violence, measured by the homicide rate, may be influenced by other factors than the plausibly exogenous driver which is the Mexican Drug War, such as local income or labor-market shocks, which might confound causal interpretations.3 , 4 To isolate the causal impact, I employ an instrumen- tal variable approach, leveraging key triggers of escalated drug violence: (a) The federal–state military 1 Angrist and Kugler (2008) highlight the role of demand-related factors in fueling violence, demonstrating that a plausibly exogenous rise in cocaine prices can incite violence in Colombia. 2 For context, the homicide toll in Mexico in 2010 exceeded the combined tolls of war-torn Iraq and Afghanistan. In 2010, Mexico recorded 26,000 homicides, while the Iraq Body Counts reported 4,167 civilian violent deaths, and Williams (2012) notes 2,777 civilian and 711 soldier deaths from violence in Afghanistan during the same period. 3 In a Beckerian model of rational utility, changes in labor-market opportunities affect crime participation, particularly in property crimes, as empirically supported by Raphael and Winter-Ebmer (2001). Draca and Machin (2015) conclude in their review that relative labor-market opportunities are less likely to be a significant determinant of violent crime or intentional homicide. Conversely, recent studies by Dix-Carneiro, Soares, and Ulyssea (2018) and Dell et al. (2019) show the influence of trade-induced labor-market conditions on violence. This paper’s findings remain robust when controlling for trade exposure in local labor markets. 4 Dube and Vargas (2013) highlight the role of local income shocks in armed conflict intensity. Such shocks could cor- relate with plant-level outcomes and conflict intensity, potentially skewing results. This study addresses this concern 634 Utar Figure 1. Homicide Rates across Selected Metropolitan Areas. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: The number of homicide occurrences and population information are sourced from the Mexican National Institute of Statistics and Geography (INEGI). Note: The population figures in the figure titles correspond to the year 2010. Homicide rates are calculated using annual population figures and represent annualized monthly rates of homicides. The x-axis scale and labels of the top graphs correspond to those of the bottom graphs. The World Bank Economic Review 635 collaboration targeting drug trafficking organization leaders (kingpin strategy), and (b) the surge in drug enforcement in Colombia, which reduced cocaine supply. The former fragmented the drug trafficking organizations (DTOs) and sparked violence, particularly where military operations were successful (Dell 2015; Espinoza and Rubin 2015; Merino 2011), while the latter intensified violence by elevating cocaine prices and the associated rents, fueling competition among DTO offshoots (Castillo, Mejia, and Restrepo 2020). Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 To rule out possible confounding effects of the Great Recession and other industry-specific shocks, the analysis controls for detailed industry-specific aggregate shocks, and the findings are robust to using product-by-year fixed effects. The findings also hold after controlling for a wide range of local labor- market, border, or state-specific time-varying economic and financial factors and employing a battery of alternative instruments commonly used in the literature. Results based on the instrumental variable strategy show that a surge of violence in a metropolitan area leads to a significant decline in plant-level output, employment, and capacity utilization. A doubling of the homicide rate in a metropolitan area causes an 8 percent decline in plant-level output and a 5 percent decline in employment, an impact that is neither temporary nor short term, as the violence of the drug war causes reduced product scopes and chance of survival. The estimates show that the Mexican Drug War accounts for about a quarter of all plant exits over the sample period. Analyzing job types within firms uncovers a stark contrast between the level of resilience among blue- collar versus white-collar employment. The negative employment effect is entirely driven by blue-collar employment, causing a significant increase in the share of white-collar employment within firms. At the same time, unchanged average wages at the plant level obscure the opposing impacts of violence on wages by job type: a significant negative effect on average wages of white-collar employees, and a significant positive effect on average wages of blue-collar employees. The findings show that the negative effect on blue-collar employment is similar whether measured in hours worked or in the number of workers, and it is driven by payroll employment as opposed to temporary workers. These results are consistent with a violence-induced labor-supply shock affecting particularly blue-collar workers. This labor-market channel is particularly strong in plants with lower-wage, labor-intensive, and particularly female-intensive workforces. The firm-level results on employment and wages are consistent with the regional- and household-level studies on the Mexican Drug War that show the costs of drug-war violence were disproportionately borne by low-income neighborhoods (Ajzenman, Galiani, and Seira 2015; Jarillo, Magaloni, Franco, and Robles 2016). Dell (2015) examines the impact of the change in the Mexican government’s drug enforcement policy on violence and drug trafficking. She establishes a causal relationship between drug crackdowns and increased violence and finds that drug crackdowns were not effective in decreasing drug trafficking activities. Although Dell (2015)’s primary focus was not on the economic impact of the drug war, her supplementary analysis, utilizing confidential data on drug trafficking routes, revealed that the drug war adversely affected female labor-force participation but not male. Studies also show the Mexican Drug War has a negative effect on housing prices (Ajzenman, Galiani, and Seira 2015), school attendance, and performance (Jarillo et al. 2016) , and labor-market participation (Robles, Calderón, and Magaloni 2013; Velásquez 2020).5 My paper contributes to and complements this literature by bringing the firm- level consequences of the Mexican Drug War into the light. The firm-level results are consistent with production workers, particularly women who are at the lower end of the wage distribution, dropping out of the labor force disproportionately or transitioning to economic activities that can be conducted at home, as the risk of exposure to violence outweighs the benefit of working. These findings show that by focusing on the Mexican Drug War’s exogenously increased violence and controlling for crop production, precious metal extraction, and oil production at the local labor-market level. 5 Ashby and Ramos (2013) find no association between manufacturing foreign direct investment (FDI) and the Mexican Drug War. 636 Utar firms are affected obliquely through the labor market in an interesting mechanism opposite to most other economic shocks (that hit the firms first and, by consequence, the labor market). In the case of the violence shock, results suggest that it is the other way around. The Mexican Drug War affects firms’ employment and output not only through the local labor mar- ket. The findings show that firms in international trade and firms with more geographically diversified sales or purchases have lower output elasticity with respect to violence.6 Output resilience to violence Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 is also higher among capital-intensive, high-wage, and bigger firms, limiting the affect of drug violence on aggregate output. The disproportionate impact on domestic-market-focused plants suggests a limited role of international trade in acting as a deterrent to violence and also speaks into a nascent literature studying the linkages between globalization and civil war (McLaren 2004; Martin, Thoenig, and Mayer 2008). The literature linking conflict and crime with economic outcomes primarily focuses on aggregate out- comes such as regional income or stock-market returns (Abadie and Gardeazabal 2003; Guidolin and La Ferrara 2007; Pinotti 2015).7 Understanding how an economy reacts to violence and organized crime, and how permanent the effect can be, requires identifying channels through which organized crime and violence impact an economy. Micro-level empirical studies can zoom in on the way firms’ and workers’ behaviors interact with violence and potentially shed light on these channels. Of recent firm-level studies, Ksoll, Macchiavello, and Morjaria (2023) examine the aftermath of escalated ethnic violence, following Kenya’s 2007 presidential election, on flower firms, revealing significant negative effects on exports, pri- marily attributed to increased worker absences. Rozo (2018) uses micro-data from Colombia and shows that the reduction in violence following President Uribe’s election led to market expansion, and Klapper, Richmond, and Tran (2013) focus on civil unrest in Cote d’Ivoire following the coup d’etat in 1999 and find that the conflict led to a drop in firm productivity. Amodio and Di Maio (2018) study Palestinian firms during the Second Intifada and show that firms were affected by the conflict indirectly via border closure, causing a decline in the use of imported materials.8 In this paper, I move the focus to an emerging country with relatively developed institutions and ex- tensive data and show that a violent environment has very heterogeneous effects on firms, significantly affecting the resource reallocation between firms. The findings show employment resilience to violence de- creases among smaller, lower-paying, female-intensive plants. Within firms, employment decline is driven by blue-collar workers as opposed to higher paid, white-collar workers. At the same time, output resilience to violence is higher among capital-intensive, diversified, exporting and importing firms. As the impact is disproportionately borne on lower-paying, female-labor intensive, locally selling, and locally sourcing manufacturing establishments, despite its limited short-run aggregate output consequences, drug violence affects the long-run development of domestic industrial capability in affected areas. Laws and institutions shape an economy’s environment and incentives, either facilitating or impeding productive activities. A growing body of literature delves into the economic consequences of weak local state institutions, lawlessness, and more recently, organized crime (Acemoglu, Robinson, and Santos 2013; 6 Gorrín, Morales-Arilla, and Ricca (2023) examine the effect of drug violence in Mexico on export growth, finding a negative impact. However, lacking firm-level data on output, input, employment, and plant locations, Gorrín, Morales- Arilla, and Ricca (2023) are unable to assess the relative significance of the violence’s impact on exports versus domestic sales or distinguish the violence-induced labor-supply impact. 7 Abadie and Gardeazabal (2003) show that economic outcomes and stock-market returns in the Basque Country were negatively affected by the outbreak of terrorist events. Similarly, Pinotti (2015), using synthetic control methods, finds lower GDP per capita in southern Italian regions exposed to organized crime. On the other hand, Guidolin and La Ferrara (2007) emphasize that violence is not necessarily perceived as negative by investors by showing that Angolan diamond-firm returns were actually hurt by the end of civil war. 8 More recently, Korovkin and Makarin (2021) and Del Prete, Di Maio, and Rahman (2023) have highlighted the role of production networks and market competition, respectively. The World Bank Economic Review 637 Besley, Fetzer, and Mueller 2015; Acemoglu, ˘ De Feo, and De Luca 2020; Alesina, Piccolo, and Pinotti 2018). For example, Alesina, Piccolo, and Pinotti (2018) show how organized crime uses violence to undermine political support for anti-crime initiatives. Globally, organized crime, predominantly linked to the illegal drug trade, is invariably associated with violence. Contributing to this literature, this paper illustrates the detrimental effects of a violent environment due to organized crime on manufacturing in an emerging economy setting and highlights the significant role of local labor markets in affecting firm Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 outcomes and organization. The remainder of this paper is structured as follows: The next section lays out the framework of the empirical analyses with background information on the history of organized crime in Mexico and the drug war. It also describes the data and presents a number of key facts on drug-war areas and firms located in the affected areas. The empirical strategy is explained in Section 3. Section 4 presents and discusses findings on firms’ output, employment, capacity utilization, product scope, and within-firm compositional shifts. Section 5 examines heterogeneity in firms’ responses at both the intensive and extensive margins and explores the channels through which drug violence affects firms. A number of robustness analyses are discussed in Section 6, followed by concluding remarks in Section 7. Supplemental analyses and a detailed description of the data sets are provided in the supplementary online appendix. 2. Violent Conflict and Firms: Sources of Variation and Measurement 2.1. Organized Crime in Mexico—A Brief History Organized crime in Mexico is centered on the transit of illegal drugs into the United States, which is the world’s largest cocaine market, with an estimated value of 38 billion USD in 2008 (World Drug Report 2010). Due to its 1,969-mile-long border with the United States, the largest cocaine market in the world, Mexico has been an ideal location for drug trafficking.9 Starting in the 1970s, the popularity of cocaine grew in the United States, and criminal organizations began to gain more power and influence on a national level in Mexico.10 Mexican organized crime groups possess crucial competitive advantages, notably rapid and efficient transit routes within Mexico, coupled with connections to cocaine suppliers in Central America and con- sumers in the United States. Beyond their dominance in the US cocaine market, Mexican DTOs also exert control over the majority of marijuana, heroin, and methamphetamine supply.11 Their activities within the United States are primarily limited to drug trafficking, with minimal involvement in other forms of illicit business (Finckenauer, Fuentes, and Ward 2001). 2.2. Change in the Drug Enforcement Policy and Subsequent Surge of Violence—Identifying Variation Prior to the mid-2000s, Mexico’s anti-drug efforts primarily centered on eradicating marijuana and opium crops in mountainous regions. Following President Calderón’s election in December 2006, the Mexican 9 In 2008, an estimated 500 metric tons of pure cocaine was on the market, with 480 metric tons consumed during that year. Of these, the United States accounted for 165 metric tons of pure cocaine consumption, while the entire North American market collectively consumed 196 metric tons. The second largest market was the Western European market (EU and EFTA), which, all together, consumed 124 metric tons (World Drug Report 2010). 10 Two major trafficking routes to the United States were used in the 1970s: the Caribbean and Mexico. When the United States gained control over the Caribbean route in the 1980s, Mexican DTOs gained more influence by controlling the major cocaine transit route to the United States. According to the US State Department’s 2013 International Narcotics Control Strategy Report (INCSR), more than 90 percent of the cocaine seized in the United States has transited the Central America/Mexico corridor. 11 Cocaine itself constituted 40 percent of the total illicit drug market share (Kilmer et al. 2014). In 2000, 73 percent of the net coca cultivation was performed in Colombia (National Drug Control Agency 2015). Other source countries are Bolivia and Peru. 638 Utar government radically shifted its strategy in combating powerful drug cartels from ineffective crop erad- ication programs to a proactive pursuit of cartel leadership, known as the “kingpin strategy” with the purpose of reducing organized crime and violence within the country.12 The kingpin strategy was developed by the US Drug Enforcement Administration (DEA) in 1992. It aimed to target and eliminate, either through capture or elimination, the commanders, controllers, and key leaders of major DTOs. Commencing in December 2006, the Calderón administration orchestrated Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 large-scale military deployments for federal–state joint military operations (Operativos Conjuntos Mil- itares). Figure S1.1 in the supplementary online appendix shows the states with military operations. This concerted effort achieved notable success in dismantling major criminal organizations by apprehending key leaders or neutralizing them in the course of arrest operations.13 Paradoxically, despite the success of the new strategy in weakening the major cartels, it also had the unfortunate and unanticipated consequence of increased violence. Killing and capturing DTO leaders sparked internal conflicts for influential and lucrative leadership positions within these organizations, pit- ting different factions against each other. As the organized crime groups splintered and power dynamics shifted within the cartels, violent confrontations erupted as they vied for control over the drug routes previously dominated by now-weakened rivals.14 Within just a few years following the start of the drug crackdowns, there was a significant proliferation of DTOs, as factions from some existing DTOs estab- lished new criminal organizations (Bagley and Rosen, 2015). An additional factor that potentially exacerbated the surge in violence after 2008, as documented in the literature, is the decline in the cocaine supply in the market. Castillo, Mejia, and Restrepo (2020) show that heightened government seizures in Colombia, the primary source country for Mexican drug cartels, played a significant role in the reduction of cocaine supply. This, in turn, resulted in increased cocaine prices in the United States and a corresponding increase in drug-related violence, especially in areas around the strategic drug trafficking routes to the US market.15 Consequently, after decades of maintaining relatively stable rates of violent crime, nationwide homicide rates in Mexico nearly tripled from 2007 to 2010 (see fig. S1.2 in the supplementary online appendix). However, not every region of Mexico was equally impacted by this sudden upswing in violence. My spatial unit of analysis is the metropolitan area, comprising an employment core and the surround- ing areas that have strong commuting ties to the core.16 This enables an analysis based on well-defined local labor markets rather than administrative divisions, thereby mitigating the potentially confounding influence of urban and rural disparities on the findings. Furthermore, metropolitan areas are less suscep- tible to spillover effects due to the prevailing absence of shared boundaries. Figure 1 shows homicide rates within selected local labor markets (metropolitan areas), and fig. S1.3 in the supplementary online appendix presents the overall evolution of violence across all metropolitan areas. Notably, there exists a substantial spatial variance in homicide rates. A comparison between fig. S1.1 and fig. S1.3 reveals that states hosting military operations are more likely to encompass metropolitan areas that subsequently witness a surge in violence. Figure S1.4 confirms this. This escalation of violent conflict, 12 Despite the fact that DTOs are not cartels in the sense that they do not control prices by colluding, the term “drug cartel” is used colloquially to refer to DTOs. Drug cartels and DTOs are used interchangeably in this paper. 13 The average annual number of troops assigned for battling drug trafficking increased 133 percent to 45,000 during the Calderón administration compared to the preceding Fox administration (Grayson 2013). 14 See Lindo and Padilla-Romo (2018) for a study illustrating that the kingpin strategy led to a roughly 60 percent increase in the homicide rate. 15 Cocaine production in Colombia decreased 43 percent from a potential 510 pure metric tons in 2006 to 290 pure metric tons in 2009 (National Drug Intelligence Center (2011)). 16 INEGI, in collaboration with the National Population Council (CONAPA) and the Ministry of Social Development (SEDESOL), established 59 such local labor markets. The World Bank Economic Review 639 plausibly exogenous to local market conditions, allows me to study the causal impact of an increase in violence in the local environment on detailed establishment-level outcomes. 2.3. Drug Violence as a Local Disamenity Shock News reports highlight two primary ways the drug war directly endangers workers’ lives. First, there is the Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 risk of direct harm or involvement in drug trafficking, driven by the substantial profits of drug cartels in the United States, estimated to be between 19 and 29 billion USD (source: Mexico Drug War Fast Facts– CNN Library). This financial incentive may lead economically disadvantaged workers to engage in drug trafficking. Second, workers, particularly those in impoverished neighborhoods, risk becoming unintended casualties of either drug gangs or government forces. Due to their location, they can more easily find themselves in harm’s way, as reported in various news sources (see, e.g. Cardona 2010, and supplementary online appendix S1). Recent studies also confirm the disproportionate effect of the Mexican Drug War on poorer neighborhoods Ajzenman, Galiani, and Seira 2015; Jarillo et al. 2016). Figure S1.5 in the supplementary online appendix shows that unskilled production workers, who tend to be lower paid, are far more likely to be victimized during the drug war than other typical occupations within manufacturing. As workers facing an increased risk of being a direct or indirect target of either DTOs or military/police forces, it is natural to expect an effect of such violence on labor-market dynamics. Indeed, by inflicting fear and anxiety, crime changes daily routines and behavior, and unsafe commuting and the likelihood of wit- nessing violence are likely to be important factors affecting the decisions of broader worker populations to participate in the labor market. The literature finds that such indirect costs of crime are indeed larger than the direct costs (Becker and Rubinstein 2011; Dustmann and Fasani 2016). In a related context, Melnikov, Schmidt-Padilla, and Sviatschi (2020) find that the presence of gangs in El Salvador increases costs of mobility and restricts labor choices for people who live in neighborhoods controlled by gangs. This emphasizes the broader implications of such violence on labor-market dynamics. Lower-wage, blue-collar workers may be especially susceptible to commuting risks, as they often live in relatively unsafe areas and travel during nights and early mornings to match production shifts, as also suggested by fig. S1.5 in the supplementary online appendix. The risk of commuting also rises with the length of the commute, and Brueckner, Thisse, and Zenou (2002) demonstrate in a model that connects worker skills with urban spatial structures that lower-skilled workers tend to live farther from their work- places. Women’s participation in the labor market may be particularly sensitive to increased commuting risks, given their generally more elastic labor supply. Additionally, heightened criminal activity in a neigh- borhood can compromise children’s safety at school, leading to increased absenteeism, as documented by Jarillo et al. (2016). This, in turn, can escalate parental responsibilities at the expense of labor-market engagement. Figure S1.6 in the supplementary online appendix shows the evolution of manufacturing employment in the same metropolitan areas presented in fig. 1. The manufacturing employment either declined or stayed constant between 2005 and 2010 in all of the highly exposed metropolitan areas, whereas all four of the similarly sized non-exposed metropolitan areas experienced a net increase in manufacturing employment over the same period. 