Abstract
We examine the effect of the increase in violence that Mexico experienced after launching an aggressive campaign against drug-trafficking organizations on immigration into the United States. We instrument for violence using electoral cycles and consider two channels through which violence impacts migration: local and transit violence. Violence at the municipality of residence increased migration. Conversely, violence on the route to the United States deterred individuals from migrating. Back-of-the-envelope calculations show that between 2007 and 2012, local and transit violence had an overall positive effect on migration. Violence was responsible for a 1.53 percentage point increase in the migration rate.
I. Motivation
In recent years, Mexico has experienced a dramatic increase in violence driven in large part by the violent struggle between and within drug trafficking organizations for control of the profitable drug-trade business. This increase in violence began in 2006 when President Felipe Calderon launched an aggressive military campaign against drug trafficking organizations. The campaign sparked competition, fragmentation, and alliances among criminal groups, leading to instability and a staggering amount of violence. There were approximately 66,000 drug-related homicides, or 6.4 homicides per 10,000 inhabitants, during Calderon’s administration from 2006 to 2012. By comparison, there were only 3,638 drug-related homicides (0.12 homicides per 10,000 inhabitants) in the United States during the same time period and only 8,900 drug-related homicides (0.9 homicides per 10,000 inhabitants) under President Fox’s administration (Heinle, Molzahn, and Shirk 2005). The degree of violence inMexico was also high relative to better-known conflicts. For example, the homicide rate associated with the war against drug cartels in Mexico exceeds the 6.3 civilian death rate reported during the war in Afghanistan over the same period of time and the 1.6 civilian death rate reported during the armed conflicts in Pakistan between 2001 and 2014.1
Over the course of Calderon’s administration, the flow of immigrants from Mexico to the United States decreased significantly. Estimates from the Pew Hispanic Center show that in 2010, for the first time in four decades, the net flow of immigrants from Mexico to the United States was zero. Furthermore, for the period 2009–2014, the net flow of immigrants was negative (Passel, Cohn, and Gonzalez-Barrera 2012; Gonzalez-Barrera 2015). While there has been some research investigating the role of the recession and state immigration laws in causing this decline (Hoekstra and Orozco-Aleman 2017; Bohn, Lofstrom, and Raphael 2014; Chiquiar and Salcedo 2013; Orrenius and Zavodny 2009a, 2009b), little is known about the degree to which the drug war contributed to this reduction in immigration. The purpose of this paper is to assess the extent of the drug war’s role in the decline in immigration from Mexico into the United States.
Violence may have contributed to the change in immigration patterns in a number of ways. Violence imposes a social and economic burden on individuals and businesses that could alter migration incentives. Estimates suggest that the annual cost of violence in Mexico is between 1.0 and 1.5 percent of GDP (BBVA 2014). At a macro level, it decreases personal consumption, domestic investment, and foreign direct investment. At a micro level, it can affect individuals’ employment, job performance, and earnings. Violence also imposes significant emotional costs on individuals. While there are not official statistics, specialists estimate that one million people have been forced to flee their place of residence to find safer areas for themselves and their families (BBC 2012). For these reasons, violence in Mexico could be expected to increase migration to the United States.
However, another channel through which violence can affect migration decisions is through changes in migration costs. The costs of migration have increased with violence as drug cartels have come to control smuggling routes into the United States and as violence has spread to Mexican highways. During recent years, migrants have been frequently subjected to abuses, including extortion, theft, and death at the hands of criminal groups. Moreover, criminal organizations have hijacked cars and passenger buses and have installed checkpoints to control the presence of members of rival organizations, generating fear among travelers.
To estimate the impact of these two channels through which drug violence could affect migration to the United States, we use data on homicides from the National Institute of Statistics, Geography, and Information (INEGI) and data on Mexican immigrants from the Survey of Migration at Mexico’s Northern Border (EMIF). A concern when studying the effect of drug violence on migration decisions is the presence of endogeneity between migration and homicide rates due to unobserved heterogeneity. If the likelihood of observing drug violence in a specific municipality is correlated with the probability of its residents to migrate, the results would be biased. To address the endogeneity of homicides we use an instrumental variables approach.We instrument for violence using electoral cycles at the municipality level. The increase in drug violence during Calderon’s administration has been broadly associated with municipal elections (Dell 2015). This is because drug cartels attempted to influence elections to make sure whoever was elected did not interfere with their criminal activities. Importantly, however, there is little reason to expect that municipal elections would have independent effects on migration incentives. This is because local spending at the municipal level is determined largely by federal transfers rather than discretionary decisions made by the municipal governments themselves.
We measure two different types of violence in order to assess the impact of local violence where prospective migrants live and transit violence on the routes taken to the United States–Mexico border. Local violence is measured using homicide rates at the municipality of residence. For transit violence, we construct two indices that capture the violence migrants are exposed to while traveling to the United States while in their own state, as well as the violence through which they must travel when crossing through other Mexican states.
Results indicate that local violence has increased migration outflows from Mexico to the United States. We calculate an elasticity of migration with respect to homicides of 1.38. Moreover, we find that the elasticity of migration to homicides is highest in northern and central Mexico, which have the highest and lowest average homicide rates, respectively.We attribute this to the fact that economic losses due to violence were likely most severe in those regions, as they had the highest growth rates over the last decade.2 In contrast, we find that transit violence was responsible for a 1.83 percentage point decline in migration between 2007 and 2012. However, our analysis indicates that the impact of local violence outweighs the impact of transit violence.We estimate that the net effect of the increase in violence was to increase migration rates by 1.53 percentage points, or 50 percent relative to the level observed in 2007.
II. Theoretical Framework and Existing Literature
The effect of violence on migration flows is theoretically ambiguous. The neoclassical theory of migration generally treats migration as an individual decision regarding income maximization. It focuses on wage differentials between receiving and sending countries and migration costs. For example, Borjas (1987) develops a twocountry model in which an individual migrates if the expected earnings at the destination, net of migration costs, are higher than the earnings at home. Consequently, if an increase in violence reduces expected earnings in Mexico, the theory predicts that more individuals will find it optimal to migrate, increasing the outflow of workers into the United States.
