Trafficking Networks and the Mexican Drug War

Transcription

1 Trafficking Networks and the Mexican Drug War Melissa Dell September, 2011 Preliminary and incomplete. Abstract: Drug trade-related violence has escalated dramatically in Mexico during the past five years, claiming 40,000 lives and raising concerns about the capacity of the Mexican state to monopolize violence. This study examines how drug traffickers economic objectives influence the direct and spillover effects of Mexican policy towards the drug trade. By exploiting variation from close mayoral elections and a network model of drug trafficking, the study develops three sets of results. First, regression discontinuity estimates show that drug trade-related violence in a municipality increases substantially after the close election of a mayor from the conservative National Action Party (PAN), which has spearheaded the war on drug trafficking. This violence consists primarily of individuals involved in the drug trade killing each other. The empirical evidence suggests that the violence reflects rival traffickers attempts to wrest control of territories after crackdowns initiated by PAN mayors have challenged the incumbent criminals. Second, the study accurately predicts diversion of drug traffic following close PAN victories. It does this by estimating a model of optimal routes for trafficking drugs across the Mexican road network to the U.S. When drug traffic is diverted to other municipalities, drug trade-related violence in these municipalities increases. Moreover, female labor force participation and informal sector wages fall, corroborating qualitative evidence that traffickers extort informal sector producers. Finally, the study uses the trafficking model and estimated spillover effects to examine the allocation of law enforcement resources. Overall, the results demonstrate how traffickers economic objectives and constraints imposed by the routes network affect the policy outcomes of the Mexican Drug War. Keywords: Drug trafficking, networks, violence. I am grateful to Daron Acemoglu and Ben Olken for their extensive feedback on this project. I also thank Abhijit Banerjee, Dave Donaldson, Esther Duflo, Rachel Glennerster, Gordon Hanson, Austin Huang, Panle Jia, Chappell Lawson, Nick Ryan, and seminar participants at the Inter-American Development Bank, the Mexican Security in Comparative Perspective conference (Stanford), MIT, NEUDC, and UC San Diego for helpful comments and suggestions. Contact mdell@mit.edu.

2 1 Introduction Drug trade-related violence has escalated dramatically in Mexico during the past five years, claiming 40,000 lives and raising concerns about the capacity of the Mexican state to monopolize violence (Guerrero, 2011; Aguilar and Castaneda, 2009). Recent years have also witnessed large scale efforts by the Mexican government to combat drug trafficking. While drug traffickers are economic actors with clear profit maximization motives, there is little empirical evidence on how traffickers economic objectives have influenced the direct or spillover effects of Mexican policies towards the drug trade. This study uses variation from close mayoral elections and a network model of drug trafficking to examine these relationships. Mexico is the largest supplier to the U.S. illicit drug market (U.N. World Drug Report, 2011). While Mexican drug traffickers engage in a wide variety of illicit activities, the largest share of their revenues derives from trafficking drugs from Mexico to the U.S. (Guerrero, 2011). Official data described later in this paper document that in 2008, drug trafficking organizations maintained operations in two thirds of Mexico s municipalities and illicit drugs were cultivated in 14% of municipalities. Drug trade-related violence in Mexico has escalated dramatically during the past five years. Recent years have also witnessed a major crackdown on drug trafficking spearheaded by the conservative National Action Party (PAN). It has been controversial whether the state s policies have been responsible for the marked increased in violence, or whether violence would have risen substantially in any case as a result of the diversification of drug cartels into new criminal activities (Guerrero, 2011; Rios, 2011; Shirk, 2011). Given that transporting drugs to the U.S. is the primary economic activity of Mexican drug trafficking organizations, the study begins by specifying and estimating a network model of the drug trade. The model s central ingredient is the reasonable assumption that traffickers objective is to minimize the costs incurred in trafficking drugs from producing municipalities in Mexico across the Mexican road network to points of entry into the U.S. The model assigns a cost to each edge in the road network that is proportional to the physical length of the edge and the amount of drugs trafficked on that edge. Congestion parameters determining how the amount of drug trafficked on an edge affects the cost of traversing that edge are estimated using the simulated method of moments and cross-sectional data on illicit drug confiscations occurring at the beginning of the sample period. This study uses the trafficking model, together with close mayoral elections and data on drug trade-related outcomes between 2007 and 2009, to develop three sets of results. The first set of results uses a regression discontinuity (RD) approach to examine whether the close elections of mayors from the PAN - which has spearheaded Mexico s crackdown on the drug 1

3 trade - affects drug trade-related violence. It also explores some mechanisms that could lead violence to increase in response to crackdowns initiated by PAN mayors. Second, the study tests whether drug traffic is diverted following close PAN victories and examines whether the diversion of drug traffic is accompanied by violence and economic spillover effects. Finally, the study uses the trafficking model and estimated spillover effects to examine the allocation of law enforcement resources. These results are now discussed in more detail. The PAN has spearheaded the war on drug trafficking in Mexico, and evidence discussed in Section 2 suggests that PAN mayors have contributed to this crackdown. Prior to the inauguration of new authorities, municipalities where the PAN barely lost are statistically indistinguishable from places where they barely won along a large number of dimensions, suggesting the use of regression discontinuity to explore the impacts of local policy on the drug trade. RD estimates exploiting the outcomes of close elections show that drug traderelated violence in a municipality increases substantially after the close election of a PAN mayor. 1 Following the inauguration of PAN authorities, the probability that a drug traderelated homicide occurs in a municipality in a given month is around 13 percentage points higher than after a non-pan mayor takes office. This is a large effect, given that six percent of municipality-months in the sample experienced a drug trade-related homicide. Non-drug trade related homicides do not change following close PAN victories, indicating that the RD estimates do not simply reflect a reclassification of violence by PAN authorities. The violence consists primarily of individuals involved in the drug trade killing each other. Analysis using data on the industrial organization of drug trafficking suggests that the violence reflects rival traffickers attempts to wrest control of valuable territories after crackdowns initiated by PAN mayors have challenged the incumbent criminals. These results are consistent with qualitative and descriptive evidence suggesting that government crackdowns on drug trafficking in Mexico have caused massive increases in conflicts between traffickers (Guerrero, 2011; Shirk, 2011). They also relate to quantitative work by Angrist and Kugler (2008) documenting that exogenous increases in coca prices increase coca cultivation and violence in rural districts in Colombia. The evidence for Colombia suggests that when cocaine prices increase, combatant groups fight over the additional rents. In Mexico, crackdowns likely reduce rents in the short-to-medium run, but by weakening the incumbent criminal group they could also reduce the costs of taking control of a municipality. Controlling the municipality could offer substantial rents from trafficking and other criminal activities when the environment becomes more favorable in the future. The paper s second set of results explores whether close PAN victories exert spillover 1 See Lee, Moretti, and Butler (2004) for a pioneering example of the use of close elections to isolate exogenous variation in policy. 2

4 effects. When policy leads one location to become less conducive to illicit activities, organized crime may relocate elsewhere. For example, internationally financed coca eradication policies in Bolivia and Peru during the late-1990s led cultivation to shift to Colombia, and large-scale coca eradication in Colombia in the early 2000s has since led cultivation to re-expand in Peru and Bolivia, with South American coca cultivation remaining unchanged between 1999 and 2009 (Isacson, 2010; Leech, 2000; UN Office on Drugs and Crime ). 2 On a much more local level, work by Rafael Di Tella and Ernesto Schargrodsky (2004) documents that the allocation of police officers to Jewish institutions in Buenos Aires following a terrorist attack substantially reduced auto theft in the immediate vicinity of these institutions but may also have diverted some auto theft to as close as two blocks away. Accounting for spillovers is important in assessing the impacts of crime policies. It may also improve the allocation of law enforcement, a point which this paper explores empirically. To the best of my knowledge, this paper is the first to empirically estimate spillover patterns in drug trafficking activity. I begin by testing whether the network model, fitted using data from the beginning of the sample period, accurately predicts spillover patterns resulting from close PAN victories in later periods. Following a close PAN victory in a municipality, I assume that it becomes more costly to traffic drugs through the municipality. Thus, close PAN victories change the edge costs in the trafficking model, and for every period I use the model to predict optimal trafficking routes given that period s edge costs. This exercise produces a timevarying set of predicted routes that can be compared to monthly panel data on actual illicit drug confiscations, as illustrated in Figure 1. The model is predictive of changes in confiscations within municipalities over time, with the presence of a predicted drug trafficking route increasing the probability that illicit drugs are confiscated in a given municipalitymonth by around 1.5 percentage points, relative to an average probability of illicit drug confiscation of 5.3 percent. This relationship is highly statistically significant. Moreover, the model s predictions are robust to varying the form of the congestion costs (including ignoring congestion all together), to imposing a variety of costs for trafficking drugs through a municipality that has experienced a close PAN victory, to dropping municipalities from the sample that are located near municipalities that have experienced a close PAN victory, and to using different measures of illicit drug confiscations. Placebo checks also support the validity of the model. When drug traffic is diverted to other municipalities, the probability of drug trade-related 2 Another famous example is the highly successful Fascist crackdown on the Sicilian mafia in the 1920s, which led a number of mafiosi to establish and expand organized criminal operations abroad (Varese, 2011; Chubb, 1989). Once the Fascists were deposed and crime policy relaxed, intact mafiosi networks returned (Gambetta, 1993). 3

5 violence in these municipalities increases. Moreover, female labor force participation and informal sector wages fall, corroborating qualitative evidence that traffickers extort informal sector producers. Specifically, when a municipality acquires a predicted trafficking route, the probability that a drug trade-related homicide occurs in a given month increases by 1.5 percentage points, relative to a baseline probability of 4.5 percent. Wages earned by adult men in the informal sector fall by around 2.5 percent and female labor force participation declines by around one percentage point, relative to a baseline participation rate of 51 percent. Formal sector wages and male labor force participation are not affected. When traffickers re-route operations to new municipalities, they may become involved in conflicts with other traffickers using these routes or violence may result from their attempts to establish other criminal activities in the municipality. The economic spillover results are consistent with qualitative evidence, discussed in Section 2, that Mexican drug trafficking organizations extort informal sector producers via protection rackets. Because informal sector workers are often economically vulnerable, the impacts on welfare may be substantial. Finally, the study uses the trafficking model and estimated spillover effects to examine the allocation of law enforcement resources. The exercise provides an illustrative example of how the framework developed in this paper can enrich decisions regarding the allocation of scarce public resources. The next section provides an overview of Mexican drug trafficking and Mexican policies towards the drug trade, and Section 3 develops and estimates the network model of drug trafficking. Section 4 examines whether local politics influences drug trade-related violence. Section 5 documents that close PAN victories divert drug traffic elsewhere and uses the network model to explore whether there are resulting violence and economic spillover effects. Section 6 utilizes the trafficking model and estimated spillover effects to examine the allocation of law enforcement resources. Finally, Section 7 offers concluding remarks. 2 Background Mexican drug traffickers dominate the wholesale illicit drug market in the United States. According to the U.N. World Drug Report, Mexico is the largest supplier of heroin to U.S. markets and the largest foreign supplier of marijuana and methamphetamine. Fourteen percent of Mexico s municipalities regularly produce opium poppy seed (used to make heroin) or cannabis. Mexico also serves as a major transhipment hub for cocaine en route from Andean producing regions, with 60 to 90 percent of the cocaine consumed in the U.S. transiting through Mexico. The U.S. State Department estimates that wholesale earnings of Mexican 4