2.4. Data and Descriptive Analysis The primary data set utilized in this study is the Encuesta Industrial Mensual Ampliada (EIMA) 2005– 2010, a comprehensive monthly survey of plants conducted by INEGI, covering approximately 90 per- cent of the manufacturing value added across the nation. The survey includes all plants in Mexico with more than 300 employees. For smaller plants, the inclusion criteria are as follows: within each detailed manufacturing activity, clase, plants are sorted by their production capacity based on the Economic Cen- sus 2004, and the survey continues from the top until at least 80 percent of production in each specific 640 Utar product category is represented.17 This survey method naturally includes a greater proportion of larger plants, which is a common aspect of such surveys. In the following sections, I illustrate that smaller plants are particularly affected by a violent environment. Consequently, the estimates presented here should be viewed as a conservative estimate of the actual impact. I focus on plants located in metropolitan areas which capture 85 percent of the manufacturing employ- ment in EIMA. Table S2.1 in the supplementary online appendix reports the distribution of plants across Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 57 metropolitan areas in the sample.18 Table S2.2 in the supplementary online appendix presents summary statistics for this sample. The average plant employs 238 workers and produces three product varieties.19 On average, firms employ one non-production (white-collar) employee for every two blue-collar workers. Figure S2.1 shows the distribution of plants in the initial year of 2005 across the three-digit industries. The sample covers a wide variety of plants, and the distribution of plants across industries reflects the overall pattern of Mexican manufacturing with a relatively high share of food manufacturing, plastics, chemicals, non-metallic mineral products, and automotive (transportation equipment) sectors. I match EIMA with the annual survey of manufacturing plants, Encuesta Industrial Anual (EIA), which provides detailed balance sheet information of the same manufacturing plants before the drug-war period of 2003–2007. As both EIA and EIMA are based on the same survey design and are run in parallel, 90 percent of the plants surveyed in EIMA can be matched with EIA.20 Maquiladoras, which are export- processing plants typically owned by foreign companies and supplying into the US market, are not part of either EIMA or EIA.21 While entry of new plants is not observed due to the fixed survey design, exit is observed in the data at a monthly frequency as the exiting plants drop from the sample. For detailed technological and organizational pre-shock characteristics, I also utilize Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005, which is a representative establishment-level survey on technological and organizational capabilities of plants. Detailed technological and employee characteristics obtained from this nationally representative survey are mapped at the four-digit industry level to EIMA, the main data set used in the analysis.22 Plants in the ENESTyC report geographic distribution of their annual sales as well as their use of imports from across the world. I use this information to construct entropy measures of sales and input diversification and study heterogeneity of the output elasticity of violence with respect to firm diversification. I begin by examining data to understand how local labor markets susceptible to heightened violence correlate with firm attributes. As a first step, I compute the mean homicide rates and numbers before and after the drug war for each labor-market area during 2005–2006 and 2008–2010. Labor-market areas 17 The manufacturing sector is divided into 230 economic activities, or clases, each identified by a unique six-digit num- ber. For example, 311320 refers to “Preparation of chocolate and chocolate products from cacao,” and 311330 to “Preparation of chocolate products from chocolate.” 18 There are 59 designated metropolitan areas in Mexico as of 2010. Plants in EIMA operated across 58 metro areas. Puerto Vallarta, which is a beach resort area primarily driven by tourism, is the only metro are not covered by EIMA. Additionally, the analysis excludes the metropolitan area that was affected by the Tabasco flood. In late 2007, a major flood struck the state of Tabasco, affecting over a million residents. The state capital faced bankruptcy, and numerous businesses were adversely affected. Given that this event likely influenced the opportunity cost of crime, plants in the flood-affected area are not included in the analysis. 19 Throughout the paper, a product variety refers to 9-digit SCIAN products, e.g. “Chocolate covered raisins produced from purchased chocolate,” (SCIAN 311330025). 20 Unfortunately, EIA was replaced with a new survey based on a new sampling in 2008; therefore I rely on EIA for initial, pre–drug war, characteristics of the plants. 21 Maquiladoras are not considered as part of the domestic manufacturing industry as they have been subject to a different legal framework. INEGI has carried out a separate survey for them (see Utar and Torres Ruiz 2013 for more details). 22 In principle, plants surveyed within ENESTyC can also be matched with the plants in EIMA. However, the resulting data set is relatively small and significantly biased toward big plants, hence the choice of utilizing this data set at the industry level. The World Bank Economic Review 641 Table 1. Pre-shock (2005) Plant Characteristics High-intensity Other metropolitan drug-war metros areas Plant-level variables Mean SD Mean SD Diff. t -stat Log output 11.31 1.99 11.22 1.95 0.09 1.28 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log N of employees 4.57 1.33 4.56 1.31 0.01 0.26 Log capital per worker 5.00 1.41 4.84 1.42 0.16∗ 2.91 Log labor productivity −1.09 1.12 −1.14 1.15 0.04 1.05 Capacity utilization rate 74.00 18.78 70.63 20.12 3.37∗ 4.53 N of varieties 3.05 2.85 3.21 3.12 −0.16 −1.42 Export dummy 0.42 0.49 0.34 0.47 0.08∗ 4.49 Import dummy 0.48 0.50 0.48 0.50 0.00 −0.08 Share of payroll workforce 0.88 0.31 0.89 0.30 −0.01 −0.93 Homicide rate 12.16 6.52 7.35 6.35 4.82 1.75 Source: Data on import and capital per worker are from the Encuesta Industrial Anual (EIA); other data are from the Encuesta Industrial Mensual Ampliada (EIMA) Note: Values are measured in 2010 thousand Mexican pesos. Labor productivity is measured as the value of production per hour unit of labor. There are 908 plants in the six metropolitan areas defined as “high-intensity drug-war zones” and 4,575 in “other locations.” ∗ indicates significance at the 5 percent level or below. are classified as high-intensity drug-war zones if their pre- and post-period homicide rate and number dif- ferences exceed the mean differences. This categorization identifies six such zones: Acapulco, Chihuahua, Juárez, La Laguna, Monterrey, and Tijuana. While the empirical analysis utilizes a continuous measure of violence exposure, this discrete classification helps in highlighting potential inherent differences between plants in areas affected by the drug war and those in unaffected areas prior to the conflict. Table 1 presents the characteristics of plants in these two areas as of 2005. The average plant size, whether measured by output value or employment, shows little variation between the two groups. Simi- larly, there are no notable differences in labor productivity or the number of product varieties per plant. Violence-exposed areas are, on average, closer to the US border, and as a result significantly more plants export in areas affected by heightened violence after President Calderón’s launch of the war on drug car- tels. The average importing likelihood, in contrast, is the same across both regions. Table 1 also shows that plants in regions soon to be impacted by drug-war violence tend to be more capital intensive and have a higher utilization rate, likely linked to a greater proportion of exporters in these areas. Finally, the table also reports that the average homicide rate was higher in soon-to-be-affected areas, though not by a significant margin. The six highly exposed metropolitan areas are key sites for manufacturing activities, accounting for 21 percent of the total manufacturing employment compared to the other metropolitan areas.23 Figure S2.2 in the supplementary online appendix shows the distribution of plants in 2005 across three-digit industries, separately, in the six highly exposed metropolitan areas and in the other metropolitan areas. Food manufacturing constitutes the largest manufacturing sector in both areas (as in Mexico overall), and there is no substantial difference in the industry specialization patterns across the two areas. 3. Empirical Strategy This section outlines the empirical approach used to identify the impact of heightened violence on plant- level outcomes. Leveraging a longitudinal data set at the plant level enables a focus on variations within plants while mitigating potential bias arising from unobservable plant and location characteristics affect- ing the results. 23 Author’s calculation using EIMA. 642 Utar Consider the following estimation equation at the plant-year level: ln Yikjt = α0 + α1 Violencejt + Xjt + τkt + ηi + ikjt . (1) In this equation, Yikjt denotes the outcome of plant i in industry k within local labor market j in year t . The variable Violence jt stands for the natural logarithm of the intentional homicides per thousand individuals within the local labor market j.24 The variable Xjt is a vector of time-varying local labor- Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 market characteristics, while τkt denotes industry-by-year fixed effects. Lastly, ηi denotes plant fixed effects that can be correlated with plant or metropolitan area characteristics. By making comparisons within a plant over time, observable and unobservable time-invariant char- acteristics, such as productivity and technology differences across firms, or location characteristics that make the location less or more attractive to legal and illegal businesses (e.g. infrastructure, ports, and economic development), are controlled for. The vector Xjt includes pre-existing trends in homicide rates across local labor markets. For this, time dummies are interacted with the year 2002 level of homicide rates of the local labor markets. Furthermore, Dube and Vargas (2013) study how different types of commodity shocks affect civil-war outcomes and show that a sharp fall in coffee prices during the 1990s in Colombia led to an increase in violence differen- tially in municipalities cultivating more coffee by decreasing the opportunity cost of conflict. To account for potential confounding effects arising from such shocks, the vector Xjt also includes metropolitan-level employment shares of crop production. Moreover, an increase in oil prices (or precious metals) can inten- sify conflict by increasing contestable income. Consequently, a positive oil price shock would likely have a positive impact on both a local economy reliant on oil production and conflict intensity. This could lead to an underestimation of the violence effect on plant-level outcomes. To proxy for such local shocks, I use metropolitan-level employment shares of oil and natural gas extraction as well as metal mining including gold, silver, copper, and uranium. I show below that the inclusion of these variables does not significantly affect the results. Industry-by-time fixed effects account for aggregate changes affecting manufacturing establishments similarly, and also industry-specific time trends that may affect certain regions disproportionately, perhaps due to a potential geographic concentration of industries. It is especially important to take industry-specific business trends into account due to the possible differential impact of the Great Recession. For this reason, equation (1) controls for trends for each five-digit manufacturing industry (168 of them in the data). These industries are narrowly defined and can be considered product lines.25 Moreover, standard errors are allowed to have arbitrary patterns of correlation within each metropoli- tan area, and also separately within each four-digit industry, and are two-way clustered for each metropoli- tan area and industry. Equation (1) allows an estimation of violence elasticity, α1 , that is based on the variation in within-plant outcomes specific to local markets that experienced heightened violence over 2005–2010. 3.1. Instrumental Variable Strategy The variation in homicide rates during the sample period, primarily attributed to the drug war, may also be affected by other factors, potentially linked to plant-level outcomes. For instance, an area’s expanded productive capacity might attract unskilled migrants, potentially driving socioeconomic disparities that could, in turn, escalate crime rates. Additionally, the intensity of drug-related violence might be influenced 24 Throughout the estimation analysis, the homicide rate refers to the number of homicides per thousand inhabitants, instead of the conventional per hundred thousand inhabitants. 25 Some examples of five-digit industries are the following: “Manufacture of cement for construction,” “Concrete manu- facturing,” “Manufacture of cement and concrete pipes and blocks,” “Manufacture of prestressed products,” “Prepara- tion of breakfast cereals,” “Manufacture of chocolate and chocolate products from cocoa,” “Manufacture of chocolate products from chocolate.” The World Bank Economic Review 643 by the performance of local plants. To eliminate the concern that homicide rates might be correlated with the error term, thereby ensuring that the findings are attributed to the plausibly exogenous surge in violent conflict resulting from unforeseen consequences of the policy shift in Mexico and intensified drug enforcement in Colombia, I adopt an instrumental variable (IV) approach. This involves constructing an instrument grounded in these two pivotal factors igniting the Mexican Drug War. When the Calderón administration opted to deploy military force against drug cartels in 2007, Mexi- Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 can states were given the option to participate in joint military operations with federal forces against these criminal groups (operativos conjuntos militares). Eight of the 32 states chose to support the federal mil- itary initiatives. Figure S1.1 in the supplementary online appendix illustrates the states that collaborated with the federal government to carry out joint military operations targeting drug cartel leaders.26 These states largely overlap with the principal drug trafficking corridors. The deployment of the federal army in these states serves as an indicator of the kingpin strategy’s execution and, consequently, the inadvertent spike in violence, given the military’s primary role in enacting this strategy. Federal army operations resulted in captures or killings of drug-cartel leaders, and that in turn triggered fights between cartels (Dell 2015; Lindo and Padilla-Romo 2018). Figure S1.4 shows the homicide rate increased dramatically after 2007 in states with federal military operations as opposed to other states, and the increase in homicide rate was driven by drug-related homicides.27 Let MOis be an indicator for plant i if they are located in one of the eight states that collaborated with the federal government’s military operations. Interacting it with the start of the policy turn, 1(Year ≥ 2007), generates MOis t , which is zero for all plants before 2007 and takes the value 1 on and after 2007 for plants in areas that implemented the military-led drug crackdowns.28 The indicator MOis t helps to isolate the spatiotemporal variation in the homicide rate that can be attributed to the intensified violence due to the military operations. Beginning in 2000, Colombia enacted policies to curb coca cultivation and prevent drug shipments from leaving the country (Mejia and Restrepo 2016). These measures significantly cut down the cocaine supply, particularly in the latter half of the 2000s. The decline in cocaine supply from Colombia and the resulting change in cocaine prices intensified drug violence by increasing rent opportunities (Castillo, Mejia, and Restrepo 2020; Angrist and Kugler 2008).29 To capture the time variation in the strength of Colombian drug enforcement, I use the cocaine seizures in Colombia, normalized with the annual cocaine- cultivated land in Colombia. Figure S4.4 in the supplementary online appendix plots this measure and shows it increases significantly after 2008. Interacting this time-varying variable with the locations sus- ceptible to violence outbreak in Mexico due to the government’s kingpin strategy, I obtain the instrument coke MexColis t ≡ MOis t × DECt . (2) coke Here, DECt measures the annual amount of cocaine seized by Colombian forces, normalized by the annual amount of net cocaine-cultivated land in Colombia. It captures the time variation in the strength of the Colombian drug enforcement agencies. Violence is expected to affect plant-level outcomes not necessarily because death tolls decrease the number of available workers, but because increased violence as proxied by the homicide rate potentially affects the incentive structure of economic agents and their optimal behavior. Additionally, homicides are recorded as they occur, but the occurrence date is not necessarily the date at which the news about the incidence of violence reaches the local population. To further break down any endogeneity concern 26 The states that participated were Michoacán, Guerrero, Baja California, Nuevo León, Tamaulipas, Chihuahua, Sinaloa, and Durango. The remaining states did not participate. 27 Also see fig. S1.3 for the spatial distribution of urban violence over the years. 28 Fourteen metropolitan areas in the sample are affected by the military-led crackdowns. 29 The dealer-level price of cocaine per pure gram increased between 2005 and 2010 by 46 percent in the United States (author’s calculation from the National Drug Control Strategy data). 644 Utar and allow plants and workers time to react to violence, both the annual homicide rate in Mexico and the cocaine seizures in Colombia are constructed with six months lag from June t − 1 toward June t .30 Moreover, while it is important to establish that the results are not affected by time-varying metropolitan variables of crop, oil, gas, and metal mining sectors, including them may introduce endogeneity concerns. I show below that including these variables does not affect the results, and thus they are not included in the default 2SLS specification. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Assuming a strong correlation between the homicide rate and the instrument which is based on the Mexican and Colombian policy triggers of the drug war, the (untestable) exclusion restriction is valid as long as the Colombian drug enforcement and the Mexican kingpin policy affect the Mexican manufactur- ing plants via their effects on violence conditional on the pre-existing trends in violence, trends and shocks that can vary at the detailed five-digit industries, and plant fixed effects (i.e. E [ ikjt Ijt | Xjt , τkt , ηi ] = 0). To make sure that the exclusion restriction is not violated due to, for example, a possible impact of increased security expenses on manufacturing plants, I augment equation (1) with the growth in local security expenses. The instrumental variable results with the inclusion of the local security expenses pro- duce robust findings (see table S3.1). Note also that if the military-led drug crackdown creates a demand shock for a particular sector, this is taken care of by five-digit industry-by-year fixed effects. It is also shown that the results are robust to the inclusion of product-by-year fixed effects. To ensure that the instrument is not correlated with other state-level policies potentially influencing plant-level outcomes, equation (1) is also augmented with a battery of time-varying state characteristics. First, states joining the military operations may be getting favored transfers from the federal budgets that in turn have an impact on plant-level outcomes. To address this, I control for total federal transfers for states as well as federal transfers that are specific security expenses. I also augment equation (1) with state-level GDP, state-level manufacturing production, as well as state-level public expenditures. These additional state-level time varying factors do not significantly affect either the first stage or the second stage of the results (see tables S3.2–S3.5), suggesting the instrument is not correlated with state-level economic and financial factors that may affect manufacturing plants. To address the possibility that large firms may have affected the probability of military intervention in their area, I also regress the firm size (both in terms of employment and production) on the probability that a firm’s location experiences military intervention, finding no significant relationship.31 4. Industrial Activity and Violent Conflict This section shows that the surge of violence induced by the drug war causes a marked decline in man- ufacturing activities. The drug war also affects the composition of employment and alters wages within establishments. 4.1. Job Losses in Manufacturing due to Drug Violence Table 2 shows the impact of drug-related violence on plant-level employment. Column (1) begins with the ordinary least squares (OLS) estimation of equation (1), revealing a negative and statistically sig- nificant employment effect. Although the drug war introduces a quasi-natural fluctuation in Mexico’s homicide rate, this violence metric also encompasses non-drug-related homicides. Particularly in areas not directly affected by the drug war, this measure is likely to correlate with the local economy’s changing characteristics or shifts in local labor markets over time. Moreover, firms might react differently to the 30 Table S3.24 in the supplementary online appendix reports the concurrent and the lagged effect of the homicide rate at a monthly frequency. 