However, violence could also increase migration costs. Migration costs include not only monetary costs, such as transportation expenses, smuggling fees, and the subsistence cost for the migrant in the host country while finding a job, but also nonmonetary costs, such as the psychic cost of changing one’s environment (Sjaastad 1962) or the emotional cost of being away from family (Taylor 1996). Hence, according to the theory, an increase in migration costs would decrease the number of individuals willing to migrate to the United States.
Some studies have explored the relationship between violence and internal migration. Schultz (1971) finds that migration rates are increasing in the incidence of political homicides in Colombia during the period 1951–1964. Engel and Ibáñez (2007) and Ibáñez and Vélez (2008) identify violence generated by illegal armed groups as an important determinant of internal displacement in Colombia. Morrison (1993) and Morrison and May (1994) study the determinants of internal displacement in Guatemala and find that political violence was a key determinant of migration decisions during the late 1970s and early 1980s.
Other studies have focused on the relationship between violence and international migration. Moore and Shellman (2006) use data on a sample of countries from 1976 to 1995 and find that state violence targeting civilians produces international refugees, whereas civil war and high levels of dissident violence tend to produce internal displacement. Bohra-Misra and Massey (2011) study how armed violence during a period of civil conflict in Nepal influenced domestic and international migration. They find that violence has a nonlinear effect on migration. People migrated only under conditions of extreme violence in which the threats to safety are perceived to exceed the risks of travel.
A few studies have analyzed the effect of violence on immigration into the United States. Stanley (1987) finds that fear of political violence was an important motivator of the Salvadorans who migrated to the United States after 1979. He also finds that military sweeps, rather than death-squad killings, were more strongly associated with undocumented migration to the United States. Shellman and Stewart (2007) predict the factors associated with forced migration and find that trends in Haitian emigration to the United States between 1994 and 2004 were strongly predicted by surges in political violence.
More recent studies have analyzed the effects of the increase in drug violence in Mexico on different economic indicators, such as economic activity and unemployment (Robles, Calderon, and Magaloni 2013), labor force participation (Dell 2015), income growth (Enamorado, Lopez-Calva, and Rodriguez-Castelan 2014), and foreign direct investment (Ashby and Ramos 2013). Moreover, two studies have looked at the effect of drug violence on migration flows. Ríos (2014), using an ordinary least squares (OLS) strategy and 2010 Mexico census data, finds evidence of internal displacement due to increases in violence. One of the limitations of this study is that it cannot distinguish between those who relocate to other cities within Mexico and those who migrated to the United States. Basu and Perlman (2013) also study the effect of violence on migration using 2010 census data. They do not find evidence of international migration after instrumenting for violence at the municipal level using kilometers of federal highways. One explanation for the unexpected finding is that the census does not account for entire families that moved abroad, which might be a likely phenomenon in the event of an increase in drug violence. Moreover, the deadliest years in the war against drug cartels during Calderon’s administration were 2010 and 2011, the effect of which would have been omitted from the analysis of both of these studies.
Our study contributes to the existing literature in several ways. First, we identify two channels through which violence can impact migration rates. Our research focuses on the conditions of violence not only at the source of migration (the individuals’ municipality of residence), but also along the travel route to the United States. Several studies have documented the increase in violence against migrants committed by drug trafficking organizations (Guerrero-Gutierrez 2011; Ríos and Shirk 2011; Walser, McNeill, and Zuckerman 2011; Amnesty International 2010). However, to our knowledge, this is the first study that analyzes the effect of violence along migration routes on the outflowof migrants fromMexico to theUnited States. Second, to instrument for violence, we exploit the large increase in drug-related violence associated with municipal electoral cycles, a phenomenon observedmainly between 2006 and 2012. Finally, by using data onmigration from EMIF, we are able to capture individuals migrating with their families. Those migrants are not captured by other datasets frequently used to study migration flows, such as the Mexican Census or the National Survey of Occupation and Employment (ENOE).
III. Data
We use the number of homicides at the municipality level between 1999 and 2012 from INEGI.3 Figure 1 shows the number of homicides per year. Between 1998 and 2004, Mexico experienced a decrease in the homicide rate; however, that trend was reversed in 2007 with the beginning of the war against drug cartels. Table 1 shows homicide rates in 1999, 2005, and 2011. The homicide rate went from 0.96 homicides per 10,000 inhabitants in 2005 to 2.42 in 2011. Mexico’s annual average homicide growth rate between 2005 and 2011 was 25.4 percent, which is significantly higher than those observed in other countries from Latin America (Honduras 16.0 percent, El Salvador 2.1 percent, Chile 1.0 percent, Nicaragua -1.1 percent, Colombia -2.5 percent), North America (United States -2.7 percent, Canada -2.8 percent), Europe (Italy -1.7 percent, France -4.2 percent, Spain -5.6 percent), Asia (India -0.4 percent, Nepal -3.2 percent, Japan -6.7 percent), and Africa (Algeria 2.7 percent, Morocco -2.1 percent, South Africa -3.6 percent).4
The data on Mexican immigrants are sourced from the EMIF. The EMIF is a crosssectional survey conducted in Mexican border cities that measures the flows of migrants between Mexico and the United States.We use the subsample of Northward-bound migrants born in Mexico with destinations of either Mexican border cities or the United States. The surveys were conducted yearly from 1999 to 2012.
Number of Homicides by Year
Homicide Rates in 1999, 2005, and 2011 (Homicides per 10,000 Inhabitants)
We restrict the sample to include only migrants with the intention of entering the United States.5We use the number of migrants who left their municipality of residence and arrived in border cities with the intention of entering the United States as a proxy for the number of migrants entering the United States. Migration rates by municipality are calculated using the number of migrants observed each year (using sample weights) as a proportion of the population aged 15–64. Data on population are from the 2000 Mexican Census. Table 2 shows average migration and homicide rates during the period of analysis estimated for western, southern, central, and northern Mexico.6
Average Migration and Homicide Rates by Region, 2007–2012
To verify how migration rates constructed using the EMIF compare to those calculated using other data sources, we use information from the National Survey of Occupation and Employment (ENOE). The evidence shows that the number of migrants captured by the EMIF is not systematically different than those captured by the ENOE. A detailed comparison of the migration rates calculated using both surveys is presented in Online Appendix 1.7
To explore the relationship between local and transit violence and the perceptions of safety in different settings (in the municipality, the state of origin, or on the roads), we use Mexico’s 2011 and 2012 National Survey onVictimization and Perception of Public Security (ENVIPE). This survey provides estimates of the number of crime victims and perceptions of public safety at the federal, state, and municipal levels. Finally, data on municipal elections were obtained from the states’ electoral institutes.