6 drug traffickers in U.S. markets range anywhere from $13.6 billion to $48.4 billion annually. 3 While this margin of error is extremely large, there is consensus that the U.S. market is a much greater source of revenue than Mexico s domestic illicit drug market, which is worth an estimated $560 million annually (Secretaría de Seguridad Pública, 2010). Data on drug addiction also emphasize the importance of the U.S. market. According to the U.S. National Survey on Drug Use and Health, 14.2 percent of Americans (35.5 million) have used illicit drugs during the past year, as contrasted to 1.4 percent of the Mexican Population (1.1 million) (Guerrero, 2011, p. 82; National Addiction Survey, 2008). While the main source of revenues for Mexican trafficking organizations is from the U.S. illicit drug market, these organizations are also diversified into a host of other illicit activities, including domestic drug sales, protection rackets, kidnapping, human smuggling, prostitution, oil and fuel theft, money laundering, weapons trafficking, and auto theft (Guerrero, 2011, p. 10). Drug trafficking organizations have substantially expanded their operations in some of these activities over the past several years (Rios, 2011). Most notably, extortion has increased substantially, with complaints to authorities tripling between 2004 and 2009 (Secretariado Ejecutivo del Sistema Nacional de Seguridad Pública, 2011). Mexico s national business association, Coparmex, estimates that extortion of small business owners in Mexico rose by more than 700 percent in Such illicit activities affect the lives of many citizens in no way affiliated with the drug trade, and as of 2011, public opinion surveys found that public security was more likely to be chosen as the largest problem facing the country than concerns about the economy. Due to Mexico s weak criminal justice system, it is estimated that only 25% of crimes are reported and only 2% are prosecuted (Shirk, 2011). At the beginning of this study s sample period, trafficking operations in Mexico were controlled by six major drug trafficking organizations. The number of major trafficking organizations had expanded to 16 by 2011 (Guerrero, 2011, p. 10). As documented in detail by Guerrero (2011, p ), alliances between these organizations have been highly unstable over the course of this period. Within cartels, most decisions about dayto-day trafficking operations are decentralized. This ensures that no single player will be able to reveal extensive information if he or she is captured by authorities, or if part of the cartel splinters into a separate organization. According to official Mexican data obtained from confidential sources, 68 percent of Mexico s 2,456 municipalities had a drug trafficking organization operating within their limits in early Historically, Mexico was dominated by a single party, the PRI (Instituionalized Revolutionary Party), and both local and federal authorities took a passive stance towards illicit 3 Estimates by U.S. Immigration and Customs Enforcement, the U.S. Drug Enforcement Agency, and the Mexican Secretaría de Seguridad Pública are broadly similar and also contain a large margin of error. 5

7 drug trafficking. There exist a number of well-documented instances in which state and federal officials engaged in drug trade-related corruption (see for example Shannon, 1988). While there were periodic crackdowns on drug trafficking by the Mexican federal government, these operations were limited in size and scope. 4 The PRI s dominance began to erode in the 1990s. The first opposition president was elected from the National Action Party (PAN) in 2000, and Mexico is now a competitive multi-party democracy. Soon after taking office in December 2006, PAN president Felipe Calderón deployed 6,500 federal troops to the state of Michoacan to combat drug trade-related violence. The government s efforts against drug trafficking have continued to increase since this time, with approximately 45,000 troops involved by Since the start of Calderón s presidency, violence has increased dramatically in Mexico. Over 40,000 people were killed by drug trade-related violence between 2006 and mid-2011, and drug trade-related violence in Mexico as a whole has increased by at least 30% every year during this period (Rios, 2011b). According to government estimates, 85% to 90% of the violence consists of people involved in the drug trade killing each other. It has been controversial whether the state s policies, by spurring conflicts between traffickers, have been responsible for the marked increased in violence, or whether violence would have risen substantially in any case as a result of the diversification of drug cartels into new criminal activities (see Guerrero, 2011; Rios, 2011; and Shirk, 2011 for discussions of both sides of this debate). This study presents causal evidence linking local politics to increases in drug trade-related violence. Local authorities command the majority of Mexico s law enforcement officers. In total, there are 2,139 independent state, local, and federal police agencies in Mexico, 2038 of which are municipal police agencies. 90% of Mexico s approximately 500,000 police officers are under the command of state and municipal authorities (Guerrero, 2011, p. 20). Mayors, who are elected every three years at different times in each of Mexico s 31 states and Federal District, name the municipal police and set policies regarding police conduct. Municipal police have a limited mandate, focusing on automobile traffic violations and minor disruptions to public order. It is rare for them to be involved in the confiscation of illicit drugs, and they do not have the training or weaponry typically required to make high level drug arrests. However, despite their limited enforcement capacity, municipal police are frequent targets of intimidation by drug traffickers and form the largest group of public servants killed by drug trade-related violence (Guerrero, 2011). Traffickers attempts to control local law enforce- 4 Notable examples include Operation Condor in the 1970s to eradicate illicit drug crops in northern Mexico and the deployment of federal troops to Nuevo Laredo, Tamaulipas by PAN president Vicente Fox in the early 2000s. 6

8 ment stems from the fact that they can serve as critical sources of local information for more sophisticated federal police attacks on drug trafficking operations. On the other hand, local police can also serve as valuable allies for traffickers. Organized criminals need information on who is passing through the municipalities they control, so that they can protect local criminal operations and anticipate attacks by their rivals, the military, and the federal or state police. Municipal police, who s main duty is to be out and about patrolling automobile traffic, are an ideal source for such information. 5 There is various anecdotal evidence that operations involving the federal police and military have been most effective when the relevant local authorities are politically aligned with the party controlling the federal government (Guerrero, 2011, p. 70). For example, while drug trade-related violence initially increased in Baja California in response to a large-scale intervention by the federal police, the violence has since come back down, and the state is frequently showcased as a success story of federal intervention. The governor of Baja California belongs to the PAN, which is the party controlling Mexico s executive branch, and the federal intervention began under the auspices of a PAN mayor in Tijuana who was enthusiastic to cooperate with federal authorities. On the other hand, in Ciudad Juarez both the mayor and governor belong to the opposing PRI party, conflicts and mistrust between municipal and federal police have been rampant, and violence continues to reach record levels. In addition to the evidence regarding federal-local coordination, there is also anecdotal evidence that PAN mayors and governors are more likely to request assistance from the federal police. 6 There are several potential explanations for this anecdotal evidence. Authorities from the same party may be able to more effectively coordinate law enforcement operations, PAN authorities could be less corrupted by the drug trade, or it could be that the preferences of PAN authorities lead them to take a tougher stance on organized crime. Disentangling local authorities underlying motivations is very challenging given the paucity of relevant data and inherent empirical difficulties in separately identifying a large number of potential explanations. 7 The PAN s role in spearheading the war on drug trafficking - combined with 5 For example, in a recent meeting on national security, Mexican president Felipe Calderón argued: The military report that when they enter a city, they tune into the frequency of the municipal police radio and hear them reporting to the criminals every step they [the military] take. And right now they are on this avenue arriving at the traffic light on that corner, and they have six trucks and bring this many weapons. And these municipal police patrols attempt to block their [the military s] access (El Pais, August 26, 2010, translation mine. 6 Unfortunately, disaggregated data on federal police assignments and requests by mayors for federal police assistance are not made available to researchers. 7 I have analyzed official government data on corruption made available by confidential sources. This data records drug trade-related corruption of mayors in 2008, as measured primarily by intercepted calls from traffickers to political officials. While the data are likely quite noisy, to my knowledge they are the best source of information on drug trade-related corruption available. Corruption was no more common in municipalities where a PAN candidate had been elected mayor by a narrow margin than in municipalities 7

9 the various plausible reasons why the party affiliation of local politicians may influence the drug trafficking environment - motivates this study s focus on close elections involving the PAN. I now turn to an examination of the extent to which the outcomes of these elections influence the municipal drug trafficking environment. 3 A Network Model of Drug Trafficking This section develops and estimates a model of the network structure of drug trafficking, which will be used in the analysis of the direct and spillover effects of local policy towards the drug trade. In this model, traffickers minimize costs from trafficking drugs from origin municipalities in Mexico across the Mexican road network to U.S. points of entry. They incur costs from the physical distance that must be traversed as well as from congestion resulting from other traffickers using similar routes. The congestion costs are unknown, and thus they are estimated using simulated method of moments and cross-sectional data on confiscations from the beginning of the sample period. Every choice of the model s congestion parameters generates a set of moments that summarize model-predicted routes, and I estimate the congestion parameters by matching these moments to their counterparts generated from data on actual illicit drug confiscations. Subsequently, I test whether the model fitted on this pre-period cross-section is predictive of how illicit drug confiscations change within municipalities later in the sample period, in response to close PAN victories occurring elsewhere in Mexico. Close PAN victories increase the costs of trafficking drugs through a municipality by a pre-specified amount, and taken together these close elections generate month-to-month variation in predicted trafficking routes throughout Mexico. If the model is predictive of how PAN victories divert drug traffic and if the outcomes of close elections are effectively random (this is explored in Section 4), the variation in these routes can be used to locate violence and economic spillover effects. The model will also be used later in the paper to explore why crackdowns by the Mexican state lead violence to increase and to examine the allocation of law enforcement resources. Section 3.1 specifies the drug trafficking environment, traffickers objective function, and the equilibrium concept. Section 3.2 estimates the model s parameters, using official municipal-level data on the value of illicit drug confiscations, which were obtained from confidential sources. The model s applications to the direct and spillover effects of local policy will be developed in Sections 4 and 5, respectively. where the PAN candidate had lost by a narrow margin. 8

10 3.1 Model Specification Let N = (V, E) be an undirected graph representing the Mexican road network, which consists of sets V of vertices and E of edges. 8 Traffickers transport drugs across the Mexican road network from a set of origin municipalities, O to a set of destination municipalities, D. Each origin i has a location in latitude-longitude space, produces a given supply of drugs, S i, and has a trafficker who decides how to transport the municipality s supply of drugs to U.S. points of entry. I model trafficking decisions as made by local traffickers because, as discussed in Section 2, trafficking decisions are typically decentralized so that no single player will be able to reveal extensive information if he or she is captured or defects. The baseline model focuses on domestically produced drugs, and origins are identified from confidential Mexican government data on drug cultivation (heroin and marijuana) and drug labs (meth). Opium poppy seed and marijuana have a long history of production in given regions with particularly suitable conditions, and thus we can be confident that the origins for domestically produced drugs are stable and accurate throughout the sample period. In contrast, cocaine typically enters Mexico via fishing vessels and go-fast boats that transport bulk shipments of cocaine from Colombia to Mexico (U.S. Drug Enforcement Agency, 2011). Thus, the origins for cocaine routes are more flexible, less well-known, and may have changed substantially during the sample period. Moreover, government policies may divert cocaine traffic away from Mexico altogether. 9 For these reasons, the model focuses on domestically-produced drugs. In practice, we know little about the quantity of drugs cultivated in each producing municipality, and hence I make the simplifying assumption that each municipality produces a single unit of drugs. Destinations consist of points of entry into the U.S. via terrestrial border crossings or major Mexican ports. While drugs may also enter the U.S. between terrestrial border crossings, the large amount of legitimate commerce between Mexico and the U.S. offers ample opportunities for drug traffickers to smuggle large quantities of drugs through border crossings and ports (U.S. Drug Enforcement Agency, 2011). 10 Each of the destinations has a given size, approximated by the number of commercial lanes for terrestrial border crossings and the container capacity for ports. Moreover, all destinations pay the same international price for a unit of smuggled drugs. 8 There are 17,453 edges in the Mexican road network. 9 There is some evidence that shipments of cocaine through Haiti have increased in recent years (U.S. Drug Enforcement Agency, 2011). 10 There are 370 million entries into the U.S. through terrestrial border crossings each year, and 116 million vehicles cross the land borders with Canada and Mexico (U.S. Drug Enforcement Agency, 2011). More than 90,000 merchant and passenger ships dock at U.S. ports each year, and these ships carry more than 9 million shipping containers. Commerce between the U.S. and Mexico exceeds a billion dollars a day. 9