31 The coefficients of interest are −0.001 and −0.007 for employment and production, with t -statistics −0.01 and −0.08, respectively. The full results are available. The World Bank Economic Review 645 Table 2. Drug Violence Decreases Manufacturing Employment (1) (2) (3) (4) Specification OLS 2SLS 2SLS 2SLS Dep. var.: Log employment Log homicide rate −0.024∗∗ −0.069∗∗∗ −0.070∗∗∗ −0.070∗∗∗ Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (0.009) (0.022) (0.022) (0.023) Plant FEs     2002 homicide rate × Year FEs     Time-varying local market characs. – –  – 5-dig. industry × Year FEs    – Product × Year FEs – – –  No. of observations 30,605 30,605 30,605 30,605 No. of clusters (LM) 57 57 57 57 First stage Instrument (MexCol) – 0.395∗∗∗ 0.396∗∗∗ 0.394∗∗∗ (0.086) (0.086) (0.085) Kleibergen–Paap F -excluded instrument – 21.146 21.094 21.722 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (1) by ordinary least squares (OLS) (column 1) and two-stage least squares (2SLS) (columns 2–4). The dependent variable is the logarithm of the number of employees. Log homicide rate is the logarithm of the number of homicides per thousand inhabitants of each metropolitan area. Time-varying local market characteristics include metropolitan area-level employment shares of crop production; metal mining including gold, silver, copper, and uranium; and the metropolitan area-level employment share of oil and natural gas extraction. Robust standard errors, reported in parentheses, are two-way clustered by local market (metropolitan area) and four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. surge in violence from the drug war compared to variations in ordinary homicides, suggesting a potential non-linear relationship between the homicide rate and plant-level outcomes. Therefore, while the OLS estimates might be attenuated, the two-stage least squares (2SLS) estimates, derived using an instrument designed to specifically capture the impact of drug-war violence, are expected to remain robust. In column (2), using MexColis t as an instrument for the logarithm of the homicide rate yields a co- efficient that is both larger and more precisely estimated. This suggests that plant employment is more adversely affected by the increased violence from the drug war than by ordinary crime. It also implies that confounding factors, which positively correlate with both the local economy and non-drug-war-related homicides, may cause an underestimation of the impact of violence in OLS analyses. The first-stage results confirm a strong correlation between the instrument and the homicide rate, with robust instrumentation evidenced by high F -statistics (Kleibergen–Paap F -statistic), underscoring the instrument’s validity (Staiger and Stock 1997). According to the estimate in column (2), a doubling of the homicide rate results in a 4.5 percent reduction in plant-level employment. To address the issues highlighted by Dube and Vargas (2013), column (3) incorporates metropolitan- level employment in crop production, mining of precious metals (gold, silver, copper, and uranium), and oil and natural gas extraction. The inclusion of these time-varying local market characteristics does not alter the observed impact of violence on plant-level employment.32 This is reassuring, since it suggests that the chosen instrument accurately isolates the variation in homicide rates linked to the onset of the Mexican Drug War. To address the uneven impact of the Great Recession across local labor markets in Mexico and miti- gate concerns about confounding factors like trade competition, column (4) incorporates product-by-year fixed effects, alongside plant fixed effects and pre-trends in homicide rates. The 2SLS estimate in column 32 Incorporating time-varying controls for metropolitan sectors like crops, oil, gas, and metal mining could raise endogene- ity concerns. The primary 2SLS analysis, aimed at isolating drug-war triggers, does not include these sector employment shares. Their inclusion in OLS analyses does not modify the outcomes, affirming the instrument’s effectiveness in captur- ing the homicide rate’s variation from 2005 to 2010 attributable to the Mexican Drug War, as table 2 also demonstrates. 646 Utar (4) remains the same, indicating that the impact of drug-war violence is not confounded by the Great Recession or trade competition. State-Specific Threats to Identification Table S3.2 in the supplementary online appendix. shows the results when we additionally include a battery of state characteristics to address state-level confounding factors. These state-level variables are per capita GDP, manufacturing GDP per capita, total transfers Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 from the federal budget per capita, and per capita transfers from the federal budget specific for public security and safety. Due to the unavailability of public finance data for the federal district, the inclusion of these variables (columns 4–8) results in a reduced number of observations, and this partially reinforces the effect of violence. Otherwise, including these time-varying state characteristics does not have a major effect on the results and the first-stage results remain strong. The magnitude of the estimate in column (4) shows that doubling the homicide rate in a metropolitan area leads to a 4.85 percent decline in plant-level employment. Since the nationwide homicide rate tripled between 2007 and 2010 while the aggregate manufacturing employment declined by 7 percent over the same period, the findings imply a significant role of the Mexican Drug War on the aggregate employment decline. 4.2. Violence and Plants’ Production, Utilization, and Productivity The section examines how increased violence affects key aspects of plant-level production, specifically focusing on output (production value), average price, the range of products offered, capacity utilization, labor productivity, and export. Table 3 presents these 2SLS results. In column (1) the dependent variable is the logarithm of the value of production. The analysis reveals that violence from the drug war significantly diminishes manufacturing output. With a magnitude of −0.112, this effect is more pronounced than the employment elasticity observed in table 2, and indicates that doubling the homicide rate decreases plant- level output by close to 8 percent. In column (2), the dependent variable is the average price across products sold per plant, and the results show no significant effect, indicating the impact on the value of production is driven by reduction in actual output rather than reduction in prices. Table 3. Mexican Drug War and Decline in Manufacturing Plants (1) (2) (3) (4) (5) (6) (7) Production Average Product Capacity Labor Export Export value price scope utilization productivity intensity (in log) (in log) (in log) rate (in log) Violence −0.112∗∗∗ 0.037 −0.045∗∗ −4.131∗∗∗ −0.062∗ −0.018 −0.009 (0.033) (0.022) (0.020) (1.071) (0.035) (0.023) (0.010) Plant FEs        Pre-trends in homicide rate        5-dig. industry × Year FEs        No. of observations 30,605 28,589 30,605 29,926 30,605 30,605 30,605 No. of clusters (LM) 57 57 57 57 57 57 57 K-P F -excluded instrument 21.15 20.86 21.15 20.32 21.15 21.15 21.15 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (1) using two-stage least squares (2SLS). Violence is the logarithm of the number of homicides per thousand inhabitants of a metropolitan area. Capacity utilization is the percentage rate of utilization of the fixed assets of the plant. All dependent variables, except Capacity utilization, Export indicator, and Export intensity are in logarithm. Output is the total value of production. Output price is the average unit price of a plant’s product varieties. Labor productivity is the value of output per hour worked. Export is an indicator variable that takes 1 if a plant exports in year t . Export intensity is the share of export revenues over the total sales. Pre-trends in homicide rate is the interaction of year dummies with the year 2002 homicide rate for each metropolitan area. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area and four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. The World Bank Economic Review 647 Column (3) of table 3 presents the effect on the product portfolio of plants. The dependent variable is the logarithm of the number of distinct product varieties that a plant produces. The results show that the reduction in output due to the drug war is accompanied by a significant drop in the number of varieties produced. Since product introductions and re-introductions typically require a significant organizational, capacity, and technological adjustment, the significant negative effect on the product scope of firms sug- gests that the decline in production may have long-term implications. The estimate in column (3) shows Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 a drop in the number of varieties by approximately 3 percent in response to doubling the homicide rate in the metropolitan area. In column (4), the dependent variable is the percentage of capacity utilization, the rate at which man- ufacturing plants utilize their fixed assets.33 The results show a significant negative effect of violence on capacity utilization. The stronger response of output to violence in comparison to employment, to- gether with a significant drop in the utilization rate, implies a drop in plant efficiency. This is confirmed in column (5) of table 3—violence causes reduced productivity as measured by the output per hour worked. Foreign demand might be less affected by the drug war’s disruptions compared to domestic demand, yet challenges such as highway disruptions could still hinder Mexican firms’ international trade. Martin et al. (2008) suggest that international trade could act as a buffer, substituting domestic trade during civil conflicts. In column (6) of table 3, the outcome variable is an indicator variable for exporting. The results show that the likelihood of export is not affected significantly by the drug war. Similarly, the share of export in total sales remains unaffected (column 7). Further results on exported products (see column (5) of table S3.7 in the supplementary online appendix) also reveal that the domestic market drives the decline in the number of products. Contrasting these findings, Gorrín et al. (2023) show a significant decline in Mexican exports in response to the Mexican Drug War. While Gorrín et al. (2023) cannot compare the impact on exports with the domestic revenues, as their data only have export information, my results, based on both domestic and foreign sales at the plant level, are evidence of how domestic demand suffers disproportionately under the violent environment. The following section focuses on the compositional changes in the plant-level workforce to further elucidate the sources of decline in employment. 4.3. Drug Violence Changes Employment Composition Within Firms I have shown above that drug-war violence causes a significant reduction in both plant-level output and employment. The negative effect on employment may be due to a lower labor demand, but drug violence can also affect employment via its effect on the local labor supply. In this section, I study the impact on employment composition and wages at the plant level to understand potential drivers of the employment effect of violence. Figure 2 shows the plot of the results from estimating equation (1) by 2SLS on production and non- production employment and wages and the full results are presented in table S3.6 in the supplementary online appendix. Figure 2a shows the employment elasticity estimates separately for the total number of blue-collar workers and for the total non-production (white-collar) workers. The dependent variables include both employees on payroll and contractual workers that are not on the firm’s payroll. The left- hand side of fig. 2a focuses on the number of white-collar and blue-collar workers and the right-hand side focuses on the same variables measured in hours worked. Figure 2a shows a strikingly asymmetric impact of the drug war on blue-collar versus white-collar (non- production) workers. The drug war leads to a significant reduction in blue-collar employment, whereas the 33 The utilization rate shows the relationship between the volume of production that is currently being obtained and the volume of production that could potentially be generated given the conditions of infrastructure, machinery, equipment, technical, and organizational procedures that are currently used in the establishment. 648 Utar Figure 2. Impact of Drug Violence on Employment Composition and Wages. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (a) (b) (c) (d) Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (1) by two-stage least squares. The log homicide rate is instrumented as described in equation (2). The full results are presented in table S3.6. Bar heights indicate the value of the coefficient estimate for the log homicide rate. The dependent variables are given as the plot and bar titles. Solid frames indicate statistical significance at the 10 percent level or less. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. effect on non-production workers is slightly positive, albeit not statistically significant. The 2SLS elasticity estimate for blue-collar employment in fig. 2a is −0.10 and statistically significant at the 1 percent level. This reduction in blue-collar employment mirrors the magnitude of the impact observed on overall plant- level production. It shows that doubling the homicide rate in a metropolitan area causes a 7 percent decline in the number of blue-collar employees. At the same time, the level of non-production employment is not significantly affected by the drug-war violence. To further elucidate this point, fig. 2a also shows margins of firm’s employment response. Hours re- sponse to a labor demand shock should be at least as pronounced as the response in employment count, since adjustments in total hours can reflect both a reduction in the number of workers and a decrease in hours per worker. Moreover, reducing one’s hours tends to be less costly and more flexible than laying the person off altogether. Interestingly, fig. 2a reveals no differential impact on employee counts and hours worked in respond to drug violence; indeed the difference (if any) runs in the opposite direction. Adjustment costs are also typically higher for payroll employment compared to less permanent, con- tractual workers due to firms’ requirement to pay social security contributions and severance payments The World Bank Economic Review 649 at the termination of a payroll contract.34 So laying off contractual employees is generally less expen- sive, offering firms a more cost-effective initial response to uncertain or temporary shocks. Therefore, when labor-market frictions such as severance payments exist, a more pronounced response is expected in non-payroll employment compared to payroll employment following a labor demand shock. The dependent variables are the blue-collar and white-collar employees on payroll in fig. 2b. As be- fore, the left-hand side of the figure plots the elasticity of employment across the two groups based on Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 employee counts, while the right-hand side of the figure plots the employment elasticity measured in hours. The results show that a negative effect on blue-collar employment is driven by the payroll, permanent, employees. Further, the extent of reduction both in blue-collar hours worked on payroll and the number of blue-collar employees on payroll is similar. Indeed, the point estimates for the blue-collar employment elasticity is smaller when measured in hours worked (−0.11 versus −0.12), which can happen, for ex- ample, if the number of blue-collar workers decreases while the hours worked by the remaining workers increase. So far the results point to a possible violence-induced blue-collar labor-supply shift. To fully understand whether the underlying employment changes are driven by labor-supply shift rather than a violence- induced labor demand reduction, we turn our attention to wages. Figure 2c shows the effect of drug-war violence on plant-level wages. The overall stability in average wages amid increased violence masks a significant rise in blue-collar wages and a decline in white-collar wages. Blue-collar wage elasticity is 0.11, significant at the 5 percent level, meaning that doubling the violence in a metropolitan area leads to a 7.7 percent wage increase for blue-collar workers. For white- collar workers, the elasticity is −0.08, indicating a 5.6 percent wage decrease under the same conditions. Taken together, these findings are consistent with drug violence impacting participation decisions of blue- collar workers and indicate that a violence-induced reduction in labor demand cannot be the sole driver of the decline in blue-collar employment. The increase in blue-collar workers’ wages does not have to be driven by an actual worker-level increase in the wages of blue-collar workers. If it is the lower-wage individuals among blue-collar workers who leave the workforce, the average blue-collar wages at the firm level will increase. But notice that this type of selection cannot explain the changes in the white-collar wages since the drug violence does not have a significant impact on the white-collar employment. Accordingly, the results imply that violence increases the relative wages of blue-collar workers (i.e. it decreases the skill premium.)35 The negative effect on skill premium is accompanied by an increase in skill intensity. Figure 2d shows that drug violence increases the share of white-collar or non-production employees in total employment. That is, drug violence works as a negative labor-supply shock on blue-collar workers. As blue-collar workers become relatively scarce in the local labor market, blue-collar employment decreases with a significant increase in the relative wages of blue-collar workers. These findings suggest that a violent environment has the ability to influence the technology of firms— the way production is organized. Firms use production technologies that are more intensive in the use of the relatively more abundant labor type, white-collar workers, in response to violence-induced local labor-supply shocks.36 34 In Mexico, firms have the option to hire workers either directly, as payroll employees, or indirectly, through external companies, as contractual employees. Payroll employees are entitled to social security contributions and severance pay- ments upon contract termination, obligations that firms must fulfill. Conversely, for contractual employment, firms do not have such obligations. 35 Both white-collar and blue-collar wages in columns 3–4 in panel C of table S3.6 are average wages across workers on the payroll. 36 Note that all adjustment to a local labor-supply shock could also take place between firms or between industries by inducing a decrease in scale of those production units that are intensive in the use of the now relatively scarce labor input 650 Utar The greater vulnerability of blue-collar workers to the drug war, compared to their white-collar counterparts, raises important questions. As fig. S1.5 suggests, production workers face higher life risks. Ajzenman, Galiani, and Seira (2015) and Jarillo et al. (2016), along with the discussions in supplementary online appendix S1.2, highlight the disproportionate impact of drug-war violence on res- idents of poorer neighborhoods, typically home to lower-paid workers. This disproportionate exposure to violence on lower-paid workers could elevate their reservation wage, below which the risks of working Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 surpass the financial benefits. The decline of blue-collar employment might also stem from the growth of the illegal sector, which com- petes for the manufacturing workforce. The rising demand for brute labor force might tempt particularly male workers away from legal manufacturing jobs. If so, a stronger negative effect on blue-collar employ- ment might be observed among male-workforce-intensive firms. Conversely, women, typically earning less, and less often primary breadwinners, may have a more elastic labor-supply participation compared to male workers. It is also possible that women, already earning lower wages, more easily switch from for- mal to informal economic activities that can be conducted at home or near home. Therefore, if increased risks elevate reservation wages, leading to reduced labor-market participation or reduced participation in the formal manufacturing sector, firms with a predominantly female workforce might be more adversely affected.37 The following section will explore these potential channels by studying heterogenous firm responses to a violent environment. 5. Heterogeneous Responses to Violence This section uses the rich information on establishments provided by the annual survey (EIA) and the technology survey (ENESTyC), to study the potentially heterogeneous impact of violence to pinpoint the channels through which firms are affected and to characterize the most vulnerable firms. 5.1. Local Labor-Market Channel Figure 3 presents the sensitivity of the employment response to drug violence across plants with different susceptibility to violence-induced labor-supply shocks. For this, the estimation sample is partitioned based on the median of plants’ initial (year-2005) characteristics, and equation (1) is estimated separately for the resulting sub-samples. Table S3.8 in the supplementary online appendix presents the full results.38 Figure 3a presents the impact of violent conflict on employment for low- and high-wage plants. Plants with average monthly wages below the sample median of average monthly wages as of 2005 are classified as low wage, and plants with wages above the median value (9,300 2010 Mexican peso) as high-wage plants. The drop in employment is concentrated among low-wage plants. The elasticity estimate is −0.12 and statistically significant at the 5 percent level for low-wage plants, the estimate for high-wage plants is −0.03 and statistically insignificant. Figure 3b shows the elasticity of employment with respect to violence depending on the share of female workers among firm’s payroll workers. Firms with a female-intensive workforce experience a stronger decline in employment. The estimates suggest that doubling the homicide rate causes a 10.4 percent (= −0.15 ln(2)) decline in total employment for firms with female-intensive workforce, as opposed to a 4.2 percent (= −0.06 ln(2)) drop for other firms.39 (Rybczynski theorem). Dustmann and Glitz (2015) emphasize the importance of within-firm adjustment in response to changes in local labor supply. 37 Labor economists widely agree on the higher labor-supply elasticities among married women. Keane (2011) provides a comprehensive review of this literature. 38 See also table S3.9 in the supplementary online appendix for the heterogeneity analysis based on an interaction approach. 39 The median level of female share of workforce in 2005 is 0.20; therefore, female-intensive plants are plants with at least 20 percent female employment. The World Bank Economic Review 651 Figure 3. Heterogeneity in Employment Response to Drug War Violence. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (a) (b) (c) (d) Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Estimation of equation (1) by two-stage least squares. Solid bar frames indicate statistical significance at the 10 percent level or less. For each figure, estimation is conducted separately depending on the median level of the characteristics written on the top of each figure. All characteristics are the values as of year 2005. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. The log homicide rate is instrumented using equation (2). Full results are shown in table S3.8. Next, I use the nationwide representative establishment-level survey, ENESTyC 2005, to calculate the average annual wage for female and male unskilled production workers and then merge the data with plants in the study based on their four-digit industry classification. Figure 3c shows the effect on employment across high- versus low-female-wage industries. The findings show that establishments in industries paying lower wages to female workers see a notable decline in employment, whereas those offering higher wages to female employees do not experience an employment decrease due to violence. Figure 3d plots the employment elasticity estimates across low- and high-wage industries, this time for male workers. The employment effect of violence is precisely estimated for both groups, and the magnitudes are similar whether establishments on average have low or high wages among male blue- collar workers. These findings once again point to the drop in relatively lower-paid (mostly) female 652 Utar workers from the local labor force as a mechanism behind the labor-market effect of violence on firms.40 Unionization could be a crucial factor affecting workers’ bargaining power, and consequently their compensation and benefits, including more secure transportation for workers and a safer, better-protected work environment. These benefits may mitigate the effects of violence on workers. Further findings pre- sented in table S3.8 indicate that plants with unionization levels above the median among their production Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 workers do not face a significant reduction in employment. In contrast, plants with a low degree of union- ization among production staff experience a substantial decrease in employment. These findings reveal that the impact of drug violence on employment is not uniformly experienced across manufacturing plants; it predominantly affects low-wage, female-intensive, and less-unionized es- tablishments. The next section turns the focus to heterogeneity in output elasticity of violence. 