Other data series obtained from INEGI include municipal data on suicides, accidental deaths, births, number of registered workers with the Mexican Institute of Social Security (IMSS), amount spent on the welfare program “Oportunidades,” number of banking branches, and the value of agricultural and meat production. Data on unemployment and years of schooling are from the National Employment Survey (ENE) from 1997 to 2004 and from the ENOE from 2005 to 2012.
IV. Empirical Specification
A. Local Violence
In order to study the effect of local violence—violence in the municipality of residence— on the outflows of migrants we estimate the following equation:
where Migration_ratemt is the number of migrants from municipality m in year t as a proportion of the population aged 15–64, and Homicide_ratemt -1 is the number of homicides per 10,000 inhabitants committed in municipality m in period t – 1. To construct migration rates, we use time invariant and predetermined population from the 2000 Census to avoid any mechanical increase in migration rates due to population reductions associated with the increase in homicides.
The vector X contains a set of time-varying controls at the municipal level. The timevarying controls include the birth rate (number of births as a proportion of the population), average years of schooling, unemployment rate, logarithm of the number of registered workers with IMSS (a measure of formal employment), logarithm of cash transfer amounts under the program “Oportunidades” (the most important antipoverty program of the Mexican Government), logarithm of the value of agricultural and meat production (measures of the economic activity), and number of banking branches. Finally, αm are municipality fixed effects, βt are year fixed effects, and εmt is an error term. Standard errors are robust and clustered at the municipality level.
The decision to migrate could be delayed if individuals need time to sell their assets, save enough to cover migration costs, or find a job in the United States. Moreover, if the upsurge in violence is perceived as a temporary phenomenon, it may not affect migration decisions to the United States. To capture the effect of violence on migration if making the decision to migrate requires time, we also estimate Equation 1 using different lagged versions of homicide rates as explanatory variables.We use a two-year homicide rate, a three-year homicide rate, and a cumulative homicide rate from the beginning of the war against drug cartels. The two-year homicide rate is the number of homicides committed in municipality m during the current and last period per 10,000 inhabitants. The threeyear homicide rate is the number of homicides committed the current and last two periods in municipality m per 10,000 inhabitants. The cumulative homicide rate is calculated as the number of homicides committed in municipalitymbetween 2006 and year t – 1 per 10,000 inhabitants.
It is important to note that individuals from different regions of Mexico have different characteristics, and therefore, might have been affected differently by changes in violence. For example, municipalities ofWesternMexico have traditionally been sources of migrants. Workers from that region have larger migration networks at their final destination and are more likely to have previous migratory and work experience in the United States.
In order to account for different characteristics and risks faced by individuals migrating from different regions of Mexico, we estimate an equation with interactions of homicide rates and dummy variables for the four regions of Mexico: northern, central, southern, and western.
First, we estimate Equations 1 and 2 using OLS. However, if unobserved characteristics drive both violence and migration decisions, OLS coefficients would be biased. An example would be if violence were more likely to occur in municipalities that have more miles of highways connecting Mexico and the United States. Violence has been associated with disputes among drug cartels to control the main distribution channels of drugs between Mexico and the United States. Therefore, violence could be positively correlated with the length of the road network in a municipality. Moreover, municipalities with more miles of highways and better infrastructure might also have higher growth rates and levels of economic development, which could negatively impact the probability of migration of their inhabitants. If that is the case, the OLS estimates would underestimate the real effect of violence (negative bias). If the relationship between violence and migration is positive, OLS estimates would be biased towards zero. If the relationship between violence and migration is negative, OLS estimates would be more negative than the coefficient true value.
To overcome endogeneity issues, we construct three instruments using electoral cycles in Mexican municipalities. Municipal election dates are different across states and are not necessarily held at the same time as federal or state elections. In general, all the municipalities of a state have elections at the same time.8 Mayors cannot be reelected. In most states, municipal elections are conducted every three years; however, during the period of analysis, some states had either abnormally long or short mandates.9 We use electoral dates as an instrument for violence given the increase in assassinations, intimidation, and attacks against civilians and politicians that have been associated with election cycles between 2006 and 2012.
Since providing public security is one of the most important tasks of municipal governments, local elections may change public safety policies and practices. During the period of analysis, drug cartels attempted to influence elections with scare tactics and monetary payouts to make sure whoever was elected would not interfere with their criminal activities. The focus was usually on local officials and police departments, who can modify the operational environment for criminal activities through loose or rigid public safety oversight. These threats were quite credible, as more than 100 candidates and elected officials were killed between 2008 and 2012 by criminal groups.
Dell (2015) studies the relationship between municipal elections and drug violence in Mexico. She finds that close municipal elections increase drug-related violence, especially if a candidate from the Partido Accion Nacional (PAN), President Calderon’s party affiliation, is involved. Dell suggests that drug-related violence is a consequence of attempts by drug cartels to control territories after newly elected PAN mayors initiate efforts to weak them.
Our methodology relies on the assumption that municipal elections are not correlated with the determinants of migration except through violence. The literature on the effect of electoral cycles on migration decisions is almost nonexistent. A few studies have looked at the political effects of migration at the municipality level. Pfutze (2012) looks at the effect of migration in promoting democratization, and Diaz-Cayeros, Magaloni, and Weingast (2003) study the effect of migration on the hegemonic dominance of oneparty authoritarian regimes. However, little is known about the effect of electoral cycles on migration flows. Brenes-Camacho (2010) looks at the effect of federal elections on migration flows from Latin America and finds that federal electoral cycles affect migrants’ economic expectations, and therefore, migration flows.