11 Trafficking paths connect producing municipalities to U.S. points of entry. Let P i denote the set of all possible paths between producing municipality i and the United States, let P = i P i denote the set of all paths between all producing municipalities and the U.S., and let x p denote the flow on path p P. Each edge e E has a cost function c e (l e, x e ), where l e is the length of the edge in kilometers and x e = p P e p x p is the total flow on edge e, which equals the sum of flows across the paths that traverse it. The cost function specifies how much a trafficker must pay to traverse the edge. The congestion costs are unknown, and thus the next subsection estimates them using the method of simulated moments. Each drug trafficker s objective is to minimize the costs of transporting drugs to the U.S., taking aggregate flows as given. Since the amount of drugs transported by a single agent is small relative to the total supply of drugs, the assumption that traffickers take aggregates as given appears reasonable and simplifies the analysis considerably. In equilibrium, the costs of all routes actually used to transport drugs from a given producing municipality to the U.S. are equal and less than those that would be experienced by reallocating traffic to an unused route. This intuitive equilibrium condition for congestion games was first published by English transportation analyst John Wardrop in 1952, and is the Nash equilibrium of this game. 11 Beckmann, McGuire and Winsten (1956) proved that x WE is a Wardrop equilibrium if and only if it is a solution to: 12 min e E xe 0 c e (z)dz (1) s.t. x p = x e e E (2) p P e p x p = S i i = 1, 2,..., p P (3) p P i x p 0 p P (4) where recall that x e is the total flow on edge e, x p is total flows on path p, c e ( ) is the cost to traverse edge e, and S i is the supply of drugs at origin i. The first constraint requires that the flow of traffic on paths traversing an edge sum to the total flow of traffic on that edge, the second constraint requires that supply be conserved, and the third constraint requires 11 Formally, an equilibrium must satisfy the following conditions: 1. For all p, p P i with x p, x p > 0, e p c e(x e ) = e p c e(x e ). 2. For all p, p P i with x p > 0 x p = 0, e p c e(x e ) e p c e(x e ). 12 A version of this proof is repeated in the appendix. 10

12 flows to be non-negative. By Weierstrass s Theorem, a solution to the above problem exists, and thus a traffic equilibrium always exists. Moreover, if each c e is strictly increasing, then the equilibrium is unique. It can be seen by differentiating the objective than an increase in flow on edge e of δ e would decrease the trafficker s objective by the additional congestion costs that he must now pay to traverse that edge. Traffickers ignore the externalities that their use of a route imposes via congestion costs. This differs from the the social planner, who would minimize total trafficking costs, and thus the equilibrium routing pattern will typically not be socially optimal. Moreover, it is possible that removing an edge could decrease the total costs paid by traffickers. This well-known result, referred to as Braess s paradox after mathematician Dietrich Braess, would never obtain in the social optimum but has been documented in a number of real world traffic congestion examples. 13 While the above optimization problem does not have a closed-form solution, for a given network, set of supplies, and specification of the congestion costs c e ( ) it can be solved using numerical methods. I use the Frank-Wolfe algorithm (1956), which generalizes Dantzig s well-known simplex algorithm for linear programming to non-linear programming problems. Details about this algorithm and its implementation are provided in Appendix A, which describes the paper s estimation procedures. 3.2 Estimating congestion costs Solving the trafficking problem requires specifying the form of the edge cost function and estimating its unknown parameters. Recall from the description of the trafficking environment that the cost of each edge, c e (l e, x e ), depends upon its physical length in kilometers and the amount of drug traffic traversing that edge. In order to explore robustness to specifying the congestion costs with different functional forms, I estimate several versions of the model. In the most parsimonious version, border crossings impose a congestion cost equal to φ t (flow e /lanes) δ for terrestrial border crossings and φ p (flow e /containers) δ for ports, where {φ t, φ p, δ} are congestion parameters. δ captures the fact that as illicit traffic through a given crossing increases, the quality of hiding places may decline or the authorities may direct more (or less) attention towards the crossing per unit of traffic. Congestion values are converted to the same units as physical distance costs by the parameters {φ t, φ p }. One might be concerned that this functional form is overly restrictive. While there is not enough variation in the data to estimate a separate congestion parameter for each of the 26 points of entry into the U.S., I do estimate a version of this model with six φ parameters: one for 13 For examples of Braess s paradox in traffic congestion in Seoul, New York, Berlin, Boston, and London, see Youn, Gastner, and Jeong (2008); Easley and Kleinberg (2008, p. 71); and Knodel (1969, p ). 11

13 terrestrial points of entry in the bottom quartile of the size distribution (i.e. crossings with a single lane), three more for terrestrial points of entry in the other three quartiles (2 lanes, 3 to 9 lanes, and 10 to 17 lanes, respectively), one for ports with below median container capacity (less than 160,000 TEUs, which is divided by 10,000 to be in units comparable to the size of the terrestrial crossings), and one for ports with above median container capacity. 14 This allows the model to more flexibly capture the relationship between congestion costs and the size of the U.S. points of entry. In the final version of the model, I estimate the seven congestion cost parameters for U.S. points of entry, as well as congestion costs on the interior edges. The congestion costs on the interior edges take the form: dist e φ int flow γ e, where dist e is the length of the interior edge, and φ int and γ are additional congestion parameters. The congestion parameters can be estimated using simulated method of moments (SMM) and a cross-section of data on the value of illicit drug confiscations during the beginning of the sample period, which lasts until the first authorities elected during the sample period take office in July For every choice of the model s parameters, this generates a set of moments that summarize the patterns of model-predicted confiscations. I estimate the model s parameters by matching these moments to their counterparts generated from data on the value of actual illicit drug confiscations. Formally, let {x m } denote the flows predicted by the trafficking equilibrium problem, aggregated to the municipal level, and let θ 0 R P denote the vector of congestion parameters plus one scaling parameter κ that maps predicted flows to predicted confiscations: conf ˆ m = κx m, κ (0, ). 15 Let g(x m, θ 0 ) R L denote a vector of moment functions that specifies the difference between observed confiscations and those predicted by the model, given the congestion costs described by θ 0. The number of moment conditions L must be greater than or equal to the number of parameters for the model to be identified. The SMM estimator ˆθ minimizes the following weighted quadratic form: θ = argmin θ Θ [ M ] [ 1 M ] ĝ(x m, θ) Σ M ĝ(x m, θ) M m=1 m=1 where ĝ( ) is an estimate of the true moment function. M is the total number of municipalities in the sample and Σ M is an L x L positive semi-definite weighting matrix. Let ˆΛ denote a consistent estimator of Var[g(X m, θ 0 )]. Then, if Σ m = ˆΛ 1, the SMM estimator will be asymptotically efficient. The optimal SMM estimator can be obtained by first minimizing 14 TEU stands for twenty-foot equivalent units. An equivalent unit is a measure of containerized cargo capacity equal to one standard 20 ft (length) x 8 ft (width) container. 15 κ almost certainly differs depending on the local environment. However, it is not possible to estimate this dependence. To the extent that the model is robustly predictive of the evolution of illicit drug confiscations and other policy spillovers in later periods, this suggests that the framework is useful despite the fact that confiscations are an imperfect measure of actual illicit drug flows. (5) 12

14 (5) using Σ M = I L, the identity matrix. From this first step, ˆΛ can be calculated, and (5) can be re-optimized in a second step using Σ M = ˆΛ 1. Further details about the asymptotic properties of the SMM estimator can be found in Pollard (1989) and McFadden (1989). Predicted confiscations on a given edge of the network are not independent of predicted confiscations elsewhere in the network, introducing potentially complicated spatial dependence. To address this concern, following Conley (1999) I replace the asymptotic covariance matrix Λ with a weighted average of spatial autocovariance terms with zero weights for observations farther than a certain distance. I allow for correlation between municipalities located within 250 km of each other. More details about inference are given in Appendix A. Depending on the form of the congestion costs being estimated, there are four to ten parameters in θ. The moments match the mean model predicted and observed confiscations at ports, at terrestrial bordering crossings, and on interior links. They also match the interaction between port confiscations and container capacity, between terrestrial border crossing confiscations and commercial lanes, between interior confiscations and the length of the interior edge, and between interior confiscations and the length of the detour required to circumvent the edge. For models that estimate six separate crossing congestion parameters, the moment conditions match predicted confiscations within a 100 km radius of each U.S. point of entry to actual confiscations within a 100 km radius. The estimates for the model with three congestion parameters are similar when these moments are included in the estimation (see the appendix), but I omit them in estimating the results presented here because they are not necessary and raise concerns akin to those that arise when many instruments are used. Finally, the moment conditions match the model predicted and observed variance of confiscations across U.S. points of entry and across interior edges. As is often the case with choice problems, the SMM objective function is not globally convex and hence minimizing it is non-trivial. Standard gradient methods such as Newton- Rhapson and the Davidon, Fletcher, and Powell (DFP) algorithm may perform poorly, and thus I instead use simulated annealing, which is more suited to problems that lack a globally convex objective (Kirkpatrick, Gelatt, and Vecchi, 1983). Details about the estimation procedure are given in Appendix A. Table 1 reports the simulated method of moments estimates of the cost function parameters. Conley standard errors are in brackets and robust standard errors are in parentheses. Column 1 reports estimates for the specification with parsimonious congestion costs on U.S. points of entry, column 2 reports estimates for the specification with more flexible congestion costs on U.S. points of entry, and column 3 reports estimates for the specification with congestion costs on both U.S. points of entry and interior edges of the road network. All parameters are precisely estimated. δ, which captures the shape of the congestion 13

15 costs on U.S. points of entry, ranges from 1.49 to 1.88, depending on the model specification. This implies that the costs of congestion increase as illicit drug traffic through points of entry to the U.S. rises. This is an intuitive finding given that hiding places may become worse and authorities may direct more attention towards a given point of entry as illicit flows through that point of entry increase. When the congestion costs on points of entry are allowed to take a more flexible form in column 2, the φ Q i t s increase with the size of the border crossing. In other words, the model estimates higher congestion costs on a ten lane border crossing edge with ten units of traffic than on a one lane edge with one unit of traffic. The total costs imposed by congestion are slightly less than the total costs imposed by distance. The model in column 3, which includes congestion costs on interior edges, estimates γ = This implies that congestion costs on interior edges are nearly linear. In equilibrium, the total costs imposed by congestion on U.S. points of entry in this specification are about fifteen times larger than the total costs imposed by congestion on interior edges, suggesting that congestion at U.S. points of entry is substantially more important than congestion within Mexico. This is not surprising, given the greater law enforcement presence at U.S. points of entry and the bottlenecks that they impose. All three models estimate that total congestion costs are nearly as large as total distance costs. However, as will be documented in detail in Section 5, the inclusion of congestion costs in the model does not cause a large reallocation of drug traffic relative to a world without congestion. The model parameters are estimated to match the cross-section of confiscation values, and the model is highly predictive. In a linear regression of actual municipal-level confiscations on model predicted confiscations, all three specifications have a t-statistic of between 7 and 8. The squared correlation coefficient is between 0.02 and As would be expected, as more parameters are added to the model, the correlation coefficient increases somewhat. The model is also highly predictive of the intensive margin of confiscations. When comparing the actual confiscations indicator variable to the predicted confiscations indicator variable, the Pearson chi-squared statistic has a value of 100 for the 4 parameter model, 116 for the 8 parameter model, and 76 for the 10 parameter model, all of which have p-values near zero. Graphical analyses of model fit are presented in the appendix. Section 5 will provide a more rigorous test of whether the model is predictive. Specifically, it uses data from a later period to test whether the model can predict how illicit drug confiscations within municipalities change over time in response to close PAN victories elsewhere in Mexico. 14