5.2. Violence-Induced Local Demand Shock The findings on employment and wage effects at the plant level, along with the heterogeneous patterns found across plants, are consistent with a drug-violence-induced labor-supply shock and corroborate with household-level studies cited above on the drug war’s impact. While a violence-induced labor-supply shock, by reducing production workers, might cause a decline in output, a significant effect documented on product scope in Results in table 3 might suggest violence also affects firms’ output demand. Violence is likely to reduce the size of the market, and this effect is expected to be stronger for firms selling and sourcing locally. Table S3.10 in the supplementary online appendix presents the heterogeneity in output elasticity of violence depending on establishment characteristics as of 2005, and fig. 4 illustrates selected findings from this analysis.41 Figure 4a plots the output elasticity of violence based on plants’ exporting status as of 2005. The reduction in output attributable to the drug war is primarily observed in plants with a heavy focus on the domestic market. An elasticity estimate of −0.17 suggests that a doubling of the homicide rate leads to a 12 percent decrease in output value for non-exporting plants. Conversely, the impact on exporting plants is negligible and lacks statistical significance. Focusing on importing status as of 2005, the estimate of output elasticity is −0.20 versus −0.09 for non-importing and importing plants, respectively. This indicates that plants that source only domestic inputs experience a 14 percent (= −0.20 ln(2)) drop in output due to heightened violence, while the average impact on importing plants is 6 percent and significant only at the 10 percent level (table S3.10, panel B). These results show that domestically selling and sourcing firms’ outputs are disproportionately im- pacted by the escalation of drug violence.42 The results also suggest that the drug violence did not con- stitute a major problem in the transportation of goods, since exporters and importers (who tend to reach more distant markets) rely more heavily on transporting their goods. To confirm, panel C of table S3.10 presents the output elasticity with respect to the local drug violence depending on the share of freight 40 Note that migration out of exposed areas is likely to affect both genders similarly. Table S1.1 in the supplementary online appendix shows a modest migration response. Table S1.1 shows that people living in exposed states are more likely to emigrate to other countries in comparison to people in non-exposed states. However, in general, there is a strong overall declining trend in the number of international emigrants (namely emigrants to the United States) over the sample period, which is likely to be due to stricter policies in the United States regarding illegal immigration. Bazzi et al. (2021) show that increased sanctions of the US Border Patrol on apprehended illegal immigrants from Mexico over 2008–2012 was effective in increasing border security. Consistently, Basu and Pearlman (2017) and Aldeco Leo, Jurado, and Ramírez-Álvarez (2022) find a muted migration response to drug violence over this period. 41 See also table S3.11 for the heterogeneity analysis based on an interaction approach. 42 The analysis using the export and import intensity measures produces similar results and available. The World Bank Economic Review 653 Figure 4. Heterogeneity in Output Response to Drug War Violence. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (a) (b) (c) (d) Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Estimation of equation (1) by two-stage least squares. Solid bar frames indicate statistical significance at the 10 percent level or less. For each figure at the bottom part, a separate estimation is conducted depending on the median level of the characteristics written on the top of each figure. For export, the estimations are conducted among exporters and non-exporters, and similarly for import. All characteristics are the values as of year 2005. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. The log homicide rate is instrumented as described in (2). Full results are shown in table S3.10. expenses in total service expenses of plants.43 The results show a significant sensitivity of output to the drug violence regardless of the importance of the transportation expenses. Magnitude-wise, the effect is even larger for non-transportation-intensive plants (−0.19 versus −0.11), indicating that disruption in transportation is not a major channel through which the drug war affects firms. Next, using the information on plants’ sales and materials purchases across different regions from the ENESTyC survey, I construct entropy measures of firm diversification across four-digit industries. The sales diversification measure, which is used in the IO literature (Palepu 1985), increases with the number of geographic segments a firm operates in and decreases with the relative importance of each segment in total sales, reaching zero for non-diversified firms. Similarly, I define materials diversification measures based 43 This information, just like the plant-level import information, is obtained from the EIA and hence the estimation sample is somewhat smaller. The EIA-EIM matched sample properties are provided in supplementary online appendix S4. 654 Utar on the geographic distribution of firms’ materials purchases. ENESTyC provides information on plants’ sales and procurement of materials across eight mutually exclusive and exhaustive regions worldwide. Mexico as a whole is considered one market, as there are no details regarding sales and purchases within the domestic market. The idea is that the more diversified a firm is worldwide, the more diversified it is likely to be domestically. Plants are classified as “diversified” if their entropy index takes a value that is larger than the sample median. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Figure 4c shows that firms’ output is more resilient to violence the more diversified they are. More precisely, doubling the homicide rate leads to a 10 percent decline in the value of production among plants with a lower rate of sales diversification, while the effect is not statistically significant among diversified establishments. Similar results are obtained when focusing on geographic diversification of inputs (fig. 4d or panel E of table S3.10) with less of a stark difference in this case. Table S3.10 also reveals significant variation in how plants’ output responds to violence, depending on their technological characteristics. Panel F indicates that plants with a lower capital-to-worker ratio bear the brunt of the output reduction. Moreover, plants characterized by a higher reliance on labor, as indicated by an above-median labor cost-share (the ratio of labor expenses to total non-capital expenses), face a pronounced decrease in output (−0.263). Specifically, this −0.263 estimate suggests that doubling the violence in a metropolitan area results in a stark 18 percent decrease in output for labor-intensive plants. These findings imply that violence primarily disrupts the domestic, local market rather than inter- national markets. This observation aligns with the record-high trade via surface transport between the United States and Mexico during the drug-war period. It supports the media narrative that, despite the surge in violence within Mexico, trade activities between the United States and Mexico proceeded with relative ease (The Economist, June 26, 2010). Independent Effects of Labor-Market and Demand Channels The findings identify two primary ways the Mexican Drug War appears to impacts firms: (a) through violence-induced shocks to the local labor supply, mainly affecting low-wage, blue-collar workers, and (b) through a reduction in local demand due to violence. This section delves into the relative significance of these potential channels. The data reveal that firms with a high concentration of female workers and low wages are partic- ularly vulnerable to employment declines, while those focused solely on the domestic market are more likely to see a drop in output. Nonetheless, a decrease in blue-collar employment due to the labor-supply shock should logically result in reduced output, as firms operate with fewer production floor workers. Therefore, firms more affected by the drug war’s impact on the labor market are expected to experience a disproportionate decrease in both employment and output. Conversely, firms more exposed to a violence-induced negative demand shock might not exhibit a sim- ilarly flexible employment response to violence, particularly if the demand shock is viewed as temporary and the costs associated with hiring and firing are significant and non-convex. The question of whether a violence-induced demand shock contributes to employment reductions is addressed by examining the differential impacts on output, employment, and wages based on plants’ exporting status as of 2005, as shown in panel A of table 4. This analysis confirms that non-exporting plants exhibit a disproportionate impact in output due to increased drug violence. However, it also reveals that exporting does not fully insulate plants from the adverse effects of violence through labor-supply changes. The impact of violence on blue-collar employment and wages is comparable between exporters and non-exporters, suggesting that the labor-supply channel predominantly drives the observed changes in employment due to violence. Panel B of table 4 presents the results when violence is interacted with a characteristic that will make the plants more or less susceptible to the violence-induced labor-supply shock: specifically in this case, Table 4. Demand and Labor-Supply Channels Spec: 2SLS (1) (2) (3) (4) (5) (6) Dep. var. Log output Log Log emp Log emp Log avg. monthly Log avg. monthly employment blue-collar white-collar wages wages blue-collar white-collar Panel A. Output demand channel The World Bank Economic Review Violence −0.156∗∗∗ −0.065∗∗ −0.099∗∗∗ 0.004 0.099∗ −0.061 (0.042) (0.024) (0.022) (0.040) (0.055) (0.041) Violence × Export 0.101∗∗∗ −0.009 −0.006 0.055∗∗∗ 0.016 −0.046 (0.017) (0.027) (0.032) (0.017) (0.031) (0.044) Kleibergen–Paap F -excluded instrument 10.54 10.54 10.69 10.54 10.32 10.34 Sanderson–Windmeijer F -test (violence) 78.82 78.82 80.64 74.40 85.35 77.23 Sanderson–Windmeijer F -test (interaction) 92.35 92.35 94.18 94.32 91.24 93.65 Panel B. Labor supply channel Violence −0.139∗∗∗ −0.079∗∗∗ −0.120∗∗∗ 0.048 0.126∗∗ −0.105∗∗ (0.035) (0.021) (0.023) (0.043) (0.059) (0.050) Violence × Unskilled female wage 0.068∗∗ 0.026∗∗∗ 0.047∗∗∗ −0.050∗ −0.054∗∗ 0.065∗∗ (0.031) (0.010) (0.016) (0.027) (0.023) (0.028) Sanderson–Windmeijer F -test (violence) 52.19 52.19 52.80 52.90 53.79 57.27 Sanderson–Windmeijer F -test (interaction) 45.30 45.30 47.48 44.40 47.97 64.95 For both panels: Plant FEs       Pre-trends in homicide rate       5-dig. industry × Year FEs       No. of observations 30,605 30,605 29,480 30,118 24,745 24,761 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Estimation using two-stage least squares (2SLS). Violence is the log number of homicides per thousand inhabitants of a metropolitan area. It is instrumented with equation (2). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. 655 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 656 Utar the average wage of blue-collar female workers. Lower-wage plants should experience a disproportionate decline in output too, as they are more vulnerable to the labor-supply channel. The results confirm this and show that the labor-supply channel is not only the major driver of the employment effect but also plays a role in the output reductions. Additional results presented in the supplementary online appendix (tables S3.12–S3.13) support the conclusion that the two channels co-exist. Taken together, these findings suggest that the Mexican Drug War triggers significant reallocation Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 both within firms and among continuing firms, identifying a violence-induced supply shock as a major, though not the sole, mechanism affecting manufacturing firms. This raises the following question: Are the drug war’s effects potent enough to influence the extensive margin, leading to plant closures? The next section explores the relationship between plant exits and the homicide rate. 5.3. Drug Violence and Plant Closings Table 5 shows results from estimating a probit model on plant exits.44 Column (1) of table 5 shows a significant positive impact of the homicide rate on the probability of exit. In column (2) the pre-trends in the homicide rate are included, and the impact is lower in magnitude, though still positive and significant. In columns (3) and (4), initial characteristics of plants (the logarithm of capital per worker, the ratio of IT expenditure over total expenses, the logarithm of labor productivity, export indicator, and import indicator), and metro-level controls (employment shares of crop production, precious metal mining, oil, and natural gas extraction) are included. The coefficient in column (4) implies that a marginal change in the homicide rate from the average of 0.085 increases the likelihood of plant exit by 2 percentage points. In column (5), the homicide rate is instrumented and the Wald test confirms the endogeneity of the homicide rate (at the 10 percent level). The positive and significant coefficient confirms that the drug war’s violence significantly contributes to plant closures. Do all plants face the same risk of closure due to drug violence? The instrumental variable analysis in table S3.14 in the supplementary online appendix indicates heterogenous effects.45 According to table S3.14, smaller plants (those with up to 40 employees) are particularly susceptible to the drug war’s im- pacts. Additionally, plants with a higher proportion of female employees and those paying lower wages are more likely to close, suggesting that the labor-supply shock induced by drug violence affects plant exit probabilities as well. Furthermore, the results in table S3.14 reveal that being an exporter or importer, along with diversi- fication in sales and purchases, significantly mitigates the risk of closure due to the drug war, highlight- ing the protective effect of international engagement and diversification against the adverse impacts of violence. In sum, the Mexican Drug War leads to reallocation of resources across heterogeneous plants, both at the intensive and at the extensive margin. Particularly hard hit are smaller, low-wage, labor-intensive with a relatively higher share of female workers and locally selling and sourcing plants. While the most immediate impacts of the drug war are felt by plants that tend to be small and less productive—limiting the broader implications for aggregate output—it is crucial to recognize that firms often begin on a small, local scale. The most productive among them eventually expand and engage in international markets. The violence stemming from organized crime, by targeting these nascent plants with the potential for growth, emerges as a significant barrier to the enhancement of domestic industrial capabilities. 44 As exit is a relatively rare event, including five-digit industry-by-year fixed effects reduces much of the identifying vari- ation. As a result, in the plant exit analysis, three-digit industry-by-year fixed effects are used. 45 Since probit models cannot incorporate plant fixed effects, these findings are derived using a linear probability model. The World Bank Economic Review 657 Table 5. Drug War Leads to Plant Closings (1) (2) (3) (4) (5) Specification: Probit Probit Probit Probit IVProbit Violence (homicide rate) 0.447∗∗∗ 0.197∗ 0.253∗∗ 0.275∗∗∗ 1.157∗∗ (0.152) (0.116) (0.100) (0.102) (0.573) Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Avg. marg. eff. 0.036 0.016 0.019 0.020 0.086 Prob of exit 0.033 0.033 0.028 0.028 0.028 Plant characteristics No No    Time-varying local market No No No   characs Pre-trends in violence No     3-dig. industry × Year FE      Pseudo R2 0.067 0.068 0.065 0.065 Wald test of exogeneity 2.840 p > χ2 0.092 No. of observations 25,979 25,979 22,528 22,528 22,528 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Violence is measured as the number of homicides per thousand inhabitants in a metropolitan area. The dependent variable in all regressions is plant exit, which is an indicator variable that takes value 1 if a plant exits the next period; as a result it is not defined in year 2010. Plant characteristics include year-2005 values of log capital per worker, IT-intensity, labor productivity, exporter dummy, and importer dummy. Time-varying local market characteristics include metropolitan area-level employment shares of crop production; metal mining including gold, silver, copper, and uranium; and the metropolitan area-level employment share of oil and natural gas extraction. Pre-trends in violence are the 2002 homicide rate interacted with year dummies. Robust standard errors, reported in parentheses, are clustered by metropolitan area (57). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. 6. Additional and Robustness Analysis In this section, I go over possible confounding effects and alternative explanations to ascertain that the results stand. 6.1. Firm Selection Table 5 and supplementary online appendix table S3.14 show that plants subjected to violence are more prone to closure, especially those with lower wages, a female-intensive workforce, and a domestic market focus. Section 5 further demonstrates that, conditional on remaining in the market, these characteristics also render plants more vulnerable to violence. This suggests that the actual impact of violence at the intensive margin might be underappreciated due to selection effects. To address this, I employ the “identification at infinity” concept (Chamberlain 1986; Mulligan and Rubinstein 2008), positing that selection bias diminishes for plants with a higher likelihood of survival. By narrowing the analysis to plants more likely to endure and progressively excluding those most susceptible to exit, I examine how estimates adjust. The findings, presented in table S3.16 in the supplementary online appendix, affirm that the adverse impact of violence on output at the intensive margin is partially obscured by plant closures. Nonetheless, the table’s results suggest that the influence of selection from plant exits on the observed compositional changes within firms is limited. 6.2. Border-Specific Shocks Metropolitan areas on the US border may be exposed to differential demand or supply shocks that are confounding the analysis since these locations are also among those experiencing heightened violence. Table S3.17 presents 2SLS results when additionally border-specific aggregate shocks are allowed for. 658 Utar The results show qualitatively similar results, providing evidence that the results are not confounded by border-specific time-varying shocks. 6.3. Informal Sector To explore the possibility that the informal sector confounds the results, by affecting margins of em- ployment adjustment, the estimation equation is adjusted to allow for time-varying local shocks that are Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 correlated with the share of the informal sector in each local economy. The share of informal economy for each metropolitan area is interacted with year fixed effects and included in equation (1). The results presented in table S3.18 with controls for the size of the informal economy are qualitatively and quan- titatively similar to the main findings, highlighting the robustness of the findings against the potential confounding role of informal economic activities. 6.4. Industry- or Export-Sector-Specific Shocks The empirical strategy in this paper allows for differential time trends across five-digit industries, and the results are also robust to including product-specific business cycles (table 2, column (5)). However, even within a detailed manufacturing activity, not all plants export or sell domestically. If exporters are more likely to be affected by the Great Recession, this could lead to a heterogeneous impact of the Great Recession within industries. To investigate whether such a channel plays a role in the results, I addi- tionally include differential time trends for exporters, namely the interaction of the exporter dummy with year fixed effects, and estimate equation (1) by two-stage least squares. The results are presented in table S3.19. They show that differential time trends for exporters do not affect the analysis and in- dicate that the Great Recession does not confound the estimated effect of violence. The analysis is also conducted using the data from only two years, 2005 and 2010, removing the recession period (table S3.20), as well as constructing trade exposure of local labor markets and explicitly controlling for lo- cal labor-market trade exposure (see supplementary online appendix S3.16 and table S3.21) with robust results. In sum, I find no evidence that the results are influenced by either trade- or recession-induced employment loss. 6.5. Municipality-Level Results The findings suggest that violence primarily affects firms via local labor markets. As a result, the right geographical unit of analysis is local labor markets. The fact that the majority of the metropolitan areas are non-adjacent labor markets also mitigates spillover issues.46 However, the main results should still not differ radically when one considers municipality-level homicide rates. Table S3.22 in the supplementary online appendix presents the results when the spatial unit is taken as municipality. Due to the absence of homicides in some small municipalities in certain years, the number of observations drops a little, but the results are similar. 6.6. Alternative Instruments to Capture Drug-Related Escalation of Violence due to the Drug War The primary spatiotemporal variation in the instrument stems from the radical shift in the Mexican gov- ernment’s drug-enforcement strategy which is implemented at the state level via the military operations. While this instrument is shown to be robust to considering a wide range of time-varying state-level eco- nomic and financial factors, to ensure the findings are not in any way dependent on a particular instru- mentation strategy, an alternative instrumentation is considered. The competitive advantage of Mexican DTOs lies in trafficking routes to the United States, and the contention for these routes hence the drug 46 For example, a plant located in the center of Jesús María municipality will likely have most of its workers residing in Aguascalientes municipality which is only 4 km away. Both Jesús María and Aguascalientes municipalities are part of the same commuting zone, hence the metropolitan area. The World Bank Economic Review 659 violence is likely to escalate the closer the routes are to the US border. By interacting the distance of metropolitan areas to the US border with the post-Calderón period that takes the value of 1 on and after 2007, an alternative instrument is constructed to capture the unexpected increase in homicide rate at- tributed to the drug war. The results obtained using this alternative instrumentation strategy are shown in table S3.23 and discussed in supplementary online appendix S3.18. This instrumentation is similarly strong, as indicated by the first-stage F -statistics and generate similar findings. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 6.7. Results Using Higher-Frequency Data By employing monthly plant-level data, the analysis is also conducted at the monthly frequency (table S3.25). In general, the findings in the paper are robust to these alternative approaches in the empirical strategy. 