In terms of revenue, municipal governments in Mexico are heavily dependent on federal transfers. According to theWorld Bank, federal transfers amounted to about 90 percent of subnational public revenue in 2012 (World Bank 2013). This system and the discretionary distribution of federal transfers have lessened the states’ and municipalities’ incentives to raise their own revenue. Those factors suggest that elections at the federal level rather than the municipal level are likely to bring changes in economic policies and welfare programs, impact the finances of the municipality, and the incentives of individuals to migrate. In the next section we explore with more detail the relationship between elections and migration rates.
1. Elections and migration rates
In order to examine if federal, state, and municipal elections affect migration rates through a channel other than violence, we perform an analysis using data between 1999 and 2012. That period covers three federal elections, 79 state elections, and 8,998 municipal elections.10 If municipal elections affect migration rates only through violence, we should observe no relationship between municipal elections and migration rates before Calderon arrived to the presidency and a strong relationship after that.
We construct migration rates using data from the EMIF from 1999 to 2012. We also construct a set of dummy variables using federal, state, and municipal election dates. The variable Federal equals one for years in which federal elections are held and zero otherwise. Pre_Federal equals one the year before federal elections are held and zero otherwise. Post_Federal equals one for years after federal elections are held and zero otherwise. Similarly, we create the dummy variable State equal to one for years in which state elections are held, Pre_State equal to one for years before state elections are held, and Post_State equal to one for years after state elections are held. For federal and state elections, the excluded category consists of the three years between the year after an election and the year before the next election.
Finally, we create the dummy variable Municipal equal to one for years in which municipal elections are held, Post_Municipal equal to one for years after municipal elections are held, and Pre_Municipal equal to one for years before municipal elections are held. It is important to note that since municipal elections are conducted in general every three years, the variable Pre_Municipal also represents the second year in office of a three-year appointment. During the period of analysis some states had abnormally long mandates. Therefore, the excluded category is the extra year in between two elections (second year in the office) when elections are held every four years.
We estimate the following equation:
where Xmt is a vector of time-varying controls at the municipal level, αm are municipali-ty fixed effects, βt are year fixed effects, and εmt is an error term. Standard errors are robust and clustered at the state level. The set of time-varying controls at the municipal level includes the birth rate, average years of schooling, unemployment rate, logarithm of the number of IMSS-registered workers, and logarithm of the amount spent on the welfare program “Oportunidades.” We split the sample and estimate Equation 3 over two periods, before and after Calderon became president.
Column 1 of Table 3 shows results for the period 1999–2005, Column 2 for the period 2006–2012, and Column 3 for the period 2007–2012.11 We find that migration into the United States decreases during federal election years in all periods. This finding is consistent with an improvement in the economic conditions in Mexico during election years, which could occur if the incumbent government can increase spending on welfare programs, public transfers, or employment during electoral years to boost its election prospects.12 With respect to state elections, we find that the variables controlling for state electoral cycles are not statistically significant in any of the periods analyzed.
When we analyze the effect of municipal electoral cycles we find that municipal elections do not impact migration decisions before 2006. In contrast, Columns 2 and 3 show that during Calderon’s presidency, municipal election years and the first year in office are associated with significant decreases in migration. In short, municipal electoral cycles only impact migration decisions during the periods characterized by high levels of drug violence.13 This supports our identifying assumption that municipal elections do not affect migration rates through a channel other than violence.
2. Instruments
As discussed earlier, we use three instrumental variables in our analysis on the effect of drug violence on migration. The first instrumental variable, named Municipal, captures increases in violence over the election year,14 including Election Day and electoral campaigns. An increase in violence over this period would occur if criminal organizations commit violent acts with the intention of influencing the results of an election. For example, a cartel could choose to kill candidates directly or to increase violence in order to reduce voter turnout. The second instrument Post_Municipal captures increases in violence during the first year in office of the recently appointed mayors. Violence could increase if, for example, criminal groups commit violent acts with the intent to pressure newly appointed officials not to interfere with the operations of their organizations. Finally, the variable Pre_Municipal captures increases in violence during the second year in the office, which is also the year before the next election.
The instrument variable (IV) estimates recover the average impact of homicides in municipalities where the increase in violence is associated with election cycles. LATE could understate or overstate the treatment effect relative to the average, depending on how municipalities with violence associated to electoral cycles compare to other municipalities in Mexico. Importantly, we are able to demonstrate that our instruments are strong and effective at predicting drug-related violence. We thoroughly examine first stage estimates to avoid potential issues associated with weak instruments.
Effect of Electoral Cycles on the Probability of Migrating to the United States: OLS Results
B. Transit Violence
Another factor that could impact migration decisions is the violence along the roads to the United States–Mexico border. To test the effect of transit violence on the probability to migrate we construct two indices that measure the violence to which migrants are exposed. First, when they travel through their state of origin and, then, when they cross through different Mexican states.
The Index_Within_Statemt variable accounts for violence in the state of residence. This index is constructed using the distance that migrants have to travel to reach their principal mode of transportation to the United States and the violence to which they are exposed in this route. In order to use similar lag structures for local and transit violence we construct four indices of violence within state. Each one is calculated using a different homicide rate (one-year, two-year, three-year, and cumulative homicide rate).
To calculate Index_Within_Statemt , we first construct a homicide rate within state:
where Homicide_rate_Statemt –1 is the homicide rate of the state where the mu-nicipality is located, Homicide_ratemt –1 is the homicide rate of municipalitym, Population_Statem is the population of the state where municipality m is located, and Populationm is the population of municipality m.
We then multiply Homicide_rate_Within_Statemt –1 by the distance that migrants have to travel to reach their principal mode of transportation to the United States. Since different individuals from the same municipality can have different modes of transportation, we use, as a proxy, the distance that migrants have to travel from their municipality to the nearest train station. To avoid endogeneity issues we use the distance to the nearest train station on the north/south rail lines as they existed in the early 1900s.15 This leads to
The Index_Across_Statesmt variable measures the violence to which migrants are exposed to in their journey from their state of residence to the United States.We construct an index using each homicide rate.