16 4 Local Politics and the Drug Trade This section uses a regression discontinuity approach to test whether close mayoral elections involving the National Action Party (PAN) - which has spearheaded the war on drug trafficking - influences drug trade-related violence. While data on mayoral requests for federal police and the allocation of federal police are not made available to researchers, the qualitative evidence discussed in Section 2 suggests that PAN mayors may be more likely to crack down on drug trafficking by enlisting the assistance of federal law enforcement authorities. Consistent with this evidence, the RD estimates show that drug trade-related violence in a municipality increases substantially after the close election of a PAN mayor. This violence consists primarily of individuals involved in the drug trade killing each other, and quantitative evidence presented in this section suggests that it reflects rival traffickers attempts to wrest control of territories after crackdowns initiated by PAN mayors have challenged the incumbent criminals. The advantages of focusing on close elections is twofold. First, they offer a plausibly exogenous source of variation for examining the effect of government policy on drug traderelated violence. Municipalities controlled by different political parties are in general quite distinct, making it difficult to disentangle the impact of politics from effects of other factors correlated with politics. In contrast, this section documents that prior to close elections municipalities where the PAN barely lost are statistically indistinguishable from municipalities where they barely won along a large number of dimensions. To the extent that the close election of PAN mayors incentivize traffickers to re-route some of their operations, they could also provide a source of variation for identifying spillover effects. Correlations between local policy in one municipality and drug trade-related outcomes elsewhere could occur for several reasons, as highlighted more generally by Manski s formal treatment of spillover effects (1993). First, correlations could result from environmental factors unrelated to the local policies under consideration. Second, they could occur because traffickers choose to operate in a given geographic arrangement for reasons unrelated to policy. Finally, local drug trafficking policies in one municipality could affect drug traffic and other outcomes elsewhere. The close PAN victories examined in this section will be used to identify spillover effects in Section Data This section uses official government data on drug trade-related outcomes obtained from confidential sources. The data, unless otherwise noted, are available and classified as complete for December of when President Felipe Calderón took office - through December of 15

17 Data on drug trade related homicides and armed confrontations between authorities and organized criminals are compiled by a committee with representatives from all ministries who are members of the National Council of Public Security (Consejo Nacional de Seguridad Pública) that meets each week to classify which homicides from the past week are drug trade-related. 17 Drug trade-related homicides are defined as any instance in which a civilian kills another civilian, with at least one of the parties involved in the drug trade. The classification is made using information in the police reports and validated whenever possible using newspapers. The committee also maintains a database of how many people have been killed in armed clashes between police and organized criminals. Additionally, confidential daily data for 2007 and 2008 on all homicides in Mexico were obtained from the National Institute of Statistics and Geography (INEGI). These data do not distinguish between drug trade and non-drug trade-related homicides. This section also uses municipal level data on the presence of drug trade-related organizations. These include local gangs as well as larger trafficking organizations. The data list which of Mexico s 2456 municipalities had at least one drug trade-related organization (DTO) operating within their limits in early They provide the closest possible approximation to pre-period DTO presence available, given that systematic data about DTOs not collected before this time. Data on the identity of the DTO controlling a municipality, when known, are also available for early Electoral data for elections occurring during were obtained from the electoral authorities in each of Mexico s states. The sources for a number of other variables, used to examine whether the RD sample is balanced, are listed in the notes to Table Econometric framework and graphical analysis This section uses the following regression specification, which combines a regression continuity and differences-in-differences approach, to test whether the outcomes of close PAN elections impact drug trade-related violence: y mst = β 0 + T ms τ= T ms β τ ζ τm + T ms τ= T ms γ τ ζ τm P ANwin ms + f(spread ms )P ost mst + ψ st + δ m + ɛ mst (6) 16 Systematic quantitative data on most drug trade-related outcomes are unavailable prior to the Calderón administration. 17 Previously reported homicides are also considered for reclassification if new information has become available. 16

18 where y mst is the violence outcome of interest and {ζ τ } is a set of months-to-election and months-since-inauguration dummies. P ANwin ms is a dummy equal to 1 if the PAN won the election, and P ost mst is a dummy equal to 1 for all periods t in which the new municipal authorities have assumed power. f( ) is the RD polynomial, spread ms is the margin of PAN victory, ψ st are state x month fixed effects, and δ m are municipality fixed effects. This specification allows us to examine the data for pre-trends as well as time patterns in the post-inauguration x PAN win effects. For the graphical analysis that is presented in this section, the sample is a balanced panel that is restricted to elections where the PAN won or came in second, with a vote spread of five percentage points or less between the winner and runner-up. Months falling between the election and inauguration of new authorities - a period lasting anywhere between two to five months - are not included in the sample. This lame duck period is difficult to examine transparently using individual month dummies due to its varying length by state. However, it is fully explored in tables presented in the next subsection, which document that there are not statistically significant differences in violence during the lame duck period in municipalities where the PAN barely won as compared to where they barely lost. Identification requires that all relevant factors besides treatment vary smoothly at the threshold between a PAN victory and a PAN loss. That is, letting h 1 and h 0 denote potential outcomes under a PAN victory and PAN loss, respectively, and v denote the PAN margin of victory, identification requires that E[h 1 v] and E[h 0 v] are continuous at the PAN win-loss threshold. This assumption is needed for municipalities where the PAN barely won to be an appropriate counterfactual for municipalities where the PAN barely lost. To assess the plausibility of this assumption, Table 2 compares a number of characteristics in municipalities where the PAN barely lost to those in municipalities where they barely won. Crime characteristics include the average monthly drug-trade related homicide rate for December of 2006 (when these data were first collected) through June of 2007 (when the first election used in this paper took place) and the average probability that a drug trade-related homicide occurred in a given month during this period. They also include police-criminal confrontation deaths per 100,000 inhabitants (Dec Jun. 2007), the average probability that police criminal-confrontation deaths occurred in a given municipality month during this period, and the long-run average municipal homicide rate ( ). Political characteristics explored are municipal tax collection per capita (2005), municipal taxes per $ income (2005), dummies for the party of the mayorship incumbent, number of alternations of the party controlling the mayorship ( ), and a dummy equal to 1 if the PRI (the historically dominant party) never lost the mayorship during the 1976 to 2006 period. Demographic characteristics examined are population (2005), population density 17

19 (2005), and migrants per capita (2005). Economic characteristics include income per capita (2005), the municipal Gini index (2005), migrants per capita (2005), malnutrition (2005), mean years of schooling (2005), infant mortality (2005), percent of households without access to sewage (2005), percent of households without access to water (2005), and the municipal marginality index (2005). The marginality index incorporates information on literacy; primary school completion rates; access to electricity, sewage, and running water; household overcrowding; construction materials used in households; municipal population in rural areas; and household income. Road network characteristics are the average detour length required for this paper s predicted drug routes to circumvent the municipality, total length of roads in the municipality (2005), road density (km/km 2 ), and distance of the municipality to the U.S. border. Finally, the geographic characteristics explored are average municipal elevation, slope, surface area, low temperature ( ), high temperature ( ), and precipitation ( ). Sources for these variables are listed in the notes to Table 2. Column (1) of Table 2 gives the mean value for each variable in municipalities where the PAN barely lost, column (2) does the same for municipalities where the PAN barley won, and column (3) reports the t-statistics on the difference in means. The sample is limited to elections with a vote spread between the winner and the runner-up of five percentage points or less. In no case are there statistically significant differences between municipalities where the PAN lost and municipalities where they won, strongly supporting the assertion that place where the PAN barely lost provide an effective control group for municipalities where they barely won. Results are similar if the vote spread is limited to three or seven percentage points instead. Identification also requires the absence of selective sorting around the PAN win-loss threshold. This assumption would be violated, for example, if elections could be rigged and PAN candidates in municipalities with a different drug trafficking trajectory were able to successfully rig the results in their favor. While there are many examples of historically rigged elections in Mexico, the political system has become dramatically more competitive and open since the early 1990s and genuine legal recourse exists in the event of concerns about the fairness of elections. The balancing of the sample on the crime pre-characteristics, the electoral variables, and the many other characteristics explored in Table 2 strongly suggests the improbability of rigged elections driving the results. Finally, identification requires the absence of differential pre-trends in municipalities where the PAN barely won, as compared to those where they barely lost. In order to assess the plausibility of this assumption, I now present a graphical analysis of equation (6). Figure 2 plots the γ τ coefficients from equation (6) against time, measured relative to the election and inauguration of new authorities. The dashed lines show 95% confidence 18

20 intervals. In Panel A, the dependent variable is the probability that a drug trade-related homicide occurs in a given municipality-month, and in Panel B it is the drug trade-related homicide rate per 10,000 municipal inhabitants. 18 Drug trade-related homicides are defined as any instance in which a civilian kills another civilian, with at least one of the parties involved in the drug trade. They exclude deaths in violent clashes between police or military and drug traffickers. 85% of drug trade-related homicides involve drug traffickers killing each other. A quadratic RD polynomial is used, and results (available upon request) are highly robust to using alternative functional forms. Panels A and B of Figure 2 document the absence of statistically significant differences in the prevalence of drug trade-related homicides between municipalities where the PAN barely won versus where they barely lost, before the close elections took place. This supports the validity of the identification strategy. Following the inauguration of new authorities, the month x PAN win coefficients become large, positive, and statistically significant. The estimates in Panel A document that the probability that at least one drug trade-related homicide occurs in a municipality in a given month is around 13 percentage points higher after a PAN mayor takes office than after a non-pan mayor takes office. 19 This is a large effect, given that six percent of municipality-months in the sample examined experienced a drug trade-related homicide. trade-related homicide rate is also large. 20 The estimated effect of close PAN victories on the drug One potential concern is that municipalities with PAN mayors could have a greater propensity to describe homicides as drug trade-related. While homicides are classified by a national committee, it is still possible that the information in the police reports used to make this classification could differ depending on the party of the mayor. Hence, Panel C of Figure 2 explores the probability that a non-drug trade-related homicide occurs in a given municipality-month, and Panel D examines the non-drug trade-related homicide rate. There are no statistically significant differences in non-drug trade-related homicides between places where the PAN barely won and those where they barely lost, either before or after the 18 While it is not obvious that drug trade-related homicides should be normalized by municipal population, as drug trafficking activity is not always proportional to population, I report this outcome because this is how homicides are typically reported. 19 When the post-period is extended to a year following the inauguration of new authorities, for the first five months the coefficient on the PAN win x month dummies are similar to those in Figure 3 (results available upon request). For the period between six months and a year after the inauguration of new authorities, the coefficients become more volatile. Whether this is due to PAN authorities successfully deterring drug trafficking activity or results from these authorities becoming less tough on crime is not possible to establish in the RD context, but results presented in the next section suggest that drug traffic continues to be diverted to other municipalities beyond the first six months that a PAN mayor has been in office. 20 Plots using the counts of these variables look similar to those exploring the rates per 10,000 inhabitants and are omitted to save space. 19

21 inauguration of new authorities. This alleviates concerns that close elections simply affect the classification of violence. Next, Figure 3 explores the relationships between close PAN victories and violence from a cross-sectional RD perspective. The y axis plots the average outcome either during the post-inauguration period (Panels A and C) or pre-election period (Panels B and D). The x-axis plots the PAN margin of victory, with a negative value representing a PAN loss. The solid black line shows a local linear regression of the relevant violence outcome on the PAN margin of victory, with the bandwidth chosen according to the Imbens-Kalyanaraman bandwidth selection rule (Imbens and Kalyanaraman, 2009). The dashed lines plot 95% confidence intervals. Panel A shows that in the post-inauguration period, there is a jump upward in drug traderelated homicides at the threshold between a PAN loss and a PAN victory. In contrast, Panel B documents that during the pre-election period there is no discontinuity at the threshold between places that later would experience a PAN loss versus a PAN victory. Panels C and D show the absence of a discontinuity, during both the post-inauguration and pre-election periods, in the non drug-trade related homicide rate. The same is true when a dummy measure is used (see appendix). These findings are highly robust to using a variety of different polynomials - instead of local linear regression - to specify the RD. Overall, this evidence in Figure 3 is highly consistent with that presented in Figure Further results and robustness The graphical analysis indicates that drug trade-related violence in a municipality increases substantially after the close election of a PAN mayor. This section examines the robustness of this result. In order to do so parsimoniously, it uses the following specification, which includes a dummy for the lame duck period and a dummy for the post-inauguration period, rather than the individual month dummies of equation (6): y mst = β 0 + β E LameDuck mst + β I P ostinnaug mst + γ E LameDuck mst P ANwin ms + γ I P ostinnaug mst P ANwin ms + f(spread ms )LameDuck mst + f(spread ms )P ostinnaug mst + ψ st + δ m + ɛ mst (7) LameDuck mst is a dummy equal to 1 for all periods t between the election and inauguration of new authorities, P ostinnaug mst is a dummy equal to 1 for all periods t in which the new municipal authorities have assumed power, and all other variables are defined 20