7. Concluding Remarks To shed light on how violence and organized crime affect industrial development, I study firm- level consequences of drug-trade-related violence. Based on an instrumental variable strategy that exploits the sudden, unanticipated, and geographically heterogeneous surge in organized crime and violence in Mexico during the late 2000s, and longitudinal plant-level data from all metropoli- tan areas of Mexico, I show that violence has a significant negative impact on plant-level out- put, employment, product scope, productivity, and the capacity utilization of Mexican manufacturing establishments. A violent environment affects labor more than capital and affects blue-collar workers more than white- collar employees. The negative impact on firm-level employment is entirely driven by blue-collar workers and at the same time, firm-level wages move in opposing directions with a significant increase in the relative wages of blue-collar workers. These findings indicate that violence manifests at the firm level as a negative unskilled-labor-supply shock, and firms adjust by using the relatively abundant, skilled labor, more intensively. These results show that firms are affected through the labor market in an interesting mechanism opposite to most other economic shocks (which hit the firms first and, by consequence, the labor market). In the case of the Mexican Drug War, results suggest that it is the other way around: the violence deters workers from working and increases the reservation wage, below which the risk of working outweighs the benefit. The results also show resilience to drug violence varies substantially across firms. The employment elasticity of violence is higher among firms that are labor intensive, low wage, less unionized, and with a relatively high rate of female workers. At the same time, well-diversified, bigger, capital-intensive, ex- porting, and importing firms have better resilience to violence-induced negative demand shocks and the resulting declines in output and product scope. I find that both local labor-supply and local de- mand channels co-exist and operate both at the intensive and the extensive margins. At the extensive margin, the Mexican Drug War causes plant closings and explains one-quarter of the plant exits over 2005–2010. These results show that there are important distributional and inequality consequences of the recent rise of drug violence in Mexico. The Mexican Drug War significantly hinders development of domestic industrial capability by taking away resources from plants that rely more heavily on low-wage or female workers, and local output and input markets. While the short-run aggregate output effects of the violence may be mitigated by this (static) reallocation, the results suggest potentially important long-run (dynamic) effects on the development of domestic industrial capability. As drug violence and organized crime are important factors affecting especially developing and emerging economies, these findings shed light on the characteristics of the most affected firms and the channels through which they are affected to guide targeted industrial policies. 660 Utar Data Availability Statement The analyses in this paper is based on confidential establishment-level data from the National Institute of Statistics and Geography of Mexico, INEGI. The data can be accessed by application to INEGI’s data lab (Laboratorio de microdatos). Please direct your queries to MICRODATOS@inegi.org.mx. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 References Acemoglu, A., J. Robinson, and R. Santos. 2013 Journal of the European Economic Association, 11(S1): 5–44. Abadie, A., and J. Gardeazabal. 2003. “The Economic Costs of Conflict: A Case Study of the Basque Country.” American Economic Review 93(1): 113–32. Acemoglu˘ , D., G. De Feo, and G. De Luca. 2020. “Weak States: Causes and Consequences of the Sicilian Mafia.” Review of Economic Studies 87(2): 537–81. Ajzenman, N., S. Galiani, and E. Seira. 2015. “On the Distributive Costs of Drug-Related Homicides.” Journal of Law and Economics 58(4): 779–803. Aldeco Leo, L. R., J. Jurado, and A. A. Ramírez-Álvarez. 2022. “Internal Migration and Drug Violence in Mexico.” Working Paper, 2022-11. Banco de Mexico, Mexico City. Alesina, A., S. Piccolo, and P. Pinotti. 2018. “Organized Crime, Violence, and Politics.” Review of Economic Studies 86(2): 457–99. Amodio, F., and M. Di Maio. 2018. “Making Do with What You Have: Conflict, Input Misallocation and Firm Performance.” Economic Journal 128(615): 2559–612. Angrist, J. D., and A. D. Kugler. 2008. “Rural Windfall or a New Resource Curse? Coca, Income, and Civil Conflict in Colombia,” Review of Economics and Statistics 90(2): 191–215. Ashby, N., and M. Ramos. 2013. “Foreign Direct Investment and Industry Response to Organized Crime: The Mexican Case.” European Journal of Political Economy 30(C): 80–91. Bagley, B. M., and J. D. Rosen. 2015. Drug Trafficking, Organized Crime, and Violence in the Americas Today. Gainesville: University Press of Florida. Basu, S., and S. Pearlman. 2017. “Violence and Migration: Evidence from Mexico’s Drug War.” IZA Journal of De- velopment and Migration 7(18): 1–29. Bazzi, S., S. Burns, G. Hanson, B. Roberts, and J. Whitley. 2021. “Deterring Illegal Entry: Migrant Sanctions and Recidivism in Border Apprehensions.” American Economic Journal: Economic Policy 13(3): 1–27. Becker, G., and Y. Rubinstein. 2011. “Fear and the Response to Terrorism: An Economic Analysis.” CEP Discussion Paper No. 1079. London: Centre for Economic Performance. Besley, T., T. Fetzer, and H. Mueller. 2015. “The Welfare Cost of Lawlessness: Evidence from Somali Piracy.” Journal of the European Economic Association 13(2): 203–39. Brueckner, J. K., J.-F. Thisse, and Y. Zenou. 2002. “Local Labor Markets, Job Matching, and Urban Location.” Inter- national Economic Review 43(1): 155–71. Cardona, J. 2010. “Poor Mexicans Easy Scapegoats in Vicious Drug War.” World News October 7, 2010. Reuters. https://www.reuters.com/article/world/poor- mexicans- easy- scapegoats- in- vicious- drug- war- idUSTRE6963MF/ Castillo, J. C., D. Mejia, and P. Restrepo. 2020. “Scarcity without Leviathan: The Violent Effects of Cocaine Supply Shortages in the Mexican Drug War.” Review of Economics and Statistics 102(2): 269–86. Chamberlain, G. 1986. “Asymptotic Efficiency in Semi-parametric Models with Censoring.” Journal of Econometrics 32(2): 189–218. Del Prete, D., M. Di Maio, and A. Rahman. 2023. “Firms amid Conflict: Performance, Production Inputs, and Market Competition.” Journal of Development Economics 164: 103143. Dell, M. 2015. “Trafficking Networks and the Mexican Drug War.” American Economic Review 105(6): 1738–79. Dell, M., B. Feigenberg, and K. Teshima. 2019. “The Violent Consequences of Trade-Induced Worker Displacement in Mexico.” American Economic Review: Insight 1(1): 43–58. Dix-Carneiro, R., R. Soares, and G. Ulyssea. 2018. “Economic Shocks and Crime: Evidence from the Brazilian Trade Liberalization.” American Economic Journal: Applied Economics 10(4): 158–95. Draca, M., and S. Machin. 2015. “Crime and Economic Incentives.” Annual Review of Economics 7(1): 389–408. Dube, O., and J. F. Vargas. 2013. “Commodity Price Shocks and Civil Conflict: Evidence from Colombia.” Review of Economic Studies 80(4): 1384–421. The World Bank Economic Review 661 Dustmann, C., and F. Fasani. 2016. “The Effect of Local Area Crime on Mental Health.” Economic Journal 126(593): 978–1017. Dustmann, C., and A. Glitz. 2015. “How Do Industries and Firms Respond to Changes in Local Labor Supply?” Journal of Labor Economics 33(3): 711–50. Espinoza, V., and D. Rubin. 2015. “Did the Military Interventions in the Mexican Drug War Increase Violence?” American Statistician 69(1): 17–27. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Finckenauer, J. O., J. R. Fuentes, and G. L. Ward. 2001. “Mexico and the United States: Neighbors Confront Drug Trafficking.” National Institute of Justice Publications. https://www.ojp.gov/ncjrs/virtual-library/abstracts/mexico - and- united- states- neighbors- confront- drug- trafficking Gonzalez, E. R., and I. Llamosas-Rosas. 2021. “Observando la Evolucion del Sector Informal desde el Espa- cio: Un Enfoque Municipal 2013-2020.” Working Papers No. 2021-18, Banco de México, Ciudad de Méx- ico. https://www.banxico.org.mx/publicaciones- y- prensa/documentos- de- investigacion- del- banco- de- mexico/%7 BB8491B85- 6FB0- 3953- 5EDB- B5397BBCDDD1%7D.pdf Gorrín, J., J. Morales-Arilla, and B. Ricca. 2023. “Export Side Effects of Wars on Organized Crime: The Case of Mexico.” Journal of International Economics 144. https://doi.org/10.1016/j.jinteco.2023.103775 Grayson, G. W. 2013. “The Impact Of President Felipe Calderón’s War on Drugs on the Armed Forces: The Prospects for Mexico’s “Militarization” and Bilateral Relations.” The Strategic Studies Institute, U.S. Army War College. https://ssi.armywarcollege.edu/SSI- Media/Recent- Publications/Display/Article/3618241/the- impact- of- president- felipe- calderns- war- on- drugs- on- the- armed- forces- the- pr/ Guidolin, M., and E. La Ferrara. 2007. “Diamonds Are Forever, Wars Are Not: Is Conflict Bad for Private Firms?” American Economic Review 97(5): 1978–93. Jarillo, B., B. Magaloni, E. Franco, and G. Robles. 2016. “How the Mexican Drug War Affects Kids and Schools? Evidence on Effects and Mechanisms.” International Journal of Educational Development 51: 135–46. Kilmer, B., S. Everingham, J. P. Caulkins, G. Midgette, R. L. Pacula, P. Reuter, R. M. Burns, and B. Han. 2014. “What America’s Users Spend on Illegal Drugs: 2000-2010.” Santa Monica, CA: RAND Corporation, RR-534-ONDCP. Klapper, L., C. Richmond, and T. Tran. 2013. “Civil Conflict and Firm Performance: Evidence from Côte d’Ivoire.” Policy Research Working Paper WPS 6640. World Bank Group: Washington, D.C.https: //documents.worldbank.org/en/publication/documents-reports/documentdetail/176061468020680001/civil-confli ct- and- firm- performance- evidence- from- cote- divoire Korovkin, V., and A. Makarin. 2021. “Production Networks and War.” CEPR Discussion Paper No. 16759. CEPR Press, Paris & London. https://cepr.org/publications/dp16759 Ksoll, C., R. Macchiavello, and A. Morjaria. 2023. “Electoral Violence and Supply Chain Disruptions in Kenya’s Floriculture Industry” Review of Economics and Statistics 105(6): 1335–1351. La Botz, D. 2011. “Mexico’s Drug War Wreaks Havoc on Workers.” Labor Notes, January 03, 2011. Available at https://labornotes.org/blogs/2011/01/mexico%E2%80%99s- drug- war- wreaks- havoc- workers Lindo, J. M., and M. Padilla-Romo. 2018. “Kingpin Approaches to Fighting Crime and Community Violence: Evidence from Mexico’s Drug War.” Journal of Health Economics 58(C): 253–68. Martin, P., M. Thoenig, and T. Mayer. 2008. “Civil Wars and International Trade.” Journal of the European Economic Association 6(2/3): 541–50. McLaren, J. 2004. “Globalization and Civil War.” Mimeo, University of Virginia. Charlottesville. Mejia, D., and P. Restrepo. 2016. “The Economics of War on Illegal Drug Production and Trafficking.” Journal of Economic Behavior and Organizations 126(PA): 255–75. Melnikov, N., C. Schmidt-Padilla, and M. M. Sviatschi. 2020. “Gangs, Labor Mobility, and Development.” NBER Working Paper 27832. National Bureau of Economic Research. https://www.nber.org/papers/w27832 Merino, J. 2011. “Nexos: Army’s Presence Equals More Homicides in Mexico? (Revisited)” Revista Nexos. Last accessed March 1 2024. Available athttps://insightcrime.org/news/analysis/nexos- armys- presence- equals- more- ho micides- in- mexico- revisited/. Mora, E. 2009. “Civilians Ran a Greater Risk of Being Killed in Juárez Last Year Than in Baghdad.” CNS News. March 5. Last accessed on 11/18/2020 athttps://www.cnsnews.com/news/article/civilians- ran- greater- risk- being- killed- juarez- last- year- baghdad Mulligan, C. B., and Y. Rubinstein. 2008. “Selection, Investment, and Women’s Relative Wages over Time.” Quarterly Journal of Economics 123(3): 1061–110. 662 Utar National Drug Control Agency. 2015. National Drug Control Strategy Data Supplement 2015, The White House, Washington D.C.https://obamawhitehouse.archives.gov/sites/default/files/ondcp/policy- and- research/2015_data_s upplement_final.pdf National Drug Intelligence Center. 2011. National Drug Threat Assessment Report, U.S. Department of Justice. Na- tional Drug Intelligence Center, Johnstown, PA. https://www.justice.gov/archive/ndic/pubs44/44849/44849p.pdf P Keane, M.. 2011. “Labor Supply and Taxes: A Survey.” Journal of Economic Literature 49(4): 961–1075. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Palepu, K. 1985. “Diversification Strategy, Profit Performance and the Entropy Measure.” Strategic Management Jour- nal 6(3): 239–55. Pinotti, P. 2015. “The Economic Costs of Organised Crime: Evidence from Southern Italy.” Economic Journal 125(586): F203–32. Raphael, S., and R. Winter-Ebmer. 2001. “Identifying the Effect of Unemployment on Crime.” Journal of Law & Economics 44(1): 259–83. Robles, G., G. Calderón, and B. Magaloni. 2013. “The Economic Consequences of Drug Trafficking Violence in Mexico.” Mimeo, Stanford University. https://cddrl.fsi.stanford.edu/publication/economic- consequences- drug- traf ficking- violence- mexico Rumelt, R. P. 1982. “Diversification Strategy and Profitability.” Strategic Management Journal 3(4): 359–69. Staiger, D., and J. H. Stock. 1997. “Instrumental Variable Regressions with Weak Instruments.” Econometrica 65(3): 557–86. UNODC (United Nations Office on Drugs and Crime). 2014. “Cocaine Retail Prices (Street Prices), USD per Gram.” Technical Report. New York. Accessed June 13, 2017. ———. 2019. “Global Study on Homicide” United Nations Office on Drugs and Crime, Vienna. Utar, H., and L. B. Torres Ruiz. 2013. “International Competition and Industrial Evolution: Evidence from the Impact of Chinese Competition on Mexican Maquiladoras,” Journal of Development Economics 105: 267–87 . https: //doi.org/10.1016/j.jdeveco.2013.08.004 V Rozo, S.. 2018. “Is Murder Bad for Business? Evidence from Colombia.” Review of Economics and Statistics 100(5): 769–82. Velásquez, A. 2020. “The Economic Burden of Crime: Evidence from Mexico.” Journal of Human Resources 55(4): 1287–318. Williams, P. 2012. “The Terrorism Debate over Mexican Drug Trafficking Violence.” Terrorism and Political Violence 24(2): 259–78. World Drug Report. 2010. United Nations Office on Drugs and Crime (UNODC), United Nations Publication, Sales No. E.10.XI.13. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Firms and Labor in Times of Violence: Evidence from the Mexican Supplementary Online Appendix Drug War Hale Utar S1. The Drug War in Mexico S1.1. Military Interventions At the end of December 2006, the federal government initiated joint military operations (operativos con- juntos militares) in agreement with some states. Figure S1.1 shows the states that joined the federal gov- ernment’s policy intervention. The location of states that joined the military interventions mostly coincide Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 with the major drug trafficking routes. Merino (2011) shows a causal link between the military interven- tions and the surge in violence and, using a regression discontinuity design, Dell (2015) establishes a causal link between the change in the government’s policy and the increased violence in Mexico. Figure S1.2 shows the extent of the increase in violence after the start of the kingpin policy. Figure S1.3 shows the evolution of the homicide rate in metropolitan areas in Mexico since 2007. Figure S1.4 shows that violence increased dramatically after 2007 in states with military operations as opposed to other states, and the increase in violence was driven by drug-related homicides. S1.2. Fragmentation of Drug Cartels and Expansion of Violence into the Public Sphere The number of major cartels in Mexico increased more than 70 percent (from 7 to 12) over the period of 2006–2010 (Bagley and Rosen, 2015). Much of the violence in Mexico has been due to fights between and within drug cartels, and many of the victims were drug cartel associates. However, violence also led to widespread, random violence, especially in poorer neighborhoods of affected metropolitan areas. As Figure S1.1. States That Join the Federal Army in Military Operations Against Drug Cartels. Source: Author’s own illustration. Note: The figure shows the states that joined the federal government’s policy intervention.. Figure S1.2. Surge in Violence in Mexico. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: National Institute of Statistics and Geography of Mexico, INEGI. Note: This figure shows the monthly number of homicides. Table S1.1. Migration Pattern and Drug War Exposed states Not exposed states Post-drug-war Post-drug-war org. crime≥p75 org. crime≥p25 2005–2010 growth Mean Mean Difference t -stat Inter-state emigrants 0.6% −1 . 5 % 2.1% −0.37 International emigrants −42.1% −45.5% 3.4% −4.70 Inter-state immigrants −6.5% 7.4% −13.9% 1.97 International immigrants 13.6% 27.2% −13.6% 2.34 Source: For the migration data, Consejo Nacional de Población (CONAPO). Note: The table shows the 2005–2010 change in the state-level migration patterns across exposed versus non-exposed states. States with average organized crime rates during 2008–2010 above the 75th percentile are defined as exposed states. These are Baja California, Chihuahua, Durango, Guerrero, Michoacán, Nayarit, Sinaloa, and Sonara. States with average organized crime rate during 2008–2010 below the 25th percentile are defined as non-exposed states. These are Baja California Sur, Campeche, Chiapas, Puebla, Querétaro, Tlaxcala, Veracruz, and Yucatán. an example of DTOs’ use of violence, in October 2010 in Juárez, a group of gunmen stormed into a party in search of a specific person. The person they were looking for was not among the party, but that did not prevent them from killing 13 people aged 13 to 32, including 6 women and girls, and wounding others, which included a 9-year-old boy (Williams 2012). The following month, in the same city, another group of armed men attacked three buses operated by an automotive parts manufacturer, transporting third-shift workers during the early morning hours. This attack resulted in many fatalities and injuries among the workers, while the assailants were apparently in pursuit of a specific worker, whom they abducted from the scene (La Botz 2011).47 47 Another distressing episode unfolded in San Fernando in August 2010, where the Mexican army discovered the remains of 72 South American migrants, including both men and women, interred in a mass grave. Subsequent investigations suggested that they were killed while resisting recruitment attempts by the Zeta cartel. Figure S1.3. Expansion of Urban Violence in Mexico. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: National Institute of Statistics and Geography of Mexico, INEGI. Note: The number of homicides per 100,000 inhabitants across municipalities with at least 100,000 inhabitants or otherwise belonging to a metropolitan area. One plausible factor contributing to this violence is the drug cartels’ utilization of such acts to terrorize the public in an attempt to force the government to back down. Additionally, drug cartels may have increasingly turned to criminal activities such as kidnappings, extortions, and thefts, directly impacting the civilian population, as a means to fund their ongoing conflicts with rival cartels and the military. Figure S1.5 shows the evolution of intentional homicide victims and the probability of being killed across a selected set of occupations. Production workers are especially susceptible to violence; the number of homicide victims who are production workers increased 160 percent between 2007 and 2010. Since there will be more unskilled production workers than, say, professionals and technicians or machine operators, a difference in the level of homicide between these groups is expected. But the rate of increase in the killings of production workers is striking. The bottom graph of the figure shows the likelihood of being a homicide victim, taking the total number of workers in these occupations into account. The death risk of production workers increases substantially to almost the level of drivers.48 The figure makes it clear that unskilled production workers are far more likely to be victimized during the drug war compared to other typical occupations within manufacturing. 48 The occupation classification is economy-wide, so while unskilled production workers or machine operators are largely manufacturing occupations, professionals and technicians, for example, include professionals such as journalists, lawyers, or bankers who are likely to be employed in non-manufacturing sectors and can also be direct targets of DTO violence. Figure S1.4. Violence Across States and the Kingpin Strategy. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: Author’s own elaboration based on data from the National Institute of Statistics and Geography (INEGI), Milenio, and the National Population Council (CONAPO). Note: Information on drug-related homicides from Milenio is available starting from 2007. . S1.3. Manufacturing Employment Across Selected Metropolitan Areas The manufacturing sector employed about 6.3 million workers in 2005, which dropped to 5.7 million in 2010 (INEGI). Figure S1.6 shows how the manufacturing employment changes during the sample period in the metropolitan areas shown in fig. 1 in the text. S1.4. Migration Patterns Using the estimated state-level migration flows provided by Consejo Nacional de Población (CONAPO), table S1.1 presents the change in the pattern of migration in exposed versus non-exposed states. For the purpose of this descriptive analysis, the state-level organized crime rate is used to describe exposed versus non-exposed states. Exposed states are states with an average organized crime rate during 2008–2010 above the 75th percentile. These are Baja California, Chihuahua, Durango, Guerrero, Michoacán, Na- yarit, Sinaloa, and Sonara. Non-exposed states are states with an average organized crime rate during 2008–2010 below the 25th percentile. These are Baja California Sur, Campeche, Chiapas, Puebla, Queré- taro, Tlaxcala, Veracruz, and Yucatán. Table S1.1 shows a significant drop in the inflow of domestic immigrants into the exposed states between 2005 and 2010. Exposed states also have significantly less inflow of international immigrants in comparison to non-exposed states. Although there is an overall strong declining trend in international emigrants during the sample period, exposed states have a signifi- cantly smaller decrease in the number of people moving out of the country in comparison to non-exposed states. Figure S1.5. Occupations and Risk to Life. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: National Institute of Statistic and Geography (INEGI), Estadísticas de mortalidad, and Encuesta Nacional de Ocupación y Empleo. Note: This figure shows the annual number of nationwide homicides depending on victims’ occupations (top) and the number of homicides over the total number of people employed in that occupation (bottom). Occupations follow the Mexican Occupation Classification (CMO). A selected set of occupations is shown here. Figure S1.6. Manufacturing Employment Across Selected Metropolitan Areas. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (a) (b) (c) (d) Source: Data from the Mexican Institute of Social Security (IMSS). Note: Manufacturing employment in 2005 at each metropolitan area is normalized to 100. S2. Summary Statistics and Descriptive Analysis Figure S2.1. Distribution of Number Plants Across Three-Digit Industries. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: Author’s own illustration based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: The figure shows the year-2005 distribution of plants in the estimation sample across the three-digit NAICS industries. Table S2.1. Distribution of Plants and Industries Metropolitan areas Number of Number of 3-digit Number of 4-digit plants industries industries Zona metropolitana de Aguascalientes 95 19 41 Zona metropolitana de Tijuana 49 16 28 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Zona metropolitana de Mexicali 41 13 25 Zona metropolitana de La Laguna 134 17 45 Zona metropolitana de Saltillo 94 15 37 Zona metropolitana de Monclova-Frontera 26 11 20 Zona metropolitana de Piedras Negras 8 4 5 Zona metropolitana de Colima-Villa de Álvarez 4 3 4 Zona metropolitana de Tecomán 6 4 4 Zona metropolitana de Tuxtla Gutiérrez 14 7 12 Zona metropolitana de Juárez 39 12 20 Zona metropolitana de Chihuahua 66 14 30 Zona metropolitana del Valle de México 2,099 21 83 Zona metropolitana de León 262 17 29 Zona metropolitana de San Francisco del Rincón 48 5 7 Zona metropolitana de Moroleón-Uriangato 15 2 2 Zona metropolitana de Acapulco 9 5 5 Zona metropolitana de Pachuca 23 11 12 Zona metropolitana de Tulancingo 13 7 8 Zona metropolitana de Tula 16 8 10 Zona metropolitana de Guadalajara 496 21 70 Zona metropolitana de Ocotlán 18 7 10 Zona metropolitana de Toluca 160 20 54 Zona metropolitana de Morelia 32 13 20 Zona metropolitana de Zamora-Jacona 7 1 1 Zona metropolitana de La Piedad-Pénjamo 17 7 8 Zona metropolitana de Cuernavaca 63 14 31 Zona metropolitana de Cuautla 14 7 12 Zona metropolitana de Tepic 11 5 8 Zona metropolitana de Monterrey 611 21 72 Zona metropolitana de Oaxaca 20 7 10 Zona metropolitana de Tehuantepec 2 1 1 Zona metropolitana de Puebla-Tlaxcala 241 20 53 Zona metropolitana de Tehuacán 15 6 8 Zona metropolitana de Querétaro 140 19 48 Zona metropolitana de Cancún 6 3 3 Zona metropolitana de San Luis Potosí-Soledad de 146 19 56 Graciano Sánchez Zona metropolitana de Ríoverde-Ciudad Fernández 1 1 1 Zona metropolitana de Guaymas 7 3 3 Zona metropolitana de Tampico 54 15 26 Zona metropolitana de Reynosa-Río Bravo 13 7 9 Zona metropolitana de Matamoros 16 10 13 Zona metropolitana de Nuevo Laredo 10 6 7 Zona metropolitana de Tlaxcala-Apizaco 40 15 24 Zona metropolitana de Veracruz 24 7 16 Zona metropolitana de Xalapa 12 5 7 Zona metropolitana de Poza Rica 4 3 3 Zona metropolitana de Orizaba 27 9 19 Table S2.1. Continued Metropolitan areas Number of Number of 3-digit Number of 4-digit plants industries industries Zona metropolitana de Minatitlán 10 5 5 Zona metropolitana de Coatzacoalcos 21 4 9 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Zona metropolitana de Córdoba 26 9 16 Zona metropolitana de Acayucan 2 1 2 Zona metropolitana de Mérida 88 16 30 Zona metropolitana de Zacatecas-Guadalupe 3 2 3 Zona metropolitana de Celaya 44 14 28 Zona metropolitana de Tianguistenco 16 7 10 Zona metropolitana de Teziutlán 2 2 2 Source: The author’s illustration based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: The table shows the distribution of plants and industries in the estimation sample across the metropolitan areas. The observations from Zona metropolitana de Villahermosa are dropped due to the 2007 Tabasco flood. Table S2.2. Summary Statistics Mean Median St dev N Number of employees 238.364 99.833 491.393 30,605 Number of blue-collar employees 159.559 64.667 322.900 30,605 Number of white-collar employees 71.924 22.917 229.183 30,605 Number of days worked 280.482 295 55.582 30,605 Capacity utilization rate 70.230 75 21.110 29,926 Number of varieties 3.126 2 3.023 30,605 Log value of output 11.254 11.272 2.048 30,605 Log value of domestic sales 11.035 11.060 2.022 30,293 Log value of foreign sales 10.236 10.405 2.570 10,812 Share of foreign sales 0.111 0 0.237 30,605 Source: Data from the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 . Note: All values are expressed in 2010 thousand Mexican peso. The table shows the summary statistics of the main variables in the estimation sample (metropolitan areas). Figure S2.2. Distribution of Number Plants Across Three-Digit Industries by Exposure. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA). Note: The figure shows the year-2005 distribution of plants in the estimation sample across the three-digit industries (NAICS) across “High-intensity drug-war zones” and “Other metropolitan areas.” The high-intensity drug-war zones are the following metropolitan areas: Acapulco, Chihuahua, Juárez, La Laguna, Monterrey, and Tijuana. S3. Supplementary Analysis and Robustness S3.1. Metropolitan-Area-Level Security Expenses Table S3.1 presents results when the metropolitan area-level 2005–2010 growths in security expenses are controlled for. To do that, the growth rate in security expenses for each metropolitan area is interacted with year dummies. The results show that including the security expenses does not change the results, Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 indicating that the exclusion restrictions are not violated. Table S3.1. Main Results with the Metro-Level Control of Security Expenses Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Violence −0.109∗∗∗ 0.038 −0.046∗∗ −4.318∗∗∗ −0.061∗ −0.019 (0.033) (0.023) (0.020) (1.084) (0.034) (0.023) N 30,605 28,589 30,605 29,926 30,605 30,605 F -test of excl rest 20.89 20.59 20.89 20.07 20.89 20.89 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Violence −0.068∗∗∗ −0.103∗∗∗ 0.027 0.115∗∗ −0.084∗ 0.018∗∗ (0.022) (0.023) (0.037) (0.054) (0.048) (0.008) N 30,605 29,480 30,118 24,745 24,761 30,605 F -test of excl rest 20.89 21.19 20.94 20.65 20.55 20.89 For both panels: Plant FEs       5-dig. industry × Year FEs       Pre-trends in homicide rate       Security expenses       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and the metropolitan area-level 2005–2010 growth in security expenses interacted with year fixed effects. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. S3.2. Results with Time-Varying State Characteristics Tables S3.2–S3.5 illustrate the robustness of the results after accounting for various time-varying state- level characteristics. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Table S3.2. Results with Time-Varying State Characteristics Panel A (1) (2) (3) (4) (5) (6) (7) Dep. var.: Log employment Log homicide rate −0.069∗∗∗ −0.073∗∗∗ −0.068∗∗∗ −0.087∗∗∗ −0.078∗∗∗ −0.077∗∗∗ −0.088∗∗∗ (0.022) (0.025) (0.021) (0.023) (0.023) (0.023) (0.025) State-level characteristics (in log): State GDP per capita – −0.172 – – – – −0.202 (0.193) (0.237) State manufacturing GDP per capita – – −0.010 – – – 0.056 (0.089) (0.132) Total public expenditures per capita – – – −0.059 – – −0.053 (0.053) (0.057) Federal transfers per capita – – – – −0.081 – −0.081 (0.074) (0.075) Federal transfers for security per capita – – – – – 0.029 0.087 (0.116) (0.147) Observations 30,605 30,605 30,605 24,371 24,371 24,371 24,371 Plant fixed effects        5-dig industry × Year FE        2002 homicide rate × Year FEs        Panel B First stage: Instrument (MexCol) 0.395∗∗∗ 0.374∗∗∗ 0.398∗∗∗ 0.395∗∗∗ 0.388∗∗∗ 0.391∗∗∗ 0.390∗∗∗ (0.086) (0.075) (0.086) (0.090) (0.083) (0.083) (0.085) First-stage F 21.15 24.69 21.53 19.14 22.04 22.25 20.94 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 and state-level public finance data from INEGI. Note: The dependent variable is the logarithm of the number of employees. Log homicide rate is the logarithm of the number of homicides per thousand inhabitants of each metropolitan area. Variables State GDP, State manufacturing GDP, Total public expenditures, Federal transfers, Federal transfers for security are all state-level per capita measures, expressed in logarithmic form. Federal transfers are state revenues obtained from the federal government. Public finance variables are not available for the federal district. Robust standard errors, reported in parentheses, are two-way clustered by local market (metropolitan area) and four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.3. Results When Federal Transfers to States Are Controlled For Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity −0.125∗∗∗ −0.044∗∗ −4.202∗∗∗ −0.066∗∗ −0.021 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Violence 0.036 (0.033) (0.023) (0.021) (1.164) (0.032) (0.025) Total federal transfers −0.213 −0.061 −0.040 −4.487 −0.137 −0.050 (0.158) (0.095) (0.063) (3.747) (0.124) (0.078) N 24,371 22,677 24,371 23,833 24,371 24,371 F -test of excl rest 22.04 21.67 22.04 21.24 22.04 22.04 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Violence −0.078∗∗∗ −0.112∗∗∗ 0.017 0.102∗ −0.077 0.017∗∗ (0.023) (0.023) (0.038) (0.052) (0.047) (0.007) Total federal transfers −0.081 −0.114 −0.073 0.041 −0.059 0.003 (0.074) (0.097) (0.085) (0.148) (0.126) (0.019) N 24,371 23,509 23,989 19,538 19,536 24,371 F -test of excl rest 22.04 22.25 22.04 21.27 21.51 22.04 For both panels: Plant FEs       5-dig. industry × Year FEs       Pre-trends in homicide rate       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 and state-level public finance data from INEGI. Note: Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and the yearly per capita total transfers from the federal government (in log). Federal transfers is not available for the federal district. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area and by four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.4. Results When Federal Transfers for Security Are Controlled For Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity −0.127∗∗∗ −0.043∗∗ −4.214∗∗∗ −0.070∗∗ −0.023 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Violence 0.038 (0.034) (0.025) (0.021) (1.192) (0.034) (0.025) Federal transfers for security −0.023 0.082 0.030 0.528 −0.094 −0.077 (0.202) (0.076) (0.060) (6.616) (0.135) (0.064) N 24,371 22,677 24,371 23,833 24,371 24,371 F -test of excl rest 22.25 21.92 22.25 21.35 22.25 22.25 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Violence −0.077∗∗∗ −0.112∗∗∗ 0.012 0.102∗ −0.075 0.016∗∗ (0.023) (0.023) (0.036) (0.051) (0.046) (0.007) Federal transfers for security 0.029 −0.005 −0.163 0.019 0.104 −0.020 (0.116) (0.184) (0.189) (0.294) (0.278) (0.043) N 24,371 23,509 23,989 19,538 19,536 24,371 F -test of excl rest 22.25 22.63 22.30 22.06 22.11 22.25 For both panels: Plant FEs       5-dig. industry × Year FEs       Pre-trends in homicide rate       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 and state-level public finance data from INEGI. Note: Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and the yearly per capita transfers from the federal government for safety and security (in log). Federal transfers for security is not available for the federal district. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area and by four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.5. Results After Controlling for State-Level GDP Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity −0.117∗∗∗ 0.050∗∗ −0.042∗∗ −3.939∗∗∗ −0.062∗ −0.022 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Violence (0.037) (0.020) (0.020) (1.143) (0.036) (0.024) N 30,605 28,589 30,605 29,926 30,605 30,605 F -test of excl rest 24.69 24.60 24.69 23.60 24.69 24.69 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Violence −0.073∗∗∗ −0.108∗∗∗ 0.027 0.098∗ −0.076 0.017∗∗ (0.025) (0.025) (0.042) (0.053) (0.048) (0.008) N 30,605 29,480 30,118 24,745 24,761 30,605 F -test of excl rest 24.69 25.17 24.74 24.09 24.16 24.69 For both panels: Plant FEs       5-dig. industry × Year FEs       Pre-trends in homicide rate       Log state GDP per capita       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 and state-level public finance data from INEGI. Note: Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and the yearly state GDP per capita in logarithmic form. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. S3.3. Violence as a Negative Blue-Collar Labor-Supply Shock Table S3.6. Violence as a Negative Blue-Collar Labor-Supply Shock (1) (2) (3) (4) Panel A. Both payroll and indirect employees Dependent variable Blue-collar White-collar Blue-collar White-collar Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 workers workers hours hours Violence −0.101∗∗∗ 0.028 −0.102∗∗∗ 0.047 (0.023) (0.036) (0.030) (0.047) No. of observations 29,480 30,118 29,658 25,071 F -excluded instrument 21.44 21.20 20.29 23.25 Panel B. Employees on payroll Dependent variable Blue-collar White-collar Blue-collar White-collar workers workers hours hours Violence −0.119∗∗∗ −0.039 −0.112∗∗∗ 0.008 (0.029) (0.034) (0.028) (0.036) No. of observations 26,186 25,846 25,595 21,148 F -excluded instrument 21.10 21.39 20.25 23.46 Panel C. Monthly wages Dependent variable Avg wage Avg wage Blue-collar White-collar on payroll avg wage avg wage Violence −0.023 0.013 0.106∗∗ −0.080∗ (0.019) (0.021) (0.052) (0.047) No. of observations 29,992 26,077 24,745 24,761 F -excluded instrument 20.73 20.90 20.66 20.74 Panel D. Skill intensity and growth rates Dependent variable Skill intensity Employment Blue-collar White-collar ( NonProduction TotEmp ) growth growth growth Violence 0.017∗∗ −0.032 −0.062 −0.014 (0.007) (0.028) (0.038) (0.072) No. of observations 30,605 24,926 24,090 24,559 F -excluded instrument 21.15 12.37 12.59 12.56 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: All estimations are by two-stage least squares using the instrument as described in Section 3. Violence is measured as the logarithm of the number of homicides per thousand inhabitants of a metropolitan area. All dependent variables are in logarithmic form except the dependent variables in Panel D. Skill intensity is the ratio of total number of white-collar employees over the total employment. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and the pre-trends in the homicide rate per metropolitan area. Robust standard errors, reported in parentheses, are clustered two way by metropolitan area and industry. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. S3.4. Additional Results on Firms’ Exports Table S3.7 reports the estimation results on plant-level exporting in detail. The dependent variable in column (1) is the export dummy, in column (2) it is the share of foreign sales in total sales, in column (3) it is the total number of exported products as a share of total number of products sold, in column (4) Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 it is the logarithm of the foreign sales, and in column (5) it is the logarithm of the number of exported products. Table S3.7 confirms the results presented in the main text that exporting activities are not disproportionately affected by the drug war. Table S3.7. Export and the Drug-War Violence Specification: 2SLS (1) (2) (3) (4) (5) Dep. Var. Export Share of Share of Log export Log number of indicator foreign sales exported products revenue exported products Log homicide rate −0.018 −0.009 −0.020 −0.195 −0.015 (0.023) (0.010) (0.019) (0.137) (0.023) Plant FEs      Pre-trends      5-dig. industry × Year FEs      F -test of excluding statistics 21.15 21.15 21.15 29.28 29.28 N 30,605 30,605 30,605 10,812 10,812 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.5. Employment Elasticity of Violence Table S3.8. Heterogeneity in Employment Elasticity of Violence Dependent variable for all regressions: Log total employment Partition variable Low High Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Panel A. Log monthly wage (p50 = 9.14) ≤p50 >p50 Violence −0.119∗∗ −0.032 (0.048) (0.030) N 14,287 14,284 First-stage F -test 15.50 25.83 Panel B. Female workforce share (p50 = 0.20) ≤p50 >p50 Violence −0.059∗∗ −0.154∗∗ (0.026) (0.077) N 12,105 12,101 First-stage F -test 26.89 13.33 Panel C. Unskilled female production wage (p50 = 70,000 peso) ≤p50 >p50 Violence −0.108∗∗∗ −0.004 (0.023) (0.029) N 15,605 15,000 First-stage F -test 22.55 17.73 Panel D. Unskilled male production wage (p50 = 613,000 peso) ≤p50 >p50 Violence −0.075∗∗∗ −0.065∗∗ (0.023) (0.029) N 16,705 13,900 First-stage F -test 16.63 25.66 Panel E. Share of unionized production workers (p50 = 0.35) ≤p50 >p50 Violence −0.107∗∗∗ −0.030 (0.030) (0.026) N 15,389 15,216 First-stage F -test 18.48 23.98 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Each cell shows the two-stage least squares estimation of the log homicide rate on the logarithm of the total number of employees when the sample is partitioned according to the value of the variable on the left in the respective row. All characteristics are from the start of the period (2005). Each regression includes plant fixed effects, five-digit industry-by-year fixed effects, and the pre-trends. Unionization and unskilled wage data are from ENESTyC, female workforce information is from EIA. Robust standard errors, reported in parentheses, are clustered two way by metropolitan area and industry. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.9. Drug Violence and Heterogeneity in Employment Elasticity—Interaction Approach Spec: 2SLS (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. var.: Log employment Violence −0.065∗∗ −0.069∗∗ −0.093∗∗∗ −0.235∗∗∗ 0.067∗∗ −1.188∗∗∗ −0.040 −0.079∗∗∗ −0.052∗∗∗ (0.024) (0.027) (0.030) (0.055) (0.026) (0.347) (0.026) (0.021) (0.016) Violence × Export −0.009 – – – – – – – – (0.027) Violence × Import – −0.008 – – – – – – – (0.010) Violence × Freight share – – 0.109 – – – – – – (0.086) Violence × log K/L – – – 0.032∗∗∗ – – – – – (0.008) Violence × Labor cost share – – – – −0.746∗∗∗ – – – – (0.099) Violence × Avg. monthly wage – – – – – 0.121∗∗∗ – – – (0.037) Violence × Female workforce share – – – – – – −0.196∗ – – (0.114) Violence × Unskilled female wage – – – – – – – 0.026∗∗∗ – (0.010) Violence × Unskilled male wage – – – – – – – – −0.009∗∗ (0.005) No. of observations 30,605 26,920 26,774 26,557 26,800 28,571 26,795 30,605 30,605 S-W F -test (violence) 78.82 56.83 72.97 65.66 62.53 57.48 81.69 52.19 33.64 S-W F -test (interaction) 92.35 105.89 78.54 60.93 66.71 59.61 70.75 45.30 33.01 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Violence is the log number of homicides per thousand inhabitants of a metropolitan area. All estimations are by two-stage least squares using the instrument as described in Section 3. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. S-W F -test refers to the Sanderson–Windmeijer F -test. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 S3.6. Output Elasticity of Violence Table S3.10. Heterogeneity in Output Elasticity of Violence Dependent variable for all regressions: Log value of output Partition variable Low High Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 ≤p50 >p50 Panel A. Exporters versus non-exporters Non-exporters Exporters Violence −0.166∗∗∗ −0.022 (0.050) (0.039) N 19,775 10,830 First-stage F -test 20.08 26.49 Panel B. Importers versus non-importers Non-importers Importers Violence −0.203∗∗∗ −0.094∗ (0.071) (0.050) N 13,775 13,145 First-stage F -test 16.96 28.00 Panel C. Transport-intensive plants ≤p50 >p50 Share of freight expenses in service expenses (p50 = 0.08) Violence −0.194∗∗∗ −0.105∗∗ (0.071) (0.042) N 13,387 13,387 First-stage F -test 18.02 27.83 Panel D. Geog. diversity of sales (p50 = 0.14) Violence −0.136∗∗∗ −0.078 (0.040) (0.052) N 15,426 15,179 First-stage F -test 20.09 21.67 Panel E. Geog. diversity of materials (p50 = 0.21) Violence −0.127∗∗∗ −0.087∗ (0.045) (0.045) N 15,407 15,198 First-stage F -test 19.53 23.01 Panel F. Log capital per worker (p50 = 4.86) Violence −0.179∗∗∗ −0.084∗∗ (0.058) (0.032) N 13,282 13,275 First-stage F -test 19.95 23.42 Panel G. Labor share in non-capital expenses (p50 = 0.17) Violence −0.019 −0.263∗∗∗ (0.032) (0.090) N 13,401 13,399 First-stage F -test 24.61 19.22 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Each panel shows the two-stage least squares estimations of the log homicide rate on the logarithm of the value of production when the sample is partitioned according to the value of the variable on the left in the respective row. All characteristics are from the start of the period (2005). Each regression includes plant fixed effects, five-digit industry-by-year fixed effects, and the pre-trends. The characteristics in panels B, C, F, and G are from the EIA and the estimation is conducted among the EIA-EIM matched sample. Robust standard errors, reported in parentheses, are clustered two way by metropolitan area and industry. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.11. Drug Violence and Heterogeneity in Output Elasticity—Interaction Approach Spec:2SLS (1) (2) (3) (4) (5) (6) (7) (8) (9) Dep. var.: Log output Violence −0.156∗∗∗ −0.200∗∗∗ −0.186∗∗∗ −0.336∗∗∗ −0.007 −1.008∗∗ −0.109∗∗∗ −0.139∗∗∗ −0.130∗∗∗ (0.042) (0.055) (0.049) (0.070) (0.048) (0.486) (0.039) (0.035) (0.041) Violence × Export 0.101∗∗∗ – – – – – – – – (0.017) Violence × Import – 0.128∗∗∗ – – – – – – – (0.045) Violence × Freight share – – 0.301∗∗ – – – – – – (0.125) Violence × log K/L – – – 0.041∗∗∗ – – – – – (0.011) Violence × Labor cost share – – – – −0.661∗∗∗ – – – – (0.187) Violence × Avg. monthly wage – – – – – 0.097∗ – – – (0.049) Violence × Female workforce share – – – – – – −0.134∗∗ – – (0.060) Violence × Unskilled female wage – – – – – – – 0.068∗∗ – (0.031) Violence × Unskilled male wage – – – – – – – – 0.011 (0.012) No. of observations 30,605 26,920 26,774 26,557 26,800 28,571 26,795 30,605 30,605 S-W F -test (violence) 78.82 56.83 72.97 65.66 62.53 57.48 81.69 52.19 33.64 S-W F -test (interaction) 92.35 105.89 78.54 60.93 66.71 59.61 70.75 45.30 33.01 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Violence is the log number of homicides per thousand inhabitants of a metropolitan area. All estimations are by two-stage least squares (2SLS) using the instrument as described in Section 3. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. S-W F -test refers to the Sanderson–Windmeijer F -test. Robust standard errors, reported in parentheses, are clustered two way by metropolitan area (57) and industry (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 S3.7. Heterogeneous Impact of Violence Across Firms Tables S3.12–S3.13 complement the analysis presented in table 4 and provide evidence supporting the idea that violence affects firms through multiple channels, including the labor-supply channel. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Table S3.12. Demand and Labor-Supply Channels—Additional Results I Spec:2SLS (1) (2) (3) (4) (5) (6) Dep. var. Log Log Log Emp Log Emp Log avg. Log avg. output employ- blue-collar white-collar monthly wages monthly wages ment blue-collar white-collar Panel A. Import Violence −0.200∗∗∗ −0.069∗∗ −0.107∗∗∗ 0.024 0.099 −0.081∗∗ (0.055) (0.027) (0.026) (0.033) (0.062) (0.040) Violence × Importer 0.128∗∗∗ −0.008 −0.000 0.019 0.005 0.004 (0.045) (0.010) (0.013) (0.027) (0.055) (0.046) No. of observations 26,920 26,920 25,944 26,489 21,782 21,812 S-W F -test (violence) 56.83 56.83 53.03 57.13 40.26 41.59 S-W F -test (interaction) 105.89 105.89 113.55 104.79 117.90 115.94 Panel B. Log capital per worker Violence −0.336∗∗∗ −0.235∗∗∗ −0.287∗∗∗ 0.056 0.270 −0.115 (0.070) (0.055) (0.069) (0.071) (0.190) (0.129) Violence × log K/L 0.041∗∗∗ 0.032∗∗∗ 0.036∗∗∗ −0.004 −0.034 0.007 (0.011) (0.008) (0.013) (0.009) (0.037) (0.028) No. of observations 26,557 26,557 25,607 26,136 21,540 21,566 S-W F -test (violence) 65.66 65.66 67.59 68.03 63.40 67.12 S-W F -test (interaction) 60.93 60.93 61.96 62.60 59.01 61.80 Panel C. Female workforce share Violence −0.109∗∗∗ −0.040 −0.087∗∗∗ 0.059 0.085∗ −0.097∗∗ (0.039) (0.026) (0.023) (0.051) (0.047) (0.044) Violence × Female workforce share −0.134∗∗ −0.196∗ −0.123 −0.139 0.102 0.099 (0.060) (0.114) (0.085) (0.145) (0.137) (0.092) No. of observations 26,795 26,795 25,823 26,364 21,710 21,740 S-W F -test (violence) 81.69 81.69 80.97 79.61 71.46 70.78 S-W F -test (interaction) 70.75 70.75 70.21 71.22 63.17 73.97 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Violence is the log number of homicides per thousand inhabitants of a metropolitan area. Estimation by two-stage least squares. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). All plant characteristics are values as of 2005. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and four-digit industry level (84). S-W F -test refers to the Sanderson–Windmeijer F -test. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.13. Demand and Labor-Supply Channels—Additional Results II Spec:2SLS (1) (2) (3) (4) (5) (6) Dep. var. Log Log Log emp Log emp Log avg. Log avg. output employ- blue-collar white-collar monthly wages monthly wages ment blue-collar white-collar Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Panel D. Unskilled male wage Violence −0.130∗∗∗ −0.052∗∗∗ −0.105∗∗∗ 0.040 0.106 −0.094 (0.041) (0.016) (0.023) (0.049) (0.064) (0.065) Violence × Unskilled male wage 0.011 −0.009∗∗ 0.002 −0.007 −0.000 0.008 (0.012) (0.005) (0.006) (0.010) (0.014) (0.018) No. of observations 30,605 30,605 29,480 30,118 24,745 24,761 S-W F -test (violence) 33.64 33.64 34.01 33.64 31.26 31.75 S-W F -test (interaction) 33.01 33.01 32.85 32.86 31.48 32.08 Panel E. Average monthly wage Violence −1.008∗∗ −1.188∗∗∗ −1.057∗∗∗ −0.194 2.112∗∗∗ 0.669∗ (0.486) (0.347) (0.277) (0.201) (0.658) (0.341) Violence × Avg. monthly wage 0.097∗ 0.121∗∗∗ 0.103∗∗∗ 0.024 −0.216∗∗∗ −0.080∗∗ (0.049) (0.037) (0.029) (0.023) (0.069) (0.037) No. of observations 28,571 28,571 27,536 28,125 23,110 23,122 S-W F -test (violence) 57.48 57.48 63.03 57.60 56.40 55.92 S-W F -test (interaction) 59.61 59.61 65.63 59.05 60.71 59.80 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Violence is the log number of homicides per thousand inhabitants of a metropolitan area. Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). All plant characteristics are values as of 2005. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and four-digit industry level (84). S-W F -test refers to the Sanderson–Windmeijer F -test. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. S3.8. Plant Exit To focus on heterogeneity in exit probabilities, I estimate a version of equation (1) where I interact the various plant-level characteristics with the metropolitan-area-level violence as measured by the logarithm of the homicide rate. As exit is a relatively rare event, instead of controlling for five-digit, I control for three-digit industry-by-year fixed effects. The 2SLS estimation results are presented in table S3.14. While this approach ignores the binary nature of the exit variable, it is useful to see how the exit probabilities vary depending on initial plant characteristics. The results show that exit due to the Mexican Drug War is more likely the smaller the plant size. Exit due to violence is also more likely if the plants have a higher share of female employees (column 6) and have lower wages (column 7). These results show that violence- induced labor-supply changes also operate at the extensive margin. The reduction in local market size due to violence is also important in deriving exit as I find that exporters and importers are less likely to exit due to the Mexican Drug War (columns 2–3), and plants are less likely to exit the more diversified they are in output and input markets (columns 4–5). Table S3.14. Heterogeneous Impact of the Mexican Drug War on Plant Exit Probabilities Specification: 2SLS (1) (2) (3) (4) (5) (6) (7) (8) Dep. var.: Plant exit Log homicide rate −0.003 0.031 0.050∗∗∗ 0.037 0.049∗∗∗ 0.000 0.300∗∗∗ 0.026 (0.021) (0.022) (0.012) (0.023) (0.017) (0.012) (0.074) (0.050) Violence × Small (emp ≤ 40) 0.089∗∗∗ Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 – – – – – – – (0.029) Violence × Export – −0.029∗∗∗ – – – – – – (0.010) Violence × Import – – −0.072∗∗∗ – – – – – (0.017) Violence × Sales diversity – – – −0.135∗∗∗ – – – – (0.035) Violence × Material diversity – – – – −0.146∗∗∗ – – – (0.049) Violence × Share of female workforce – – – – – 0.065∗ – – (0.036) Violence × Avg. monthly wage – – – – – – −0.031∗∗∗ – (0.007) Violence × log K/L – – – – – – – −0.003 (0.009) S-W F -test (violence) 39.83 52.64 34.89 34.06 48.64 56.01 28.65 36.88 S-W F -test (interaction) 58.24 58.03 87.87 28.94 38.15 52.55 29.30 33.66 N 25,979 25,979 22,831 25,979 25,979 22,735 24,316 22,530 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010 supplemented with Encuesta Industrial Anual (EIA) 2005 and Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Estimation by two-stage least squares (2SLS). Exit is an indicator variable that takes the value 1 if a plant exits the next period; as a result it is not defined in year 2010. All regressions include plant fixed effects, three-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). All plant-level characteristics are as of year 2005. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). S-W F -test refers to the Sanderson–Windmeijer F -test. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.9. Results Without the Greater Mexico City While most of the metropolitan commuting zones consist of either a single municipality or a center mu- nicipality and one or two adjacent smaller municipalities, zona metropolitan del valle de México is an exception. It covers a large urban area encompassing 60 adjacent municipalities. As an important in- dustrial center of Mexico, a significant number of manufacturing firms are located in the Mexico City metropolitan area. What happens to the results if firms located in this area are excluded from the sam- ple? Table S3.15 presents 2SLS results when the estimation sample excludes firms located in the greater Mexico City. The results carry over smoothly when firms located in the greater Mexico City are excluded. Table S3.15. Robustness Analysis—Excluding Firms in Zona Metropolitan del Valle de México Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity −0.141∗∗∗ 0.044∗∗ −0.051∗∗ −4.235∗∗∗ −0.076∗∗ −0.033 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate (0.042) (0.021) (0.024) (1.405) (0.035) (0.029) No. of observations 19,111 17,795 19,111 18,701 19,111 19,111 First-stage F 21.79 21.57 21.79 21.04 21.79 21.79 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.087∗∗∗ −0.121∗∗∗ 0.019 0.091∗ −0.062 0.019∗∗ (0.029) (0.028) (0.048) (0.052) (0.057) (0.008) No. of observations 19,111 18,442 18,793 15,255 15,250 19,111 First-stage F 21.79 22.08 21.88 20.68 20.79 21.79 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       No. of metros (clusters) 56 56 56 56 56 56 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS); the sample excludes Mexico City. All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (56) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.10. Robustness Analysis—Dropping One State at a Time Figure S3.1 illustrates the changes in the estimates of employment and output elasticity when plant-level data in each state are sequentially excluded from the analysis. Figure S3.1. Effects of Violence When One State at a Time Is Dropped. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 (a) (b) Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (1) by two-stage least squares. Coefficient estimates and 95 percent confidence intervals are shown. Each coefficient estimate is obtained when one state at a time is dropped from the sample for each of the 32 states. The number observations vary between 24,371 and 30,605. First-stage F -statistics vary between 13.16 and 25.14. The dependent variables are the natural logarithm of employment in (a) and the natural logarithm of the value of production in (b). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate. The log homicide rate is instrumented as described in equation (2). S3.11. Plant Exit and the Impact at the Intensive Margin I show that plants that are exposed to the violence shock are more likely to exit, and that the likelihood of exit is stronger if the plants are smaller, more female-intensive, and oriented towards the domestic market rather than exporting and importing. I also show that such plants disproportionately downsize conditional on staying in the market. These results may imply that selection may be leading to underestimation of the true effect at the intensive margin. To gauge this, I use the “identification-at infinity” idea (Chamberlain Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 1986; Mulligan and Rubinstein 2008) that the selection bias must be lower for plants with higher survival probability and restrict the estimation sample to plants with higher survival probability and observe how the estimates change as one drops, step by step, the plants that most likely exit. Table S3.16 presents the results when plants are allocated in sub-samples depending on their average probability of exit across the sample years. The results suggest that to some extent the endogenous exit is likely to lead to understating the true impact at the intensive margin as the coefficient estimates get larger for employment and output Table S3.16. Exit Likelihood and the Impact at the Intensive Margin Specification: 2SLS (1) (2) (3) (4) (5) Exit prob All Except top 1 percent Except top 5 percent Except top 10 percent Except top 15 percent Panel A Dep. var.: Value of output Log homicide rate −0.112∗∗∗ −0.128∗∗∗ −0.121∗∗∗ −0.133∗∗∗ −0.136∗∗∗ (0.033) (0.037) (0.039) (0.039) (0.041) First-stage F 21.15 21.44 21.98 22.38 23.39 N 30,605 26,289 25,228 23,902 22,572 Panel B Dep. var.: Employment Log homicide rate −0.069∗∗∗ −0.075∗∗∗ −0.072∗∗∗ −0.068∗∗∗ −0.068∗∗∗ (0.022) (0.025) (0.024) (0.023) (0.023) First-stage F 21.15 21.44 21.98 22.38 23.39 N 30,605 26,289 25,228 23,902 22,572 Panel C Dep. var.: Blue-collar employment Log homicide rate −0.101∗∗∗ −0.110∗∗∗ −0.106∗∗∗ −0.103∗∗∗ −0.104∗∗∗ (0.023) (0.027) (0.029) (0.029) (0.029) First-stage F 21.44 21.95 22.48 22.96 23.96 N 29,480 25,344 24,304 23,002 21,691 Panel D Dep. var.: Blue-collar wages (on payroll) Log homicide rate 0.106∗∗ 0.104∗∗ 0.100∗∗ 0.081∗ 0.092∗∗ (0.052) (0.047) (0.047) (0.049) (0.045) First-stage F 20.66 20.77 21.42 21.86 22.50 N 24,745 21,336 20,377 19,192 18,043 Panel E Dep. var.: White-collar employment Log homicide rate 0.028 0.037 0.035 0.039 0.033 (0.036) (0.038) (0.038) (0.040) (0.043) First-stage F 21.20 21.65 22.22 22.58 23.37 N 30,118 25,912 24,890 23,602 22,324 Panel F Dep. var.: White-collar wages (on payroll) Log homicide rate −0.080∗ −0.083∗ −0.078∗ −0.080∗ −0.067 (0.047) (0.043) (0.042) (0.045) (0.050) First-stage F 20.74 21.09 21.78 22.20 22.89 N 24,761 21,358 20,397 19,206 18,056 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). All dependent variables are in logarithmic form. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area and by four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. impact of violence. As such, one can interpret the findings in the paper as the lower bound of the real impact. S3.12. Accounting for US-Border-Specific Shocks Table S3.17 presents 2SLS results when equation (1) is augmented with US border indicators interacted with time fixed effects to account for shocks that may be affecting US border areas disproportionately. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 The findings affirm the robustness of the results, even when accounting for differential shocks experienced by metropolitan areas along the US border. Table S3.17. Robustness Analysis—Including Border-Specific Aggregate Shocks Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Log homicide rate −0.096∗∗ 0.052∗∗∗ −0.034 −2.814∗∗ −0.064∗ −0.016 (0.037) (0.015) (0.022) (1.136) (0.037) (0.027) No. of observations 30,605 28,589 30,605 29,926 30,605 30,605 First-stage F 18.97 18.27 18.97 18.53 18.97 18.97 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.053∗∗ −0.095∗∗∗ 0.064∗ 0.114∗ −0.075 0.021∗∗ (0.023) (0.028) (0.037) (0.061) (0.059) (0.008) No. of observations 30,605 29,480 30,118 24,745 24,761 30,605 First-stage F 18.97 19.13 19.04 18.46 18.79 18.97 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       US border j × Year FEs       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). The specification additionally includes the border-by-year fixed effects. All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.13. Accounting for the Informal Economy To explore the potential influence of the informal economy on the observed results, given the absence of time-varying data on the informal economy across local labor markets during the study period, the estimated size of the informal economy for each local labor market as of 2013 (the earliest available data) is interacted with time fixed effects and incorporated into equation (1). This method allows for the assessment of differential impacts based on the share of the informal sector within each local area. The size Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 of the informal economy, expressed as a percentage of informal value-added in each metropolitan area, is estimated by Gonzalez and Llamosas-Rosas (2021) using satellite images of nighttime lights. The findings, as detailed in table S3.18, demonstrate that the presence of the informal economy does not significantly affect the results, underscoring the robustness of the initial findings against the potential effects of informal economic activities. Table S3.18. Robustness Analysis—Controlling for the Local Informal Economy Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Log homicide rate −0.122∗∗∗ 0.027 −0.039∗∗ −4.450∗∗∗ −0.054 −0.022 (0.040) (0.022) (0.018) (1.391) (0.038) (0.022) No. of observations 30,467 28,457 30,467 29,790 30,467 30,467 First-stage F 18.69 18.64 18.69 17.92 18.69 18.69 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.088∗∗∗ −0.120∗∗∗ 0.008 0.096∗ −0.067 0.017∗∗ (0.027) (0.028) (0.041) (0.055) (0.048) (0.008) No. of observations 30,467 29,345 29,984 24,637 24,654 30,467 First-stage F 18.69 19.07 18.79 18.79 18.72 18.69 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       InformalEcon × Year FEs       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. The size of the informal economy is from Gonzalez and Llamosas- Rosas (2021, table 2) and calculated for 55 metropolitan areas. Note: Estimation by two-stage least squares (2SLS). The specification additionally includes the border-by-year fixed effects. All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate for each metropolitan area j (2002 homicide rate × Year FEs), and the size of the informal economy for each metropolitan area j interacted with year fixed effects. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (55) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. Table S3.19. Robustness Analysis with Additional Differential Time Trends for Exporters Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Output Product Capacity Labor Export (in log) price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate −0.114∗∗∗ 0.035 −0.046∗∗ −4.009∗∗∗ −0.068∗ 0.019 (0.033) (0.023) (0.020) (1.027) (0.036) (0.019) No. of observations 30,605 28,589 30,605 29,926 30,605 30,605 F -test of excl. restr 21.19 20.89 21.19 20.37 21.19 21.19 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.065∗∗∗ −0.099∗∗∗ 0.031 0.104∗∗ −0.083∗ 0.017∗∗ (0.023) (0.023) (0.037) (0.052) (0.046) (0.007) No. of observations 30,605 29,480 30,118 24,745 24,761 30,605 F -test of excl. restr 21.19 21.48 21.23 20.69 20.76 21.19 Indicators for both panels: Plant fixed effects       5-digit industry × Year FEs       Pre-trends in homicide rate       Exporter × Year FEs       No. of metros (clusters) 57 57 57 57 57 57 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All dependent variables, except Capacity utilization, and Export are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and Exporter time trends (Exporter as of 2005 × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.14. Differential Time Trends for Exporters Table S3.19 shows the results when differential time trends for exporters are additionally controlled for. Here, I allow for differential time trends for each exporter by interacting plants’ exporting status in 2005 with year fixed effects. The results are robust. Table S3.20. Robustness Analysis—Estimation Based on Year 2005 and 2010 Data Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate −0.071∗∗∗ 0.037∗ −0.042∗∗∗ −2.773∗∗∗ −0.032 −0.015 (0.021) (0.019) (0.015) (0.783) (0.021) (0.018) No. of observations 10,109 9,445 10,109 9,773 10,109 10,109 F -test of excl. restr 44.60 44.70 44.60 44.35 44.60 44.60 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.044∗∗∗ −0.074∗∗∗ 0.030 0.073∗ −0.060 0.015∗∗∗ (0.016) (0.018) (0.020) (0.036) (0.042) (0.005) No. of observations 10,109 9,774 9,951 8,148 8,155 10,109 F -test of excl. restr 44.60 45.09 44.11 44.09 42.75 44.60 Indicators for both panels: Plant fixed effects       5-digit industry × Year FEs       Pre-trends in homicide rate       No. of metros (clusters) 57 57 57 57 57 57 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All dependent variables, except Capacity utilization, and Export are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.15. Analysis with Only Data from 2005 and 2010 Table S3.20 presents the 2SLS estimation of equation (1) using only data from the years 2005 and 2010. S3.16. Violence Outbreak and Trade Shocks In this section I address the concern that other local market shocks may be confounding the results. In particular, Utar and Torres Ruiz (2013) show that rising import competition in the United States has a substantial impact in Mexico via maquiladoras, export processing plants in Mexico that are tied to the US market. Recently, Dell, Feigenberg, and Teshima (2019) have found that areas that encounter decline in employment due to the Chinese import competition shock in the US market also suffer from heightened drug violence. Since the results here are robust to eliminating all potential changes happening at the product-by-year level, it is very unlikely that such effects play a role. Regardless, I conduct a further robustness check by constructing a metropolitan-area-level import competition shock due to China’s rise in the US market. Let TradeComp j be the per worker measure of change in trade competition between 2005 and 2010. Following Utar and Torres Ruiz (2013) and Dell, Feigenberg, and Teshima (2019), I use the following measure of trade competition: L jk,ini MCHUS 05–10 TradeComp j = , Lk,ini L j,ini k MCH j,2005 05–10 MCHUS = ∗ [TotMCH2010 − TotMCH2005 ], TotMCH2005 Table S3.21. Robustness Analysis—Additional Differential Time Trends for China-Shock Exposed Areas Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate −0.133∗∗∗ 0.051∗∗∗ −0.052∗∗ −3.710∗∗ −0.078∗ −0.023 (0.046) (0.017) (0.024) (1.418) (0.042) (0.029) No. of observations 30,605 28,589 30,605 29,926 30,605 30,605 First-stage F 20.78 19.81 20.78 20.06 20.78 20.78 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.077∗∗ −0.114∗∗∗ 0.031 0.111∗ −0.050 0.021∗∗ (0.032) (0.031) (0.052) (0.061) (0.059) (0.009) No. of observations 30,605 29,480 30,118 24,745 24,761 30,605 First-stage F 20.78 21.07 20.82 20.91 20.99 20.78 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       TradeComp j × Year FEs       No. of metros (clusters) 57 57 57 57 57 57 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and a measure of import competition shock due to China’s rise in the US market ( TradeComp j × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. where L jk,ini is the employment of industry k in metropolitan area j at the initial year, Lk,ini is total initial employment of industry k in Mexico, and L j,ini is total non-agricultural employment in metropolitan area j. The variable 05–10 MCHUS is the predicted change in Chinese imports in the United States in industry k between 2005 and 2010.49 A higher value of TradeComp j means that a metropolitan area has a larger initial share of employment in industries where Chinese imports in the United States are predicted to grow. I then interact TradeComp j with year fixed effects and include this in equation (1) and re-estimate the impact of violence shock as proxied by the logarithm of the homicide rate. The logarithm of the homicide rate is instrumented as described in Section 3. Estimates that are presented in table S3.21 re-confirm that the results are robust. S3.17. Municipality-Level Results Table S3.22 presents 2SLS results when local labor markets are defined as municipality. Accordingly, both the homicide rate and the pre-trends are calculated at the municipality level and the standard er- rors are clustered at the same level. The results are robust. Additional results are also available upon request. 49 Industry k denotes four-digit NAICS industry. Initial employment shares are calculated using Economic Census 2004 (INEGI). Table S3.22. Robustness Analysis—Municipality-Level Violence Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate −0.090∗∗ 0.020 −0.037∗∗ −3.239∗∗∗ −0.056 −0.020 (0.035) (0.029) (0.015) (0.946) (0.036) (0.017) No. of observations 30,048 28,053 30,048 29,381 30,048 30,048 First-stage F 25.66 25.06 25.66 24.80 25.66 25.66 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.053∗∗∗ −0.089∗∗∗ 0.029 0.101∗∗ −0.086∗∗ 0.016∗∗ (0.015) (0.020) (0.029) (0.046) (0.039) (0.007) No. of observations 30,048 28,950 29,577 24,310 24,329 30,048 First-stage F 25.66 25.81 25.95 24.69 24.98 25.66 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). Municipality-level homicide rate is used. All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). Robust standard errors, reported in parentheses, are two-way clustered by municipality (219) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.18. An Alternative Instrumental-Variable Strategy This section examines the robustness of the findings using alternative instruments. An alternative method to assess the susceptibility of the metropolitan areas to the drug-war violence in- volves considering their proximity to the United States. The competitive advantage of Mexican DTOs lies in trafficking routes to the United States, and the contention for these routes is likely to escalate the closer the routes are to the US border, hence the drug violence. By interacting the distance of metropolitan areas to the US border with the post-Calderón period that takes the value 1 on and after 2007, an alternative instrument is constructed to capture the unexpected increase in drug violence after the implementation of the kingpin strategy. Table S3.23 presents the results obtained when this instrument that varies at the local labor market is employed to predict the heightened violence attributable to the drug war. The first- stage F -statistics, reported in table S3.23, indicate a strong correlation between the instrument and the homicide rate with robust instrumentation, and the second-stage results align with those obtained using the default instrument. Table S3.23. Instrument Based on the Distance to the US Border and the Post-Calderón Term Specification: 2SLS 2SLS 2SLS 2SLS 2SLS 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Log homicide rate −0.121∗∗∗ −0.013 −0.062∗∗∗ −4.315∗∗∗ −0.058∗∗ 0.002 (0.037) (0.035) (0.021) (1.348) (0.028) (0.017) No. of observations 30,605 28,589 30,605 29,926 30,605 30,605 First-stage F 25.32 25.89 25.32 24.58 25.32 25.32 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Log homicide rate −0.078∗∗ −0.095∗∗∗ −0.006 0.088 −0.089 0.013∗ (0.030) (0.028) (0.040) (0.059) (0.058) (0.006) No. of observations 30,605 29,480 30,118 24,745 24,761 30,605 K-P F -excluded instrument 25.32 25.98 25.17 26.44 25.69 25.32 Indicators for both panels: Plant fixed effects       5-dig. industry × Year FEs       Pre-trends in homicide rate       No. of metros (clusters) 57 57 57 57 57 57 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation by two-stage least squares (2SLS). All dependent variables, except Capacity utilization, and the Export indicator are in logarithmic form. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, and pre-trends in the homicide rate (2002 homicide rate × Year FEs). The instrument is the distance of each metropolitan area to the US border interacted with an indicator for post-2007. K-P F -excluded instrument refers to the Kleibergen–Paap F -test of excluded instrument. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels respectively. S3.19. Analysis with the Monthly Data For the monthly analysis, I use the EIMA collected at the monthly frequency and monthly data on homi- cides across Mexican municipalities to construct the monthly homicide rate at the metropolitan area level. Table S3.24 presents the OLS analysis, while table S3.25 provides the full set of results from the instrumental variable analysis, where the monthly data on cocaine seizures from the Colombian Defense Ministry is constructed as the instrument. The instrument otherwise follows equation (2). In order to study the short-run impact of violence, I utilize the high-frequency plant-level data and estimate the following equation: ln Yikjtm = β0 + β1 HomicideRatejtm + Xjt + μtm + τkt + ηi + ikjtm , (S3.1) where tm denotes monthly frequency time and μtm denotes time fixed effects at the monthly frequency. To see the short-run impact and how it varies depending on the current versus lagged homicide rate, table S3.24 presents the results from estimating equation (S3.1) using OLS where the dependent variable is the logarithm of monthly employment. Table S3.24 provides both the concurrent effect of the homicide rate and the results when the homicide rate is lagged for 1 to 6 months separately. The results show that the OLS coefficient estimate is biggest when the homicide rate is lagged for 2 months and the magnitude is still bigger than the concurrent effect when it is lagged for 6 months. Column (6) of the table presents when both the concurrent and lagged homicide rates are included simultaneously. These results indicate that it makes sense to use lagged homicide rate, rather than relying on a concurrent rate. In the annual analysis, the 6-month-lagged homicide rate is used to further break endogeneity concerns, but the results are not sensitive to this choice, as also indicated by table S3.24. Table S3.24. Concurrent and Lagged Effects of Homicide Rate on Employment (1) (2) (3) (4) (5) (6) (7) (8) Specification: OLS, monthly Dependent variable: Log monthly employment Monthly homicide rate −0.563∗∗∗ −0.140 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Concurrent – – – – – – (0.129) (0.133) 1-month lag – −0.619∗∗∗ – – – – – −0.175∗ (0.121) (0.095) 2-month lag – – −0.658∗∗∗ – – – – −0.189∗∗∗ (0.092) (0.063) 3-month lag – – – −0.638∗∗∗ – – – −0.074 (0.083) (0.044) 4-month lag – – – – −0.645∗∗∗ – – −0.122∗ (0.086) (0.069) 5-month lag – – – – – −0.633∗∗∗ – −0.131 (0.082) (0.079) 6-month lag – – – – – – −0.596∗∗∗ −0.013 (0.063) (0.084) Plant FEs         Monthly time FEs         2002 homicide rate × Year FEs         5-dig. industry × Year FEs         Time-varying local market characs.         No. of observations 334,195 334,195 334,195 334,195 334,195 334,195 334,195 334,195 No. of clusters (LM) 57 57 57 57 57 57 57 57 Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (S3.1) using ordinary least squares (OLS) with the dependent variables the logarithm of monthly total employment. Estimation includes plant fixed effects, monthly time fixed effects, 5-digit-industry-by-time fixed effects, pre-trends in the homicide rate, and the time-varying local-labor-market-level strategic sector controls. Robust standard errors, reported in parentheses, are two-way clustered by local market (metropolitan area) and four-digit industry level. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. Table S3.25 presents a summary of estimation results when the analysis is conducted at the monthly frequency. In these results, the log of monthly homicide rate is instrumented with the same instrument as described in the main text, except that now I use the monthly amount of cocaine seized by Colombian forces. Here, too, both the homicide rate and the cocaine seizures are lagged by six months. Elasticity estimates shown in table S3.25 are similar to those found in the main analysis. Table S3.25. Main Results with the Monthly Data Specification: 2SLS (1) (2) (3) (4) (5) (6) Panel A Output Avg. output Product Capacity Labor Export price scope utilization productivity Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Violence −0.138∗∗∗ 0.029 −0.027 −4.189∗∗∗ −0.075∗ −0.018 (0.045) (0.021) (0.023) (1.231) (0.041) (0.020) N 334,046 310,715 334,046 326,047 334,046 334,046 F -test of excl rest 10.50 10.52 10.50 9.97 10.50 10.50 Panel B Total Blue-collar White-collar Blue-collar White-collar Skill employment employment employment wage wage intensity Violence −0.075∗∗∗ −0.121∗∗∗ 0.036 0.119∗ −0.106∗ 0.020∗ (0.026) (0.038) (0.050) (0.063) (0.062) (0.012) N 334,046 321,513 327,680 267,200 267,336 334,046 F -test of excl rest 10.50 10.74 10.51 10.22 10.22 10.50 For both panels: Plant FEs       5-dig. industry × Year FEs       Monthly time FE       Pre-trends in homicide rate       Source: Author’s analysis based on the Encuesta Industrial Mensual Ampliada (EIMA) 2005–2010. Note: Estimation of equation (S3.1) using monthly frequency data by two-stage least squares (2SLS) where the dependent variable is the logarithm of the homicide rate. The instrument is the monthly version of the instrument described in the text. Both the homicide rate and the Colombian cocaine seizure data used in the instrument are lagged by six months. All regressions include plant fixed effects, five-digit industry-by-year fixed effects, pre-trends in the homicide rate (2002 homicide rate × Year FEs), and monthly time fixed effects. Robust standard errors, reported in parentheses, are two-way clustered by metropolitan area (57) and by four-digit industry level (84). ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10 percent, 5 percent, and 1 percent levels, respectively. S4. Data Appendix S4.1. Plant-Level Data S4.1.1. EIMA 2005–2010 La Encuesta Industrial Mensual Ampliada (EIMA) is a monthly survey of manufacturing plants carried out by INEGI. It constitutes the basis of gross domestic product and economic indicators on, among others, employment, production, and productivity. It includes 230 economic classes of activity (clases de Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 actividad) and covers 7,328 establishments that produce 86 percent of the nationwide manufacturing value-added. Industries in the data are classified based on the North American Industry Classification System, SCIAN 2002.50 It was developed jointly by the US Economic Classification Policy Committee (ECPC), Statistics Canada, and Mexico’s Instituto Nacional de Estadistica y Geografia (INEGI) to allow for a high level of comparability in business statistics among the North American countries. Each of 230 economic classes within the manufacturing sector has a unique six-digit number. For example, 311320 refers to “Preparation of chocolate and chocolate products from cacao” and 311330 refers to “Preparation of chocolate products from chocolate.” For each detailed manufacturing activity, clase, plants are ranked according to their production capacity as of the Economic Census 2004 and they are included in the survey from the top until at least 80 percent of all production within each detailed product category is covered. If a plant employs 300 or more employees, they are always included in the survey. EIMA provides information on the number of white-collar and blue-collar workers, wages, hours and days worked, and plant capacity utilization. Importantly, EIMA reports quantity and value of production, sales, and export separately for each product that a plant produces. For example, within economic activity 311330 “Preparation of chocolate products from purchased chocolate,” there are more than 30 products specified, e.g. chocolate covered almonds (311330023) and chocolate covered raisins (311330025). Using this information, it is possible to construct plant-level prices for each product. In recent years there have been important changes in the way companies are organized. One of the most important is related to outsourcing of personnel. EIMA captures information both of the personnel dependent on the corporate name and that provided by a personnel service provider, so that both of these two components of the personnel employed in the manufacturing sector are in the data set. Plant-level wages, salaries, and benefits are deflated using the consumer price index and expressed in thousand 2010 peso. Plant-level sales and production values are deflated using the industry-level producer price deflators and expressed in thousand 2010 peso. The consumer and producer price indices are from Banco de Mexico. S4.1.2. Employment Classifications INEGI uses separate categories for workers who are directly employed and on company payroll versus workers who work in the establishment but are not on payroll. In addition, within each of these categories, workers are classified depending on their job types. Workers whose tasks are directly related to production activities, such as unskilled and skilled workers and technicians working on the production floor, are classified as blue-collar workers. Workers who work on non-production activities and activities that are auxiliary to production, such as workers in administrative and management positions and service workers, are classified as white-collar workers. S4.1.2.1. Aggregation of Monthly Plant-Level Data Plant-level information on the monthly value of production, sales, monthly wage payments are aggregated into annual data by summing them through their values over the 12 months. Employment information (such as the total number of employees, the total number of white-collar employees, the total number of 50 Sistema de Clasificación Industrial de América del Norte. blue-collar employees, the total number of hours worked, the total number of blue-collar hours, the total number of white-collar hours) are averaged across the 12 months between January and December. The monthly data end in June 2010. For the year 2010, the data are annualized by assuming that the second half of 2010 is the same as the first half of the year. S4.1.3. EIA 2005 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 La Encuesta Industrial Anual (EIA) is an annual survey of manufacturing plants carried out by INEGI. It provides detailed balance sheet information for the manufacturing plants, including information on employment, fixed assets, wages, itemized expenses, itemized income, value of production, and invento- ries. The industry classification of plants is based on the North American Industry Classification System (NAICS), 2002. This survey runs on the same sample rules over 2003–2007 with the EIMA, used in this study, hence EIA and EIMA can be matched at the plant level using the unique plant identification system. I enrich the initial plant characteristics with the data from EIA 2005. These data include gender compo- sition of workforce, capital items, and detailed account of expenditures. When the EIA data set is used, the analysis is conducted in the matched EIMA-EIA sample, which is somewhat smaller than the main sample but with very similar characteristics overall. Table S4.1 presents the summary statistics from the EIMA-EIA matched sample. Table S4.1. Comparison of EIMA and the EIMA-EIA Matched Sample EIMA sample EIMA-EIA matched sample N = 30,605 N = 26,920 Mean Median St dev Mean Median St dev Number of employees 238.36 99.83 491.39 240.88 101 481.69 Number of blue-collar employees 159.56 64.67 322.90 161.70 65.33 330.54 Number of white-collar employees 71.92 22.92 229.18 72.37 23.08 207.89 Number of days worked 280.48 295 55.58 280.97 295.00 55.24 Capacity utilization rate 70.20 75 21.11 70.02 75.00 21.23 Number of varieties 3.13 2 3.02 3.13 2.00 3.03 Log value of output 11.25 11.27 2.05 11.27 11.29 2.05 Source: Encuesta Industrial Mensual Ampliada (EIMA) and Encuesta Industrial Anual (EIA). Note: All values are expressed in 2010 thousand Mexican peso. The table shows the summary statistics of the main variables in the estimation sample (metropolitan areas). Table S4.2. Unionization Rates Across Selected Industries Industry Unionization rate (production workers) Sawmills and wood preservation 0.06 Seafood product preparation and packaging 0.06 Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Leather and hide tanning and finishing 0.17 Architectural and structural metals manufacturing 0.17 Other nonmetallic mineral product manufacturing 0.22 Pesticide, fertilizer, and other agricultural chemical manufacturing 0.23 Agriculture, construction, and mining machinery manufacturing 0.28 Pharmaceutical and medicine manufacturing 0.28 Textile furnishings mills 0.38 Lime and gypsum product manufacturing 0.38 Iron and steel mills and ferroalloy manufacturing 0.42 Converted paper product manufacturing 0.46 Fiber, yarn, and thread mills 0.47 Pulp, paper, and paperboard mills 0.51 Fabric mills 0.51 Source: Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (ENESTyC) 2005. Note: Unionization rate is the number of union member production workers over the total number of production workers. Author’s calculation. S4.1.4. ENESTyC 2005 The Encuesta Nacional de Empleo, Salarios, Tecnología y Capacitación en el Sector Manufacturero (EN- ESTyC) is a representative establishment-level survey of manufacturing firms conducted in 1995, 1999, 2001, and 2005. This study employs ENESTyC 2005, which is representative based on 2004 economic census information and covers 9,920 manufacturing establishments and 685 maquiladoras. This survey is used to derive the sales and material entropy measures as it reports the geographic distribution of sales and material purchases (see below for details). The survey also reports wages across detailed occupation– gender categories within plants, as well as unionization rates across different types of employees within plants. Using this data I calculate the average wages paid among unskilled female and male workers and the unionization rates among the production workers. Table S4.2 presents the unionization rate among production workers across selected industries. The plant-level match between EIMA and ENESTyC is possible for a subset of ENESTyC establishments. Since in this match there is a systematic bias toward bigger firms, rather than using the plant-level match I incorporate the ENESTyC characteristics with the main data set via the establishments’ four-digit industries of operation. S4.2. Spatial and Regional Data S4.2.1. Distance to the US Border I select more than 130 points along the US border with latitude and longitude information and obtain the position of each locality (village) in Mexico (degrees/minutes/seconds (DMS) ) from INEGI. After converting the DMS measure to decimal degrees, I use the Haversine formula to calculate the great circle distance from each urban Mexican village (locality) to around 130 US border points.51 I then take the distance between each municipality’s position and the closest border point. S4.2.2. Homicide Rates Information on the number of homicides by municipality and month is obtained from INEGI. Homicide rates used in the descriptive analysis throughout the paper are calculated as the number of homicides in 51 I also use the Pythagorean theorem to calculate the distance in kilometers, obtaining very similar results. Figure S4.1. Organized Crime Related Violence in Mexico. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: Data from INEGI, Milenio, and Reforma. Note: The figure shows the number of total and organized-crime-related homicides.. 100,000 people. Homicide rates used in the regressions are re-scaled and they are the number of homicides in 1,000 people. Municipality-level annual population numbers are calculated using the census data for years 1990, 1995, 2000, 2005, and the annual state-level population estimates of INEGI. INEGI also provides the number of intentional homicides by occupation of victims at the nationwide level. These data are used in preparing the data underlying fig. S1.5. The two newspapers Reforma and Milenia also provide the state-wide number of organized-crime- related homicides since the start of the drug war (see fig. S4.1). Since the data on organized-crime-related homicides do not cover the pre-drug-war time period and do not have detailed geography information, it is not suitable in this analysis. Also note that my IV strategy aims at capturing the variation in the homicide rate that is related to the organized crime as it focuses on the Mexican Drug War. S4.2.3. Metropolitan Area-Level Data The analysis makes use of a set of time-varying metropolitan-area-level variables. These are the annual information on the metropolitan-area-level employment shares of crop production, metal mining includ- ing gold, silver, copper, and uranium, as well as oil and natural gas extraction. The sources of annual data on municipality-level employment across industries are the records of contributions to the Mexican Institute of Social Security (IMSS). The industry classification used in these data is the Mexican version of the North American Industrial Classification System (SCIAN) in its 2007 revision. INEGI is the source of the additional municipality-level variables, which include the number of strikes, the number of regis- tered vehicles, the number of traffic accidents, the number of traffic accidents due to bad road conditions, and high-school success rate. Whenever used in the plant-level analysis, these data are aggregated at the metropolitan level using the key provided by INEGI, matching municipalities with metropolitan areas. Per capita security and public expenditure data come from Ted Enamorado, Luis F. López-Calva, Carlos Rodríguez-Castelán, and Hernán Winkler’s study, titled “Income Inequality and Violent Crime: Evidence Figure S4.2. Cocaine Seizure in Colombia. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: The cocaine seizure information is from the Colombian Ministry of Defense. Note: The figure plots the evolution of the annualized cocaine seizure in Colombia. from Mexico’s Drug War,” published in 2016 in the Journal of Development Economics. The data are reported at a five-year frequency between 1990 and 2010 in real terms as of August 2010. I use the data between 2005 and 2010. Using the metropolitan-area and municipality-level population information, I converted the data into per capita values for each metropolitan area. The size of the informal economy for each of the 55 metropolitan areas is derived from the work of Gonzalez and Llamosas-Rosas (2021),52 utilizing estimates based on satellite images of nighttime lights as of 2013. S4.2.4. State-Level Data The annual state-level GDP and manufacturing GDP data are sourced from INEGI and cover all states for the years 2005–2010. State-level public finance data are available for the 31 states, excluding the federal district. S4.3. Time-Series Data S4.3.1. Cocaine Data Cocaine prices are purity-adjusted prices of a gram of cocaine in the United States. The quarterly data are obtained from the annual reports of the National Drug Intelligence Center. The annual data source is the US Office of National Drug Control Policy, the data obtained from the United Nations Office on Drugs and Crime (UNODC 2014). Cocaine-seizure data are from Castillo, Mejia, and Restrepo (2020). The source of data is the Colom- bian Ministry of Defense, Acción Social, Comando General de las Fuerzas Militares, Fuerza Aérea Colom- biana, Armada Nacional y Naciones Unidas. The seizure data are reported at the monthly frequency between 1999 and 2012. Figure S4.2 shows the evolution of the annualized cocaine seizure in Colombia. 52 Gonzalez and Llamosas-Rosas (2021) . “Observando la Evolucion del Sector Informal desde el Espacio: Un Enfoque Municipal 2013-2020,” Working Papers No. 2021-18, Banco de Mexico. Figure S4.3. Net Coca-Cultivated Land in Colombia. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: Data are from the US Department of State, Bureau of International Narcotics and Law Enforcement Affairs, 2015 International Narcotics Control Strategy Report [INCSR] (March 2015). I also obtain information on the net coca-cultivated land between 1986-2012 in the Andean region from the 2015 Data Supplement of National Drug Control Strategy, an annual report prepared by the Office of National Drug Control Policy. Figure S4.3 shows the evolution of coca-cultivated land over 2004 to 2012. Figure S4.4 shows the moving average of annualized cocaine seizure normalized by net coca-cultivated land in Colombia. S4.3.2. Occupation Data The data on the total number of employees per occupation is obtained from INEGI. The occupation information is used to calculate the risk to life per occupation presented in the supplemental appendix. The underlying source of this data is the National Survey of Occupation and Employment (Encuesta Nacional de Ocupación y Empleo, ENOE). S4.4. Construction of Variables S4.4.1. Construction of Entropy Measures of Diversification The nationwide representative survey ENESTyC 2005 reports the percentage of sales for each plant, as well as material use for each geographic region in the world. These regions are (a) domestic, (b) United States, (c) Canada, (d) Caribbean and Central America, (e) South America, (f) Europe, (g) Middle East and Asia and (h) Africa, Australia, and New Zealand. The entropy measure of diversification DivSales is defined as follows. Let Pi be the share of the ith geographic segment in the total sales of the firm. Then DivSalesi = N 1 1 Pi ln ( Pi ). This is a weighted average of the shares of the segments, the weight for each segment being the logarithm of the inverse of its share. The measure, which is used in the IO literature Figure S4.4. Cocaine Seizure in Colombia. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Source: The cocaine seizure information is from the Colombian Ministry of Defense. Net cocaine-cultivated land data are from the US Department of State. Note: The figure plots the moving average of annualized cocaine seizure (in kilograms) divided by net coca-cultivated land (hectares). (Palepu 1985; Rumelt 1982), gets larger the more segments a firm operates in, and the less the relative importance of each of the segments in the total sales. It takes the value zero for non-diversified firms. Similarly, a diversification measure of materials, DivMatsi , is calculated for each firm i. I then map this information with the plants in my sample using the four-digit industry classification. S4.4.1.1. The top four industries with the highest sales diversity measure, DivSales, are the following: (1) Motor vehicle parts manufacturing (2) Resin, synthetic rubber, and artificial synthetic fibers and filaments manufacturing (3) Basic chemical manufacturing (4) Nonferrous metal (except aluminum) production and processing S4.4.1.2. The bottom four industries with the lowest sales diversity measure, DivSales, are the following: (1) Other furniture related product manufacturing (mattresses and box springs) (2) Other food manufacturing (corn snacks, tortilla chips, peanuts, french fries,...) (3) Cement and concrete product manufacturing (4) Animal slaughtering and processing S4.4.1.3. The top four industries with the highest material diversity measure, DivMats, are the following: (1) Motor vehicle manufacturing (2) Electrical equipment manufacturing (3) Motor vehicle parts manufacturing (4) Semiconductor and other electronic component manufacturing S4.4.1.4. The bottom four industries with the lowest material diversity measure, DivMats, are the follow- ing: (1) Cement and concrete product manufacturing (2) Lime and gypsum product manufacturing (3) Sawmills and wood preservation (4) Bakeries and tortilla manufacturing S4.4.2. Construction of Trade Exposure Variables In constructing trade exposure variables at the metropolitan level, I use employment information from the Mexican Census 2004 (Censos Economicos 2004) and international trade data from the United States. Downloaded from https://academic.oup.com/wber/article/39/3/632/7754927 by World Bank Publications user on 04 August 2025 Censos Economicos 2004 provides employment information at municipality and industry level. Indus- try classification in the 2004 census is the Mexican version of NAICS (SCIAN). The United States and Mexican versions of NAICS are identical at the first four digits. Import information for the United States is obtained from the US Census (https://usatrade.census.gov/). The data include all goods that physically arrive into the United States, whether they are consumed domestically or are used further in production. The import value excludes transportation, insurance, freight, and other related charges incurred above the price paid. The data employ the North American Industry Classification System (NAICS) definitions for industries. To calculate the trade competition exposure variable for each metropolitan area, I first cal- culate the predicted change in Chinese imports in the United States in industry k between year 2005 and year 2010 for each four-digit NAICS industry. I divide this measure by the total non-agricultural number of employees in metropolitan area j to obtain the per-worker measure of the predicted change in Chinese imports in the United States. À la Bartik, I then use the ratio of employment of industry k in metropolitan area j in the census year 2004, Ek j0 , to the total initial Mexican employment for industry j, Ejo = k Ekj0 to map the change in the Chinese imports in the United States with the Mexican metropolitan areas. C The Author(s) 2024. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.