To calculate Index_Across_Statesmt we find the shortest route from each municipality to the nearest port of entry to the United States.16 Once we identify the states that migrants from each municipality will cross, we construct weights to account for the share of the trip that occurs in each of the states crossed.
The weights are calculated as follows:
where weightmS is the weight assigned to state S that will be crossed by residents of municipality m.We divide the surface of each state (Surfaces ) by the sum of the surfaces of the S = 1,...,n states that will be crossed by residents of municipality m. Then, using the weights and the homicide rate of each state crossed, we calculate a weighted homicide rate during the trip.
Finally, we multiply the weighted homicide rate by the total distance to the border (in thousands of kilometers) from the municipality of origin to the nearest crossing point. This leads to
We include Index_Within_Statemt , Index_Across_Statesm t, and the in-teraction of both indices into Equation 1 to obtain:
V. Results: Effect of the Violence on the Outflows of Migrants
A. Local Violence
Table 4 shows OLS results. The dependent variable is the migration rate and each column shows regressions using different homicide rates as independent variables. It is important to note that our OLS results may be biased if unobserved factors jointly determine migration decisions and increases in homicide rates. Our findings show that a one-unit increase in the one-year, two-year, and three-year homicide rate increases migration rates by 0.0003, 0.0002, and 0.0001, respectively. In order to quantitatively compare these results, we calculate elasticities of migration with respect to homicides. We estimate an elasticity of migration of 0.03, 0.043, and 0.03 with respect to the oneyear, two-year, and three-year homicide rate, respectively.17 Panel B of Table 4 shows regressions by region. We find that rises in the different homicide rates increase migration rates mainly in northern and central Mexico.
Table 5 shows results using instrumental variables. A unit increase in the one-year, two-year, three-year, and cumulative homicide rates increases migration rates by 0.138, 0.0085, 0.0078, and 0.0064, respectively. The elasticity of migration increases as we increase the lag on the homicide rate. The elasticity of migration with respect to the oneyear, two-year, three-year, and cumulative homicide rate are 1.38, 1.82, 2.35, and 1.89, respectively.
The top panel of Table 5 includes first stage results. The F-statistics and underidentification tests are included. The F-statistics of the first stage in all regressions are sufficiently high (between 15.32 and 36.41). The F-statistic exceeds the Stock andYogo critical values for maximal IV relative bias at the five percent level (13.91) in all regressions. Using the Kleibergen-Paap rk LM statistic we reject the null hypothesis of underidentification in all regressions, so there is high confidence in the validity of the instruments.18
Effect of Violence in the Probability of Migrating to the United States: OLS Results 2007–2012
It is interesting to note that the coefficients of our three instruments are negative. Violence was higher during the additional year between two elections when elections were held every four years (excluded category). That year represents the second year in the office of a four-year appointment. This result is consistent with the findings of Dell (2015), who reports that drug violence increased significantly in the period following an election. Drug trafficking organizations may have more incentives to control authorities and influence policies or their implementation in their favor when the length of the mayor’s appointment is longer. The coefficients for variables Post_Municipal and Pre_Municipal (first and second year in the office, respectively) have, in general, the largest coefficients (less negative), indicating more violence during these periods. The variable Municipal has smaller coefficients (more negative), indicating that violence decreased during election years.
Effect of Violence in the Probability of Migrating to the United States: IV Results 2007–2012
The large differences in the magnitudes of OLS and IV coefficients could be a signal that the homicide rate is negatively correlated with the error term biasing downwards the OLS coefficients. While there is consensus that net effect of drug trade on the economy is negative, the presence of drug cartels in poor, rural, underdeveloped communities can have a positive impact on the economic conditions of their residents.19 Moreover, the distribution of violence has spread across the country as government crackdowns force cartels to divert their drug activity and move to different areas (Dell 2015). Drug cartels may have a preference to locate in areas with infrastructure and access to major transportation routes to the United States. Both factors may be positively correlated with the economic conditions of their residents, and negatively correlated with their probability to migrate.
Table 6 shows results of our analysis across regions. Central and northernMexico are more responsive to increases in homicide rates than southern and western Mexico. We calculate an elasticity of migration with respect to the two-year homicide rate of 3.98 for centralMexico, 3.43 for northern Mexico, and 0.7 for southern Mexico. The elasticity of migration with respect to the cumulative homicide rate is 3.86 for ventral Mexico, 2.68 for northern Mexico, and 1.69 for southern Mexico. In western Mexico, violence does not seem to impact migration flows.
The top panel of Table 6 shows F-statistics and test for weak identification and underidentification.20 The F-statistics of the first stage are, in most of our specifications, higher than ten. Therefore, we reject the null that the instrument is weak.21 Moreover, the null hypothesis of underidentification is rejected.
B. Transit Violence
To identify two different channels through which violence can impact migration, Table 7 shows regressions including the index of violence within state, index of violence across states, and their interaction.
Panel A of Table 7 shows first stage results. We obtain a low F-statistic when we use the homicide rate (Column 1).22 However, when using the two-year, three-year, and cumulative homicide rates, the F-statistics are greater than ten (Columns 2–4). In those specifications, the F-statistics exceed the Stock and Yogo critical values for maximal IV relative bias at the five percent level (13.91) two times, and at the ten percent level (9.08) once.We reject the null hypothesis of underidentification in all regressions according to the Kleibergen-Paap rk LM statistic.23
Effect of Violence in the Probability of Migrating to the United States: IV Results Regional Analysis, 2007–2012
Effect of Violence in the Probability of Migrating to the United States: IV Results Local and Transit Violence, 2007–2012
Results from Panel B show that the elasticity of migration to the homicide rate increases when we account for transit violence. The elasticity with respect to the oneyear, two-year, three-year, and cumulative homicide rate went from 1.38, 1.82, 2.35, and 1.89 to 1.84, 2.75, 3.56, and 2.24, respectively.