22 as in equation (6). Pre-election is the omitted category. The unit of observation is the municipality-month. Columns (1) through (3) estimate equation (7), limiting the sample to a seven percentage point, five percentage point, and three percentage point vote spread, respectively, and using a quadratic RD polynomial. Panel A examines drug trade-related homicides, and Panel B examines non drug-trade related homicides. The probability that at least one drug traderelated homicide occurs in a municipality in a given month is around 13 percentage points higher after a PAN mayor takes office than after a non-pan mayor takes office. This is of similar magnitude to the post-inauguration coefficients plotted in Figure 2, and can be compared to the sample average probability of six percent that a drug trade-related homicide occurs in a given month. Estimates increase somewhat in magnitude as the sample is restricted to elections with narrower vote spreads. In contrast, the coefficients on lame duck x PAN win are statistically insignificant and tend to be substantially smaller than the post-inauguration x PAN win coefficients, indicating that the violence effects of close PAN victories do not occur until the inauguration of new authorities. Moreover, Panel B documents that PAN victories do not exert statistically significant impacts on non drug-trade related homicides in either the lame duck or post-inauguration periods, and the coefficients in Panel B tend to be small compared to those in Panel A. Columns (4) through (8) further examine the robustness of the estimated relationship between close PAN victories and drug-trade related violence. The dependent variable in column (4) is the number of drug trade-related homicides (non drug trade-related homicides) in a given municipality-month, and the dependent variable in column (5) is monthly drug trade-related homicides (non drug trade-related homicides) per 10,000 municipal residents. All specifications estimate that the inauguration of PAN authorities significantly increases drug trade-related homicides relative to the inauguration of non-pan authorities. The effects on non-drug trade-related homicides remain small and statistically insignificant. Moreover, columns (6) through (8) show that the estimates are similar regardless of the order of the RD polynomial. The appendix documents that results are also robust to allowing the slope of the RD polynomial to differ on either side of the PAN win-loss threshold. Finally, columns (9) through (11) estimate the effects of PAN victories on homicides in the post-inauguration period, lame duck period, and pre-election period, respectively, using local linear regression with the bandwidth chosen according to the Imbens-Kalyanaraman bandwidth selection rule. In other words, these columns estimate the magnitude of the crosssectional discontinuities that are plotted in Figure 3. The dependent variable is the average probability of homicides during the post-inauguration period (column 9), the lame duck period (column 10), and the pre-election period (column 11). There is a statistically significant 21

23 effect of PAN victories on drug trade-related homicides only in the post-inauguration period. The estimated effect of 0.17 (s.e.=0.1) is similar in magnitude to the coefficients estimated using the full panel. While the cross-sectional RD approach has less power than the panel approach, the coefficient on drug trade-related homicides is statistically significant at the 10% level. In contrast, close PAN victories do not affect drug trade-related homicides in the lame duck period or pre-election period, and they do not affect non-drug trade-related homicides in any period. These findings are highly robust to using a variety of different polynomials - instead of local linear regression - to specify the cross-sectional RD. Another outcome of interest is deaths in police-drug trade-related confrontations. Deadly conflicts between police and drug traffickers are relatively rare, occurring in only 20 municipality-months and 12 municipalities across the sample period. Half of the deadly confrontations between drug trade criminals and authorities occurred during the postinauguration period, with 60 percent of these occurring in municipalities where the PAN barely won. Due to the rarity of these events, it is not possible to conduct a rigorous econometric analysis of the links between deadly police-criminal confrontations and the outcomes of close elections. Table 3 provides strong evidence that close PAN victories increase drug trade-related violence. One potential concern, however, is that it is not the close election of a PAN mayor but rather some other characteristic of local politics that influences drug trade-related violence. Hence, Table 4 examines how different aspects of local politics relate to violence. The dependent variable in all columns is a dummy equal to one if a drug trade-related homicide occurred in a given municipality-month and equal to zero otherwise. All columns utilize the panel data, the sample is limited to elections with a vote spread of five percentage points or less, and the quadratic RD polynomial - which produces the most conservative results - is used. The lame duck x PAN win effects are not reported due to space constraints. Overall, the results strongly suggest that the violence effect is related to the PAN taking office, as opposed to a more general effect of the alternation of political power. Column (1) repeats the baseline specification from Table 3, which includes municipalities with close elections where a PAN candidate was the winner or runner-up. Column (2) reports a triple interaction specification that distinguishes between municipalities where the PAN was the incumbent versus municipalities where another party held the incumbency. 21 The estimated effect of a PAN mayor taking office - relative to a non-pan mayor taking office - is large and statistically significant regardless of whether the PAN held the incumbency. The effect is not statistically different across these two samples. Similarly, Column (3) reports 21 In Mexico, mayors cannot run for re-election, so regardless of the party of the incumbent a new politician always takes office with each electoral cycle. 22

24 a triple interaction specification which distinguishes between whether the PAN candidate faced an opponent from the PRI, which is the historically dominant party. Again, there are not statistically significant differences in drug trade-related violence in the post-inauguration period. Next, column (4) includes only close elections where the PRI and PRD - Mexico s two other major parties - received the two highest vote shares. There is no statistically or economically significant difference in drug trade-related violence following a PRI mayor s inauguration as compared to a PRD mayor s inauguration. Column (5) includes all close elections in the sample, regardless of which parties received the two highest vote shares. Here, the post-inauguration dummy is interacted with a dummy equal to 1 if there is an alternation in the political party of the mayor following the inauguration of new authorities. Not surprisingly given the results in columns (1) through (4), the coefficient on this interaction term is 0.057, which is substantially smaller than the baseline and marginally significant. Finally, column (6) repeats the baseline specification from column (1) on a sample that includes all municipalities where the PAN was the winner or runner-up. While the coefficient on PAN win x post-inauguration is positive, it is very small in comparison to the estimate from the RD sample and is not statistically significant. Politics is likely to be meaningfully different in municipalities with very competitive elections as compared to municipalities without competitive elections, and omitted variables bias in column (6) could also help explain the difference between the estimates in column (1) versus column (6). The examination of drug-trafficking spillovers below will use only the plausibly exogenous variation generated by close elections to identify spillover effects. 4.4 Trafficking Industrial Organization and Violence Eighty-five percent of drug trade-related violence consists of people involved in the drug trade killing each other. Before turning to the spillover effects of close PAN victories, I provide a brief investigation of why close PAN victories lead to such large increases in violence. The evidence presented in this section suggests that the violence reflects rival traffickers attempts to wrest control of territories after crackdowns initiated by PAN mayors have challenged the incumbent criminals. In order to shed light on why PAN victories lead to increased violence, I construct several measures of the industrial organization of the drug trade. First, I categorize municipalities into four categories using confidential government data that identifies whether the dominant drug trafficking group in a municipality is a major drug trafficking organization (DTO) or 23

25 a local gang. 22 The categories are: 1) municipality controlled by a major DTO and borders territory controlled by a rival DTO (9.5% of the sample), 2) municipality controlled by a major DTO and does not border territory controlled by a rival DTO (20% of the sample), 3) controlled by a local drug gang (33% of the sample), and 4) no known drug trade presence (37.5% of the sample). Municipalities with no known drug trade presence had not had any drug trade-related homicides or illicit drug confiscations at the time these data were compiled, and local authorities had not reported the presence of a drug trade-related group to federal authorities. Columns 1 and 2 of Table 5 report a triple interaction specification based on equation (7), where dummies for the three categories of drug trade presence are interacted with post-inauguration and post-inauguration x PAN win. No known drug trade group is the omitted category. Column 2 also includes triple interactions with several other important municipal characteristics: income per capita, population density, and distance to the U.S. The most pronounced result is that the effect of a close PAN victory on violence is very large in municipalities controlled by a major DTO that border a rival DTO s territory. A close PAN victory increases the probability that a drug trade-related homicide occurs in a given month in this set of municipality s by a highly significant 48 percentage points, and the estimated effects change little when interactions between PAN win, post-inauguration, and the three controls are included. The effect for municipalities with a major DTO that do not border municipalities controlled by rival DTOs is also somewhat larger than the effect for municipalities with no known drug trafficking presence, though the difference is smaller and no longer statistically significant when the interactions with the controls are included. The PAN win x post inauguration effects are similar for municipalities with a local gang and municipalities with no known drug trafficking presence, with a close PAN win increasing the probability of a drug trade-related homicide by around 7 percentage points. This effect is statistically significant at the five percent level. The fact that the estimates are similar for these groups of municipalities suggests that local drug trade presence is under-reported. Overall, these results support the hypothesis that government crackdowns in Mexico have increased violence through rival traffickers attempts to wrest control of territories after crackdowns have challenged the incumbent traffickers. Next, columns (3) and (4) use the cross-sectional trafficking model developed in Section 3 to predict how many trafficking paths pass through a municipality. 23 The mean number of trafficking paths in the sample is 4. For each additional predicted trafficking path, the 22 The major DTOs during the sample period are Beltran, Familia Michoacana, Golfo, Juarez, Sinaloa, Tijuana, and Zetas. 23 The model with parsimonious costs on U.S. points of entry is used. The appendix shows that results are very similar when the other specifications of the congestion costs are used. 24

26 probability that a drug trade-related homicide occurs in a given month following a close PAN victories increases by around 0.8 percentage points. 24 Finally, I use the trafficking model to calculate the total detour costs that would be imposed if trafficking routes could no longer pass through the municipality under consideration. Columns (5) and (6) interact PAN win x post-inauguration with a dummy equal to one if the total detour costs for municipality m are above the median detour costs in the sample. The effect of a close PAN victory on drug trade-related homicides is ten percentage points higher in municipalities that require high detour costs to circumvent. Together, the results in columns (3) through (6) suggest that the violence response is larger in municipalities that are more valuable to control. The characteristics examined in Table 5 are highly correlated, and the presence of drug trafficking groups is likely to be an outcome of the network structure of drug trafficking. Thus, I cannot separately identify the impacts of each of the characteristics examined. 25 Nevertheless, together the results suggest that the industrial organization of drug trafficking exerts important effects on the violence response to close PAN victories, with violence increasing the most in municipalities that are the most valuable to control. 5 A Network Analysis of Spillover Effects This section uses the network model of the drug trade and plausibly exogenous variation in the cost of trafficking drugs, induced by close elections, to identify the spillover effects of PAN crackdowns on the drug trade. First, I test whether the network model is predictive of the diversion of drug traffic following close PAN victories. Then, I use variation in predicted routes induced by the outcomes of these close elections to examine violence and economic spillover effects. 5.1 Do close PAN victories divert drug traffic? I begin by examining whether the trafficking model estimated on a cross-section of data from the beginning of the sample period can accurately predict the diversion of drug traffic in response to PAN crackdowns occurring during the remainder of the sample period. The costs 24 It would be interesting to see if there was a differential effect depending on whether the predicted trafficking path was controlled by a rival or ally trafficking organization. However, over half of the producing municipalities are controlled by local gangs not typically involved in large scale trafficking, and we do not know which larger trafficking organization(s) they ally with to traffic each shipment of illicit drugs. 25 When the characteristics are included together, they remain large and statistically significance. However, these estimates are difficult to interpret since the presence of drug trafficking groups may be endogenous to the network characteristics. 25

27 of trafficking drugs through a municipality are assumed to increase to infinity (or by some other positive proportion) following the inauguration of narrowly elected PAN mayors. Thus, from 2007 through 2009, the costs of traversing various edges in the Mexican road network change in response to close PAN victories, and this induces month-to-month variation in predicted trafficking routes throughout Mexico. Specifically, I run regressions of the following form: conf mst = β 0 + β 1 DummyRoutes mst + ψ st + δ m + ɛ mst (8) where conf mst is actual illicit drug confiscations of domestically produced drugs in municipality m in month t. Both an indicator measure and a continuous measure are explored. DummyRoutes mst is a dummy variable equal to 1 if the model predicts that municipality m contains a drug trafficking route at time t, ψ st is a month x state fixed effect, and δ m is a municipality fixed effect. I focus on the extensive margin of drug trafficking because data about the supply of drugs produced at each origin are not available. Because variation in routes may be correlated across space, the error term is clustered simultaneously by municipality and state-month (following the two-way clustering of Cameron et al. 2011). The sample excludes municipalities with close elections, since the aim of the model is to predict spillovers from these elections, and it also excludes producing municipalities, since they mechanically contain a trafficking route. Table 6 reports estimates from equation (8). For comparison purposes, Panel A predicts trafficking routes using a very simple model, where traffickers choose the shortest path between producing municipalities and the nearest U.S. point of entry, avoiding municipalities that have experienced a close PAN victory. This provides an intuitive benchmark that can be compared to the model with congestion costs, whose parameters were estimated in Section 3. Panel B uses the model with parsimonious congestion costs on U.S. points of entry to predict trafficking routes. I report this specification because it is slightly more predictive than the others. Results are highly robust to the specification of the congestion costs used, and estimates from (8) for the other two specifications of congestion costs are reported in the appendix. In column (1), the dependent variable is a dummy equal to one if domestically produced illicit drugs are confiscated in a given municipality-month and equal to zero otherwise. The value of confiscations (evaluated at Mexican illicit drug prices) in a municipality-month must be equal to at least $1,000 USD to be included in the sample. Occasionally the total value of confiscations in a municipality month is less than $1,000 and positive, and such 26

28 confiscations are very likely to be from individual consumers and not from drug traffickers. 26 Within-municipality variation in predicted trafficking routes is positively correlated with within-municipality variation in actual illicit drug confiscations, both when simple shortest paths are used to predict trafficking routes and when the model with congestion costs is used. When a municipality acquires a predicted trafficking route, actual drug confiscations increase by around 1.5 percentage points, relative to a sample average probability of confiscations in a given municipality-month of 5.3 percent. This correlation is statistically significant at the 1% level in both Panel A and Panel B. While congestion costs are about half of overall costs in the trafficking model that estimates the predicted routes used in Panel B, these costs do not change the optimal routes dramatically relative to a world without congestion costs. In column (2), the dependent variable is equal to the log of the value (in US dollars) of domestic illicit drug confiscations in the municipality-month if confiscations are positive and equal to zero otherwise. to 1,000 USD, this measure is always positive. 27 Because all positive confiscation values are at least equal The correlation between the log value of confiscations and predicted trafficking routes is large, positive, and statistically significant at the one percent level regardless of whether the shortest paths approach or the congestion costs model are used to predict the routes. Acquiring a predicted trafficking route is associated with an increase in the value of confiscated drugs of around nineteen percent. One concern with interpreting these results is that the estimated relationship between predicted trafficking routes and actual confiscations could result from the direct effects of close PAN victories rather than from diverted drug traffic. This would occur if close PAN victories mostly divert drug traffic nearby and if PAN authorities are more likely to coordinate with the military or federal police, who in turn become active in an entire region. It is much more difficult to tell a story in which close PAN victories would directly affect drug trade outcomes in municipalities located further away, in a manner that would mimic changes in predicted drug traffic. Thus, columns (3) and (4) examine whether the model remains predictive when municipalities bordering those that have experienced a close PAN victory are dropped from the sample, reducing the sample size by around 20%. The estimated coefficients are similar in magnitude to those reported in columns (1) and (2) and are statistically significant at the 5 percent level. To shed further light on the plausibility of the model, columns (5) through (8) report placebo checks. First, I assume, contrary to the regression discontinuity evidence, that the 26 Estimates are robust to using a variety of different cut-offs for the minimum value of drugs confiscated to construct the confiscations dummy variable (results available upon request). 27 Working in logs is attractive because drug confiscations are highly right-skewed, with several major drug busts resulting in tens of millions of dollars worth of confiscated drugs. Using log values makes the data more normally distributed. Moreover, using logs aids in the interpretation of the results. 27

29 costs of passing through municipalities that have experienced a close PAN loss are infinity, whereas there is no additional cost beyond traversing the physical distance to traffic drugs through a municipality that has experienced a close PAN win. This offers a very basic test of whether the model loses its predictive power when it uses the wrong shocks. Columns (5) and (6) show that the model does in fact lose its predictive power when this implausible assumption is made about the effects of close elections. Next, columns (7) and (8) test whether variation in routes induced by close PAN victories is correlated with variation in cocaine confiscations. Because the network model only uses origins for domestically produced drugs, the predicted routes measure should be uncorrelated with confiscations of cocaine as long as cocaine routes have different origins. Columns (7) and (8) document that the coefficients on the predicted routes measure are small and statistically insignificant, whether a dummy or value measure of cocaine confiscations is used as the dependent variable. These results lend further support to the validity of the model. In Table 6, the model assumes that the cost of passing through a municipality that has experienced a close PAN victory is infinity, but the true cost could be substantially less. Figure 4 explores whether the relationship between predicted trafficking routes and the log value of domestic drug confiscations is robust to assuming that a close PAN victory proportionally increases the physical distance costs of trafficking drugs through a given municipality. The x-axis gives the factor α by which the cost of trafficking drugs through a given municipality is multiplied if the municipality has experienced a close PAN victory. The point estimates and confidence intervals for the coefficients on the routes dummy for each of the specifications of the PAN cost parameter are plotted on the y-axis. 95% confidence bands are shown with a thin black line and 90% confidence bands with a slightly thicker black line. The figure contains two panels, one using shortest paths and the other using the model with congestion costs. 28 Moving from left to right across the x-axis, the first two cost factors examined are 0.25 and 0.5. These serve as placebo checks. The RD evidence indicates that a close PAN victory makes trafficking drugs more costly, whereas multiplying the cost of an edge by 0.25 or 0.5 would imply that PAN victories reduce the cost of trafficking drugs. Moving further along the x-axis, multiplicative cost factors ranging from 1.5 to 10 are explored. For comparison purposes, the far right of each figure shows the estimate when an infinite cost of passing through a PAN municipality is assumed. 29 The placebo check estimates generated by using cost factors of 0.25 or 0.50 are small, and none are statistically significant at the 10% level, 28 Figures for all three specifications of the congestion costs are available in the appendix. 29 Cost values between 0.5 and 1.5 are not informative, as values close to 1 do not generate enough variation in the edge weights over time to create within-municipality variation in trafficking routes. 28

30 lending support to the validity of the approach. The majority of the estimates for cost values greater than one are similar to those generated by the model that imposes infinite costs, further documenting the robustness of the model s predictions to alternative reasonable assumptions about the network cost function Violence and economic spillovers Overall, these results indicate that the trafficking model estimated in Section 3 accurately predicts the diversion of illicit drug traffic following close PAN victories. A simple model assuming that traffickers take the shortest distance path to the nearest U.S. point of entry is also predictive. Now, I use predictions about the diversion of drug traffic to explore whether close PAN victories have spillover effects. Due to space constraints, I focus on the model with congestion costs, which is slightly more predictive. The partial correlation coefficient between its routes measure and actual confiscations is 0.018, as compared to for the shortest path routes measure. Estimates of spillovers using the shortest path routes measure and also using routes estimated for the other specifications of the congestion costs are presented in the appendix and are similar to those discussed in the main text. Ideally, I would use predicted trafficking routes as an instrument for actual trafficking routes, but actual trafficking routes are unobserved. Proxies for actual routes, such as confiscations, may contain non-classical measurement error, biasing instrumental variables estimates. Thus, I focus on testing whether there is a reduced form relationship between predicted drug trafficking presence and violence or economic outcomes. First I explore violence spillovers. Table 7 reports estimates from equation (8), with violence measures used as the dependent variable. Column (1) shows that the presence of a predicted trafficking route increases the probability that a drug trade-related homicide occurs in a given month by 1.5 percentage points (s.e.=0.5), which can be compared to the average sample probability of 4.4 percent. This effect is statistically significant at the 1 percent level. It increases slightly, to 0.18, and remains statistically significant when municipalities bordering a municipality that has experienced a close election are excluded from the sample. This strongly suggests that the increase in violence results from diverted drug traffic. In contrast, columns (2) and (4) show that the relationship between predicted trafficking routes and drug trade-related homicides per 10,000 municipal inhabitants, while positive and large relative to the sample mean, is not statistically significant. Moreover, column (5) documents that the presence of a predicted trafficking route has no effect on non-drug trade related homicides. Finally, columns (6) and (7) report results from the basic 30 Estimates are similarly robust when the confiscations dummy is used (see the appendix). 29

31 placebo check in which close PAN losses rather than close PAN victories increase the costs of trafficking drugs. As expected, there is no longer a relationship between predicted routes and drug trade-related violence when routes are predicted using this incorrect shock. Next, I turn to economic spillovers. I use quarterly data from the National Survey of Occupation and Employment (Encuesta Nacional de Ocupación y Empleo), collected by the National Institute of Statistics and Geography (INEGI), to construct measures of labor market outcomes for 2007 through Because the economic data are quarterly and only available for a sample of municipalities, they provide less power than the monthly confiscations and violence data, which are available for all municipalities. Thus, a priori we would expect the relationship between predicted trafficking routes and the economic outcomes to be somewhat less precisely estimated than the relationships between predicted trafficking routes and illicit drug confiscations or violence. Table 8 reports results from regressing male and female labor force participation, as well as formal and informal sector wages of prime age males, on the predicted trafficking routes dummy, state x month fixed effects, and municipality fixed effects. 31 The correlations between predicted trafficking routes and male labor force participation (column 1), as well as between predicted trafficking routes and formal sector wages, are relatively small and statistically insignificant. This suggests that variation in drug trafficking presence over time - at least of the magnitude documented here - has little effect on these outcomes. In contrast, the presence of drug traffickers in a municipality impacts both informal sector wages and female labor force participation. Columns (2) and (4) document that the presence of a predicted trafficking route lowers informal sector wages by around two and a half percent and lowers female labor force participation by around 1.26 percentage points, relative to an average rate of female labor force participation of 51 percent. The coefficients tend to be of similar magnitude regardless of the specification of the congestion costs. The impact on female labor force participation is significant at the 5 percent level, and the impact on informal sector wages is marginally significant. Both of these results remain similar in magnitude and statistical significance when municipalities bordering a municipality that has experienced a close election are excluded from the sample. It is hard to think of a plausible reason why PAN mayors would directly affect economic outcomes in municipalities not immediately adjacent, given their limited budgetary and economic powers. Hence this robustness suggests that theeconomic spillovers are the result of PAN crackdowns diverting drug traffic elsewhere. Moreover, columns (7) and (8) examine the relationship between predicted routes and these outcomes when close PAN losses 31 The analysis of wages is limited to prime age males because they tend to participate inelastically in the labor force, reducing concerns about selection bias in the wage regressions. 30