Table 7 also shows that the coefficient associated with the index of violence within state is statistically significant at the one percent level in Columns 2–5. The significance of the index across states improves as we increase the lag in the homicide rate used to construct it. It is significant at the five percent level when we use the three-year homicide rate and at the one percent level when we use the cumulative homicide rate.24
Figure 2 shows the effect of transit violence on migration rates for municipalities with different characteristics using coefficients estimated with the two-year homicide rate (Column 2). The horizontal axis shows variation in the index of violence across states, including the mean and median. The different series in the graph show the effect for municipalities with different levels of violence within state: 25th percentile, median, mean, and 75th percentile. According to Figure 2, a municipality with average violence observed a 2.56 percent decrease in its migration rate due to transit violence. A municipality with median violence within and across states, suffered a 1.42 percent drop in its migration rate.
Effect of Transit Violence on Migration Rates
Next, we estimate the total effect of violence—local and transit—for different municipalities. Figure 3 shows that a municipality with average violence observed an overall increase of 1.56 percent in its migration rate. The municipality with median violence within and across states observed an increase of 2.70 percent in its migration rate. These findings suggest that the migration rate in a municipality with average local violence could have potentially increased as much as 4.1 percent if violence on the roads would have not deterred migration.
We also calculate the net effect of transit and local violence over the period of analysis. Back-of-the-envelope calculations show that violence on the roads was responsible for a 1.83 percentage point decline in migration between 2007 and 2012. We calculate an overall positive effect of violence on migration flows of 1.53 percentage points. This figure represents a 50 percent increase in migration into the United States relative to the level observed in 2007.25
Effect of Transit and Local Violence on Migration Rates
VI. Exclusion Restriction and Robustness Tests
To test the soundness of our IV strategy, we provide evidence supporting the assumption that the timing of municipal elections influence migration only through their association with the recent increase in drug violence.
First, we estimate IV regressions using Equation 1 and data between 1999 and 2006. During this period, all municipal elections were conducted every three years. We use the instruments Municipal and Post_Municipal; the excluded category is the variable Pre_Municipal. Table 8 shows the results. Municipal elections do not have predictive power during the period 1999–2006. The F-statistics obtained in the first-stage are well below ten. According to the Kleibergen-Paap rk LM statistic we cannot reject the null hypothesis of underidentification. These results contrast the findings reported in Table 5 that electoral cycles are strong instruments for drug violence during Calderon’s administration (2007–2012).
Next, we estimate the reduced-form coefficients for the period before and after the beginning of the war against drug cartels to show that electoral cycles only impact migration decisions during the periods characterized by high levels of violence. The results are shown in Online Appendix Table A10. Municipal elections do not impact migration decisions before 2006. On the contrary, during Calderon’s administration, migration decreases during the election year and the first year of office relative to the levels observed during the second year in the office of mayors with three and four year appointments. These results are also consistent with our identifying assumption that municipal elections influencemigration only through the recent increase in drug violence.
We also test if municipal elections were correlated with other types of deaths during Calderon’s administration using Equation 1. We estimate IV regressions using as dependent variables a two-year accidental death rate and a two-year suicide rate (deaths per 10,000 inhabitants).26 Results are shown in Table 9. Municipal elections do not have predictive power over other types of deaths. The F-statistics obtained in the firststage regressions are well below 10, and we cannot reject the null hypothesis of underidentification using the Kleibergen-Paap statistic. These results corroborate the validity of our IV strategy; municipal elections can be used to instrument for drug violence, but no other types of deaths.
The next step is to test the effect of homicides on migration rates among males and females following Equation 1 and using as an explanatory variable the two-year homicide rate. Table 10 shows that a one-unit increase in the two-year homicide rate increases migration rates by 0.0074 among males and has no effect on the migration rates of females. This result is in line with previous studies (Giorguli and Angoa 2016; Le Goff 2016; Morrison, Schiff, and Sjoblom 2007) reporting that males and females may respond differently to migration shocks. Female migration has been found to be more likely motivated by family reunification; their decisions are frequently a response to household migration strategies.
Finally, we test the effect of violence on return migration. An increase of violence in a municipality of origin in Mexico can also impact the inflows of migrants back from the United States. An increase of violence could have increased the emotional cost of being away for migrants who leave their families back in Mexico and perceive their family members might be at risk. In that scenario, an increase in violence could trigger an increase in return migration. On the other hand, violence could have decreased the emotional cost of being away for migrants who migrate with their whole families and feel that Mexico is too risky. In that case, an increase in violence could decrease return migration.
Effect of Violence Probability of Migrating to the United States: IV Results, 1999–2006
Effect of Other Violent Deaths on the Probability of Migrating to the United States: IV Results, 2007–2012
Effect of Violence Probability of Migrating to the United States for Males and Females: IV Results, 2007–2012, Second Stage Estimation
Effect of Violence on the Probability of Returning to Mexico: IV Results 2007–2012
The EMIF has a questionnaire conducted among southward-bound migrants returning to Mexico from the United States by their own free will. This sample includes individuals visitingMexico for a short period of time and return migrants. Return migrants are individuals traveling to Mexico to settle there permanently with no intention to return to the United States.We construct return migration rates by municipality using the number of return migrants (using sample weights) as a proportion of the population aged 15–64. Data on population are from the 2000 Mexican Census.
We use Equation 4 with the return migration rate as the dependent variable. Results shown in Table 11 indicate that local and transit violence do not affect return migration. Estimated coefficients in all specifications are not statistically significant.
These findings suggest that, while violence in Mexico can affect the emotional costs of being away for migrants, there may be other more important determinants of the decision to return, such as economic conditions in the United States, seasonality of employment, or duration of job permits. These factors may weigh heavily in the return migration decision, which would explain why return migration does not seem to respond to increases in violence.
VII. Perception and Costs of Violence: Differences across Regions
In this section we study the relationship between local and transit violence and corresponding changes in the perception of safety of individuals in the municipality, state of residence, and on the roads. This evidence can be used to corroborate the existence of two different channels affecting migration decisions of individuals. Rational choice theory suggests that a deterioration in the perception of safety could increase the incentives of individuals to migrate (Edwards 2008).
We estimate the following equation:
and use as dependent variable four different measures of the perception of public safety. The first two measures account for safety in the municipality and state of residence.We use a dummy variable equal to one if individuals feel unsafe in their municipality and zero otherwise. The second variable equals one if individuals feel unsafe in their state of residence and zero otherwise. The other two measures account for safety on the roads. We use a dummy variable equal to one if individuals feel unsafe when traveling along the roads and zero otherwise. Finally, we use a dummy variable equal to one if individuals stopped traveling to other states for fear of becoming a crime victim and zero if they did not.