32 instead of close PAN victories are assumed to increase the costs of trafficking drugs through a municipality. As expected the coefficients are smaller and no longer statistically significant. These results are consistent with qualitative evidence that drug trafficking organizations engage in widespread extortion, particularly of informal sector producers, demanding taxes in return for protection. This evidence was outlined in Section 2. Drug trafficking presence may reduce female labor force participation because women work disproportionately in the informal sector or because their labor force participation is more elastic than that of men. In summary, while the economic data do not provide as much power as the confiscations data examined in the previous section, overall the evidence strongly suggests that drug trafficking presence decreases the wages of informal sector producers and decreases female labor force participation. Given that those harmed are amongst the most vulnerable in society, the impacts on welfare may be substantial. 6 Policy Implications Note: this section is still in progress This paper s results have a number of policy implications. Most basically, they strongly support the hypothesis that policy has had important impacts on drug trade-related violence, both directly and through spillover effects. This has been a point of controversy (Guerrero, 2011; Rios, 2011; Shirk, 2011). More specifically, they suggest that this violence largely reflects rival traffickers attempts to wrest control of territories after crackdowns have challenged the incumbent criminals. The results have implications for the allocation of law enforcement resources. For example, if a major crackdown is planned in one location, monitoring of extortion and other criminal activities on alternate routes can be increased and, if necessary, law enforcement resources deployed. Additionally, the trafficking model can be used to predict bottlenecks in the Mexican road network. These are a useful policy input, because placing law enforcement resources at bottlenecks would increase the distance and congestion costs that traffickers face in transporting drugs to the U.S. The network model accounts for congestion externalities, which may be important in determining where bottlenecks are located. Because of congestion externalities, it is possible that sending law enforcement to a road that is heavily trafficked and requires a long detour to circumvent could actually decrease the total costs faced by traffickers. This well-known result, referred to as Braess s paradox after mathematician Dietrich Braess, has been documented in a number of real world traffic congestion examples, and also occurs for around fifteen percent of the edges in the trafficking network under consideration 31

33 here. 32 The trafficking model accounts for such possibilities when identifying bottlenecks. Suppose that the government s objective is to allocate law enforcement resources - such as police checkpoints - to k edges in the road network, so as to maximize the sum of costs over the minimal trafficking paths that will be used to transport drugs to the U.S. This policy will increase the cost of trafficking drugs on each of the k chosen edges, and traffickers will respond by selecting the cheapest routes given the new set of costs. While the problem specification is straightforward, solving for the optimum is difficult. Allocating police checkpoints to two edges at the same time might increase the objective function more than the summation of changes in the objective function when police checkpoints are allocated to each edge separately. The order in which one allocates resources to edges matters and the optimal solution cannot be found in polynomial time. Thus, I develop an approximate heuristic for identifying bottlenecks. While it is not feasible to guarantee optimality, it is straightforward to assess whether the approach offers a robust solution. For each of k iterations, the algorithm selects an edge to receive law enforcement resources by checking how the total costs of trafficking drugs respond to individually increasing the cost of each of the N most trafficked edges in the network, where N > k. The cost of traversing an edge is increased by a factor of four, which is the estimated size of a close PAN victory that, when combined with the trafficking model, predicts confiscations the best. Of course, other government interventions may be very different from those initiated by PAN mayors, but this approach nevertheless provide a reasonable estimate of the magnitude of a crackdown. The edge that maximizes the increase in the transport costs from distance and congestion is chosen at the end of each iteration. Law enforcement resources are allocated to this edge, implying a new set of edge costs at the start of the next iteration. Figure 5 plots the results of this exercise with k = 25 and N = 250. It highlights municipalities that contain bottlenecks in red. 33 The average drug trade-related homicide rate is plotted in the background. Wile some bottlenecks are in high violence municipalities, others are not. Thus, the identification of bottlenecks provides information that is distinct from simply allocating law enforcement based on rates of violence, which would be a reasonable alternative approach to allocating resources in the absence of a network analysis. A major critique of the Mexican government s policy towards drug trafficking is that it has tended to indiscriminately target drug traffickers, rather than focusing resources using a more systematic and theoretically informed approach (Guerrero, 2011). By incorporating spillovers and well-defined predictions about trafficker s behavior, the network approach provides unique 32 For examples of Braess s paradox in traffic congestion in Seoul, New York, Berlin, Boston, and London, see Youn, Gastner, and Jeong (2008); Easley and Kleinberg (2008, p. 71); and Knodel (1969, p ). 33 I have documented robustness to changing the details of this procedure in several ways, as detailed in the appendix. 32

34 information with the potential to contribute to a more nuanced and efficient allocation of law enforcement resources. 7 Conclusion This study examines how drug traffickers economic objectives influence the direct and spillover effects of Mexican policy towards the drug trade. By exploiting variation from close mayoral elections and a network model of drug trafficking, the study develops three sets of results. First, regression discontinuity estimates show that drug trade-related violence in a municipality increases substantially after the close election of a PAN mayor. This violence consists primarily of individuals involved in the drug trade killing each other. The empirical evidence suggests that the violence largely reflects rival traffickers attempts to wrest control of territories after crackdowns initiated by PAN mayors have challenged the incumbent criminals. Second, the study accurately predicts the diversion of drug traffic following close PAN victories. It does this by estimating a model of optimal routes for trafficking drugs across the Mexican road network to the U.S. When drug traffic is diverted to other municipalities, drug trade-related violence in these municipalities increases, female labor force participation declines, and informal sector wages fall. These results corroborate qualitative evidence that traffickers extort informal sector producers. Finally, the study uses the trafficking model and estimated spillover effects to examine the allocation of law enforcement resources. Overall, the results demonstrate how traffickers economic objectives and constraints imposed by the routes network affect the policy outcomes of the Mexican Drug War. 33

35 References Angrist, J. and A. Kugler (2008): Rural windfall or a new resource curse? Coca, income, and civil conflict in Colombia, The Review of Economics and Statistics, 90, Beckmann, M., C. McGuire, and C. Winston (1956): Studies in the Economics of Transportation, Yale University Press. Cameron, A., J. Gelbach, and D. Miller (2006): Robust inference with multi-way clustering,. Chubb, J. (1989): The mafia and politics: the Italian state under siege, Cornell Univ Press. CONADIC, Secretaría de Salud (2008): National Addictions Survey,. Di Tella, R. and E. Schargrodsky (2004): Do police reduce crime? Estimates using the allocation of police forces after a terrorist attack, The American Economic Review, 94, Easley, D. and J. Kleinberg (1988): Desperados: Latin drug lords, US lawmen, and the war America can t win, Viking. (2010): Networks, crowds, and markets: Reasoning about a highly connected world, Cambridge Univ Press. Gambetta, D. (1993): The Sicilian Mafia. Guerrero-Gutiérrez, E. (2011): Security, Drugs, and Violence in Mexico: A Survey, 7th North American Forum, Washington D.C. Imbens, G. and K. Kalyanaramang (2009): Optimal bandwidth choice for the regression discontinuity estimator,. Isacson, A. (2010): Colombia: Don t Call it a Model, Policy Document: Washington Office on Latin America. Leech, G. (2000): The Failure of Coca Eradication in Peru and Bolivia, Colombia Journal. Manski, C. (1993): Identification of endogenous social effects: The reflection problem, The Review of Economic Studies, 60,

36 Rios, V. (2011a): Evaluating the economic impact of drug traffic in Mexico, Manuscript: Harvard University Department of Government. (2011b): Understanding Mexico s Drug War, Manuscript: Harvard University Department of Government. (2011c): Why did Mexico become so violent? Criminal diversification and drugrelated violence: The case of La Familia, Manuscript: Harvard University Department of Government. Secretaría de Seguridad Pública (2010): Informe del Estado de la Seguridad Pública en México,. Secretariado Ejecutivo del Sistema Nacional de Seguridad Pública (2008): National criminal incidence,. Shannon, E. (1988): Desperados: Latin drug lords, US lawmen, and the war America can t win, Viking. Shirk, D. (2011): Drug Violence and State Responses in Mexico, Manuscript: University of San Diego. U.S. Drug Enforcement Agency (2011): Drug Trafficking in the United States,. Varese, F. (2011): Mafias on the Move: How Organized Crime Conquers New Territories, Princeton Univ Press. W, K. (1969): Graphentheoretische Methoden und ihre Anwendungen, Springer-Verlag. Youn, H., M. Gastner, and H. Jeong (2008): Price of anarchy in transportation networks: Efficiency and optimality control, Physical review letters, 101,

37 Table 1: Trafficking Model Parameter Estimates (1) (2) (3) Crossing Costs Full parsimonious flexible congestion model model costs φ t 62.34*** [2.72] (1.41) φ p 36.48*** [2.07] (1.40) φ Q1 t 3.24*** 6.33*** [0.30] [0.788] (0.25) (0.718) φ Q2 t 13.19*** 25.24*** [2.14] [1.623] (1.89) (1.489) φ Q3 t 13.86*** 13.03*** [4.37] [1.37] (4.08) (1.32) φ Q4 t 18.81*** 32.41*** [0.86] [2.58] (0.83) (2.33) φ small p 64.47*** 70.35*** [9.76] [7.24] (9.16) (6.81) φ large p 55.34*** 52.39*** [8.43] [6.66] (7.46) (6.06) φ int 0.005*** [0.0008] (0.0007) δ 1.88*** 1.57*** 1.49*** [0.05] [0.15] [0.07] (0.04) (0.12) (0.06) γ 1.06*** [0.10] (0.09) κ 0.763*** 0.91*** 0.86*** [1.69] [0.08] [0.05] (0.65) (0.07) (0.05) Notes: Column 1 reports the simulated method of moments parameter estimates for the model with parsimonious congestion costs on U.S. points of entry, Column 2 reports the parameter estimates for the model with flexible congestion costs on U.S. points of entry, and Column 3 reports the parameter estimates for the model with congestion costs on both U.S. points of entry and interior edges. Conley (1999) standard errors are in brackets, and robust standard errors are in parentheses. 36

38 Table 2: Pre-characteristics (5% vote spread) PAN PAN t stat on lost won difference Crime characteristics Monthly drug-trade related homicides (Dec Jun. 07) (-0.41) Monthly drug-trade related homicide dummy (Dec Jun. 07) (0.51 ) Monthly police-criminal confrontation deaths (Dec Jun. 07) (0.57) Monthly confrontation deaths dummy (Dec Jun. 07) (-0.01) Annual homicide rate per 100,000 inhab. ( ) (0.46) Municipal political characteristics Mun. taxes per capita (2005) (-0.22) Mun. taxes per $ income (2005) (-0.05) PAN incumbent (-0.07) PRD incumbenet (-0.59) % alternations ( ) (0.34) PRI never lost ( ) (0.97) Demographic characteristics Population (2005) (-0.35) Population density (2005) (-0.41) Migrants per capita (2005) (0.69) Economic characteristics Income per capita (2005) (0.53) Mun. Gini (2005) (1.47) Malnutrition (2005) (-0.52) Mean years schooling (2005) (-0.32) Infant mortality (2005) (-0.22) HH w/o access to sewage (2005) (-0.05) HH w/o access to water (2005) (0.62) Marginality index (2005) (0.23) Road network characteristics Average detour length (km) (-0.46) Total roads (km) (-1.28) Road density (km/km 2 ) (-1.23) Distance U.S. (km) (0.62) Geographic characteristics Elevation (m) (-0.19) Slope (degrees) (-0.89) Surface area (km 2 ) (-1.34) Average min. temperature, C ( ) (0.57) Average max. temperature, C ( ) (0.45) Average precipitation, cm ( ) (-0.82) Observations Notes: Data on population, population density, mean years of schooling, and migrants per capita are from II Conteo de Poblacion y Vivienda, INEGI (National Institute of Statistics and Geography, 2005). Data on municipal tax collection are from Sistema de Cuentas Municipales, INEGI. Data on household access to sewage and water are from CONAPO (National Population Council) (2005). Data on malnutrition are from CONEVAL (National Council for Evaluating Social Development Policy), Indice de Reazgo Social (2005). Data on infant mortality are from PNUD Mexico (UN Development Program, 2005). The marginality index is from CONAPO (2005), and incorporates information on literacy; primary school completion rates; access to electricity, sewage, and running water; household overcrowding; construction materials used in households; municipal population in rural areas; and household income. Data on distance to the U.S. and other road network characteristics are from the author s own calculations, using GIS software. Electoral data are from Mexico Electoral -Banamex and electoral results published by the Electoral Tribunals of each state. The geographic characteristics are from Acemoglu and Dell (2010). Data on homicides ( ) are from INEGI and data on drug trade-related violence are from confidential sources.