The explanatory variables include the two-year homicide rate, index of violence within state, index of violence across states,27 dummy variables for regions of Mexico, distance from the municipality to the United States, and distance from the municipality to the nearest train station. The vector Ximt contains a set of individual characteristics, including age, sex, labor force participation, type of job (farm worker, factory worker, owner/employer, worker without pay, and self-employed), and a dummy variable indicating if the individual has been victim of a crime. The regressions include fixed effects by year of the survey; standard errors are robust. Equation 5 differs from our previous specifications since the fixed effects by municipality are replaced with region dummy variables. This replacement occurs given the high demands that the identification of municipal fixed effects places on the data when only two years of information are available.
Table 12 shows OLSregression results. An increase in the homicide rate (a measure of local violence) has the strongest impact on the perception of safety in the municipality, followed by the perception of safety in the state. Transit violence within state, measured by the index of violence within state, has a strong impact on the perception of safety in the state, followed by the perception of safety on the roads, and in the municipality. Finally, the index of violence across states has a strong impact on the probability to stop traveling due to fear of violence.28
Perception of Public Safety and Losses Due to Crime
Even though we only have two waves of data on perceptions covering our period of analysis, this evidence is valuable since it helps to establish the relationship between local and transit violence to corresponding changes in perceptions of safety. Local violence makes individuals feel unsafe in their municipality of residence. Transit violence within state makes individuals feel unsafe in their state. Finally, transit violence across states deters individuals from traveling to other states. This evidence supports our assumption that local and transit violence are two different channels that impact the perceptions of individuals. These relationships authenticate the fear of violence, and the distinction between local and transit violence, as the causal channels through which migration is affected.
VIII. Conclusions
During President Calderon’s administration, Mexico experienced a dramatic increase in drug-related violence that had a substantial social and economic impact on Mexican families. In this study, we examine the effect of this violence on the outflows of migrants from Mexico to the United States between 2007 and 2012.
To identify effects, we use an instrumental variable approach that exploits the fact that violence in Mexico has been widely linked to municipal electoral cycles. Moreover, we identify two different channels through which violence impacts migration: local violence (violence in the municipality of residence) and transit violence (violence along the route to the Mexico–United States border).
Our findings show that violence has increased migration outflows from Mexico to the United States. We calculate an elastic response of migration rates to the homicide rate of 1.38.We also show that the elasticity of migration with respect to violence is higher among individuals from northern and central Mexico. This contrasts with what one would have expected based on underlying propensities of individuals from different regions to migrate, since individuals in these regions had much lower migration rates and smaller social networks in the United States. Rather, this pattern is perhaps best explained by the fact that economic losses due to violence were likely most severe in those regions, as they had the highest growth rates over the last decade. Assessing the effects that drug violence may have on future migration flows and stocks is crucial. The creation of migration infrastructure in regions where it is relatively low may have important spillover effects increasing migration from those regions in the future even if drug-related violence in Mexico subsides. Our findings also showthat violence on the roads had a negative impact on migration. Transit violence was responsible for a 1.83 percentage point decline in migration between 2007 and 2012. However, our results show that the impact of local violence outweighs the impact of transit violence; back-of-the-envelope calculations indicate that the net effect of the increase inviolencewas to increase migration rates by 1.53 percentage points, or 50 percent relative to the level observed in 2007.
Overall, this study identifies two channels through which drug violence affects migration decisions, demonstrates that migration decisions are sensitive to both local violence and transit violence, and finds that the net effect of the drug violence in Mexico has been to increase migration to the United States.
Acknowledgments
The authors thank Daniel Berkowitz, Daniele Coen-Pirani, Mark Hoekstra, Randy Walsh, and seminar participants at the 2014 Annual Meeting of the American Economic Association, the Seventh International/Development Economics Workshop at the Federal Reserve Bank of Atlanta, and the 18th Annual Meeting of the Latin American and Caribbean Economic Association, for helpful comments on this and earlier versions of the manuscript. The data used in this article can be obtained beginning January 2019 through January 2022 from Sandra Orozco-Aleman, Box 9580, Mississippi State, MS, 39762 (sorozco{at}business.msstate.edu)
Footnotes
1. Estimates from the authors based on data from Crawford (2015), FBI Uniform Crime Reporting Program Publications, and World Bank Statistics. Homicide and death rates describe the number of homicides and deaths over the period of analysis per 10,000 inhabitants according to the pre-conflict population.
2. Between 2003 and 2014, states from central, northern, western, and southern Mexico contributed on average with 40 percent, 25 percent, 19 percent, and 16 percent of GDP, and experienced average GDP growth rates of 5.0 percent, 3.1 percent, 2.1 percent, and 2.3 percent, respectively.
3. An alternative source of homicide data is the National Public Security Council (CNSP). They compiled data on drug-related homicides by municipality (homicides from armed confrontations between authorities and organized criminals and from confrontations between civilians with at least one of the parties being involved in the drug trade) between December 2006 and September 2011. We calculate a correlation coefficient of 0.95 between INEGI and CNSP data that demonstrates that the increase in homicides is driven by the increase in drug-related violence.We use data from INEGI since it covers a longer period of time; it allows us to construct lagged and cumulative homicide rates and conduct robustness tests analyzing other periods of time.
4. Estimations by the authors using data from the United Nations Office on Drugs and Crime.
5. Approximately 90 percent of the immigrants surveyed are undocumented. We eliminate from the sample migrants with a recent entry into the United States (within 30 days) to avoid the possibility of double-counting. The migrants eliminated represent 0.68 percent of the sample, with 88 percent of them being undocumented. They might be circular migrants or migrants caught and returned by the border patrol who reenter the United States.
6. We divide Mexico into four regions according to their geographical and migratory characteristics. Northern Mexico: Baja California Norte, Baja California Sur, Sonora, Sinaloa, Chihuahua, Coahuila, Nuevo Leon, and Tamaulipas. Western Mexico: Aguascalientes, Colima, Durango, Guanajuato, Jalisco, Michoacan, Nayarit, San Luis Potosi, and Zacatecas. Central Mexico: Morelos, Queretaro, Tlaxcala, Puebla, Hidalgo, D.F., Estado de Mexico, and Veracruz. Southern Mexico: Campeche, Chiapas, Guerrero, Oaxaca, Quintana Roo, Tabasco, and Yucatan.
7. Online appendixes can be found at http://jhr.uwpress.org/.
8. Between 2007 and 2012 there were 4,175 municipal elections. During that period there were 14 extraordinary elections. An extraordinary election occurs when an election is tied or voided.
9. While most states have municipal elections every three years, some had elections every four years (38 municipalities of Coahuila had elections in 2005 and 2009, 76 municipalities of Guerrero had elections in 2008 and 2012, and 112 municipalities from Michoacán had elections in 2007 and 2011), and others every two years (116 municipalities from Chiapas had elections in 2010 and 2012, eight municipalities from Quintana Roo had elections in 2008 and 2010, and 67 municipalities from Yucatán had elections in 2010 and 2012). In some states, longer or shorter mandates were set to match municipal elections with other state or federal elections. Those exceptions were set in the State’s Constitution years before the war against drug trafficking organizations began.
10. Out of the 2,456 municipalities accounted for in the 2010 Mexican Census, we exclude 424 municipalities in Oaxaca since most of them elect their leaders through customary law in nonpartisan elections, and 17 municipalities for which data on migration rates were unavailable.
11. Since Calderon arrived to the presidency on December 1, 2006, we analyze the period 2006–2012 and 2007–2012.
12. Schuknecht (1996) finds that welfare policies are effective in determining voter’s preferences, and Diaz-Cayeros et al. (2003) find that the hegemonic dominance of a one-party authoritarian regime is sustained by a credible threat of withdrawing financial resources from defectors.
13. We also evaluate if municipal resources respond to electoral cycles.We use Equation 3, but the dependent variable is now the logarithm of the municipality income and the logarithm of federal transfers. Results are shown in online Appendix Table A3. Municipal income and federal transfers to municipalities are not affected by municipal electoral cycles before 2006. However, during Calderon’s presidency, we observed higher income and federal transfers during the second year in the office of mayors with three- and four-year appointments. This evidence supports the assumption that municipal elections do not impact the economic conditions of the municipality through a channel other than violence.
14. Elections in Mexico are restricted by law to last no more than five months; therefore, Election Day and electoral campaigns are likely to be captured by the variable Municipal.
15. Distance from each municipality to the nearest train station on the north/south rail lines is described in Woodruff and Zenteno (2007) and provided by ChristopherWoodruff at: http://www2.warwick.ac.uk/fac/soc/ economics/staff/academic/woodruff/data/mexico_migration (accessed October 20, 2017).
16. We use Google maps to calculate the shortest distance from each municipality to the nearest port of entry to the United States. We use road travel and allow for driving on highways and toll routes.
17. We calculate elasticities using regression coefficients and Table 2. The percentage change in migration is the coefficient divided by the average migration rate of the period (0.0003/0.015 = 0.02). The percentage change in the one-year homicide rate is one divided by the average homicide rate (1/1.498 = 0.667). The elasticity is 0.03 (0.02/0.667).
18. We run the regressions shown in Table 5 using different years in the election cycle as an excluded variable and obtain similar results. Furthermore, if we exclude any of the instruments used in these regressions we obtain lower F-statistics and higher standard errors.
19. Benefits include employment opportunities and large cash inflows (around 2.5 billion dollars a year), which translate into more consumption and investment in the region (Ríos 2011).
20. Online appendix Tables A4–A7 show more detailed information regarding first stage regression results.
21. Since there are no Stock and Yogo critical values for maximal IV relative bias for four endogenous regressors, we use F-statistics to test for weak instruments.
22. The instrument Pre_Municipal is a weak instrument when we use the homicide rate (Column 1), and for that reason, it is not included in the first stage regression.
23. To verify that the indices within and across states are exogenous, we run endogeneity tests using the C statistic, defined as the difference of two Sargan–Hansen statistics: one for the equation where the suspect regressors (indices within and across states) are treated as endogenous and one for the equation where they are treated as exogenous. Under the null hypothesis the specified endogenous regressors can actually be treated as exogenous.We test the regressions from Table 7, and, in all specifications, we cannot reject the null hypothesis that the indices within and across states should be treated as exogenous.
24. Similar results are obtained when we include the index within state, index across states, and their interaction into Equation 2. Results are reported in Online Appendix Table A7.
25. Online Appendix Table A9 shows the average and change in migration rates during the period of analysis, two-year homicide rate, and index within and across states used to calculate the total effect of violence. The change in the index across states between 2007 and 2012 was 3.719.We multiply the change in the index by its marginal effect and obtain a decrease of 0.00425 [3.719*(-0.0020 + 0.0022* 0.3914) = –0.00425]. The effect of the change in the index within state is -0.0140. Finally, the effect of the change in the local violence is 0.0336 (0.0128*2.616). The total effect of violence on the migration rate during the period of analysis is an increase of 1.53 percentage points (0.0336 – 0.01401 – 0.00425 = 0.0153).
26. We only report results using the two-year suicide and accidental death rate; however, similar results were found when we use the one-year and three-year lagged rates.
27. Both indices were constructed using the two-year homicide rate.
28. Online Appendix Table A11 shows similar findings when we analyze the effect of local and transit violence on the perception of safety by region in Mexico.
- Received February 2015.
- Accepted February 2017.
References
In this issue
Jump to section
- Article
- Abstract
- I. Motivation
- II. Theoretical Framework and Existing Literature
- III. Data
- IV. Empirical Specification
- V. Results: Effect of the Violence on the Outflows of Migrants
- VI. Exclusion Restriction and Robustness Tests
- VII. Perception and Costs of Violence: Differences across Regions
- VIII. Conclusions
- Acknowledgments
- Footnotes
- References
- Figures & Data
- Supplemental
- Info & Metrics
- References