39 Table 3: Close PAN Elections and Violence Homicide dummy Vote spread < 5 points Vote spread less than 5 points; Homicide dummy Vote spread less than: Homicide Quartic Cubic Linear Post Lame Pre 7 points 5 points 3 points count rate RD polynomial inaug. duck election (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Panel A: Drug trade-related homicides PAN win x lame duck (0.048) (0.059) (0.079) (0.152) (0.025) (0.090) (0.088) (0.058) PAN win 0.115*** 0.132*** 0.159*** 0.553** 0.088** 0.204*** 0.204*** 0.147*** x post-inaug. (0.039) (0.047) (0.059) (0.235) (0.038) (0.062) (0.064) (0.051) PAN win 0.171* (0.101) (0.106) (0.081) R Clusters Observations Mean dep. var Panel B: Non drug trade-related homicides PAN win * x lame duck (0.060) (0.066) (0.087) (0.165) (0.055) (0.087) (0.089) (0.068) PAN win x post-inaug. (0.051) (0.059) (0.070) (0.213) (0.049) (0.068) (0.071) (0.059) PAN win (0.096) (0.120) (0.125) R Clusters Observations Mean dep. var Notes: The unit of observation is the municipality-month in columns (1) through (8) and the municipality in columns (9) through (11). PAN win is a dummy equal to one if a PAN candidate won the election. Lame duck is a dummy equal to one if the observation occurred between the election and inauguration of a new mayor. Post-inaug. is a dummy equal to one if the observation occurred in the five months following the inauguration of a new mayor. The omitted category is pre-election. Columns (1) through (8) include a lame duck main effect, a post-inauguration main effect, and month x state and municipality fixed effects. Columns (1) through (5) include a quadratic polynomial in vote spread, interacted separately with a post-inauguration dummy and a post-election dummy, whereas column (6) uses a quartic vote spread polynomial, column (7) a cubic polynomial, and column (8) a linear vote spread. All columns limit the sample to municipalities where a PAN candidate was the winner or runner-up. The coefficients in columns (9) through (11) are estimated using local linear regression. Column (1) limits the sample to municipalities with a vote spread of seven percentage points or less, columns (2) and (4) through (11) to municipalities with a vote spread of five percentage points or less, and column (3) to municipalities with a vote spread of three percentage points or less. In column (9) the sample is limited to the post-inauguration period, in column (10) it is limited to the lame duck period, and in column (11) it is limited to the pre-election period. Robust standard errors, clustered by municipality in columns (1) through (8), are in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%.

40 Table 4: Local Politics and Violence (1) (2) (3) (4) (5) (6) Elections involving PAN Alternative samples PAN PRI PRI v. Any All Baseline Incumbent Opponent PRD alternation muns. PAN win x post-inaug *** 0.118** 0.137** (0.047) (0.048) (0.066) (0.019) PAN win x post-inaug x PAN Incumbent (0.056) PAN win x post-inaug x PRI opponent (0.052) PRI win x post-inaug (0.056) Alternate x post-inaug * (0.035) R Clusters Observations Mean dep. var Post-innaug. effect 0.186*** (PAN incumbent) (0.070) Post-innaug. effect 0.107** (PRI opponent) (0.046) Notes: The dependent variable in all columns is a dummy variable equal to one if a drug trade-related homicide occurred in a given municipality-month and equal to zero otherwise. PAN win is a dummy equal to one if a PAN candidate won the election, PRI win is a dummy equal to one if a PRI candidate won the election, Alternate is a dummy equal to one for any alternation of the political party controlling the mayorship, Pan Incumbent is a dummy equal to 1 if the municipality had a PAN incumbent, and PRI opponent is a dummy equal to one if the PAN candidate faced a PRI opponent. All columns include month x state and municipality fixed effects, a post-inauguration main effect, and a quadratic polynomial in vote spread interacted with the post-election dummy. Column 2 contains a post-inauguration x Pan incumbent interaction, and column (3) contains a post-inauguration x PRI opponent interaction. In columns (1) through (5) the sample is limited to municipalities with a vote spread of five percentage points or less. Columns (1) through (3) limit the sample to municipalities where a PAN candidate was the winner or runner-up, Column (4) limits the sample to municipalities with a close election between PRI and PRD candidates, and Column (5) includes all municipalities with a close election, regardless of the political parties involved. Column (6) includes all elections where a PAN candidate was the winner or runner-up. Robust standard errors, clustered by municipality, are in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%. 39

41 Table 5: Trafficking Industrial Organization and Violence (1) (2) (3) (4) (5) (6) Dependent variable is the drug-trade related homicide dummy PAN win x post-inaug 0.073** 0.068** 0.109** 0.103** 0.102** 0.070** (0.034) (0.031) (0.043) (0.041) (0.044) (0.030) PAN win x post-inaug 0.395*** 0.394*** x borders rival (0.112) (0.112) PAN win x post-inaug 0.142** x borders allies (0.065) (0.078) PAN win x post-inaug x local gang (0.022) (0.025) PAN win x post-inaug 0.008** 0.005** x predicted flows (0.004) (0.002) PAN win x post-inaug 0.103* 0.103* x long detour (0.058) (0.058) Controls no yes no yes no yes R-squared Clusters Observations 1,639 1,639 1,529 1,529 1,639 1,639 Borders rival 0.468*** 0.462*** effect (0.115) (0.112) Borders allies 0.215*** 0.136* effect (0.076) (0.077) Local gang 0.086** 0.079** effect (0.036) (0.034) Long detour 0.204*** 0.172** effect (0.068) (0.069) Notes: The dependent variable in all columns is a dummy variable equal to one if a drug trade-related homicide occurred in a given municipality-month. PAN win is a dummy equal to one if a PAN candidate won the election, borders rival is a dummy equal to one if the municipality is controlled by a major DTO and borders territory controlled by a rival DTO, borders allies is a dummy equal to one if the municipality is controlled by a major DTO and does not border territory controlled by a rival, and local gang is a dummy equal to one if the municipality is controlled by a local gang. No known drug trade presence is the omitted category. Predicted flows gives the total illicit drug flows predicted to pass through municipality m by the cross-sectional trafficking model, and long detour is a dummy equal to one if the total detour costs when drug traffic cannot pass through municipality m is greater than the median total detour costs. All columns include month x state and municipality fixed effects, a post-inauguration main effect, and a quadratic polynomial in vote spread interacted with a post-inauguration dummy. Columns (2), (4), and (6) also include triple and lower interactions between PAN win, post-inauguration, and the following characteristics: municipal income per capita, population density, and distance to the U.S. The sample is limited to municipalities with a vote spread of five percentage points or less. Robust standard errors, clustered by municipality, are in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%. 40

42 Table 6: The Diversion of Drug Traffic (1) (2) (3) (4) (5) (6) (7) (8) Full Sample Limited Sample Placebo Paths Full Sample Domestic illicit drug confiscations Cocaine confiscations Dummy Value Dummy Value Dummy Value Dummy Value Panel A: Shortest Paths Predicted 0.016*** 0.170*** 0.015** 0.162** routes dummy (0.005) (0.050) (0.006) (0.063) (0.004) (0.038) (0.004) (0.020) R Panel B: Congestion Model Predicted 0.015*** 0.178*** 0.013** 0.159** routes dummy (0.005) (0.059) (0.006) (0.064) (0.006) (0.069) (0.004) (0.023) R Municipalities Observations 69,153 69,153 58,238 58,238 69,153 69,153 69,153 69,153 Mean dep. var Notes: The dependent variable in columns (1), (3), and (5) is a dummy equal to 1 if domestic illicit drug confiscations are made in a given municipality-month; the dependent variable in columns (2), (4), and (6) is the log value of domestic illicit drug confiscations (or 0 if no confiscations are made); the dependent variable in column (7) is a dummy equal to 1 if cocaine confiscations are made in a given municipality-month; and the dependent value in column (8) is the log value of confiscated cocaine (or 0 if no confiscations are made). Columns (5) and (6) use the placebo network, as described in the text. Columns (3) and (4) limit the sample to municipalities that do not border a municipality that has experienced a close election. Panel A predicts trafficking routes using the shortest paths model, and Panel B uses the model with parsimonious congestion costs on U.S. points of entry. All columns include month x state and municipality fixed effects. Standard errors clustered by municipality and month x state are reported in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%.

43 Table 7: Violence Spillovers (1) (2) (3) (4) (5) (6) (7) Full Sample Limited Sample Full Sample Placebo Paths Non-drug Drug-related Drug-related homicide homicide homicide dummy rate dummy rate rate dummy rate Predicted 0.015*** *** routes dummy (0.005) (0.019) (0.006) (0.025) (0.007) (0.006) (0.013) R Municipalities Observations 69,153 69,153 58,238 58,238 69,153 69,153 69,153 Mean dep.var Notes: The dependent variable in columns (1), (3), and (6) is a dummy equal to 1 if a drug trade-related homicide occurred in a given municipality-month; the dependent variable in columns (2), (4), and (7) is the drug trade-related homicide rate per 10,000 municipal inhabitants, and the dependent variable in column (5) is the non-drug trade-related homicide rate per 10,000 municipal inhabitants. Columns (6) and (7) use the placebo network, as described in the text. Columns (3) and (4) limit the sample to municipalities that do not border a municipality that has experienced a close election. All columns include month x state and municipality fixed effects. Standard errors clustered by municipality and month x state are reported in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%.

44 Table 8: Economic Spillovers (1) (2) (3) (4) (5) (6) (7) (8) Actual Network Placebo Network Male Female Formal Informal Female Informal Female Informal participation log wages participation wages participation wages Predicted ** * ** * routes dummy (0.302) (0.570) (0.012) (0.013) (0.673) (0.017) (0.636) (0.020) R Municipalities Observations 9,821 9, , ,302 7, ,633 9, ,302 Mean Dep. Var Notes: The dependent variable in column (1) is average municipal male labor force participation; the dependent variable in columns (2), (5), and (7) is average municipal female labor force participation, the dependent variable in column (3) is log wages of formal sector workers; and the dependent variable in columns (4), (6), and (8) is log wages of informal sector workers. Columns (7) and (8) use the placebo network, as described in the text. All columns include month x state and municipality fixed effects. The sample in columns (5) and (6) excludes municipalities that border a municipality that has experienced a close election. Standard errors clustered by municipality and month x state are reported in parentheses. * significant at 10%, ** significant at 5%, *** significant at 1%.

45 Figure 1: Illustration of spillovers methodology (a) Legend (b) Basic environment (c) Close PAN victory increases costs; routes change (d) Do confiscation patterns also change?

46 Figure 2: Results from a DD Perspective A. Drug trade related homicide dummy B. Drug trade related homicide rate Coefficient on PANwin x post Months from Election/Innauguration Coefficient on PANwin x post Months from Election/Innauguration C. Non drug homicide dummy D. Non drug homicide rate Coefficient on PANwin x post Months from Election/Innauguration Coefficient on PANwin x post Months from Election/Innauguration 45

47 Figure 3: Results from a RD Perspective A. Drug trade homicide dummy (Post) B. Drug trade homicide dummy (Pre) Prob. of homicide Prob. of homicide PAN margin of victory PAN margin of victory C. Non drug homicide rate (Post) D. Non drug homicide rate (Pre) Homicide rate Homicide rate PAN margin of victory PAN margin of victory 46

48 Figure 4: Varying the costs imposed by PAN victories Panel A: No congestion costs 0.3 Coefficient on Routes Dummy PAN cost infinity Panel B: Crossing Congestion (3 param) 0.2 Coefficient on Routes Dummy PAN cost infinity

49 Figure 5: Bottlenecks and Violence