Violence and Children’s Education: Evidence From Administrative Data

This paper exploits the sharp escalation of violence in Colombia in the 1980s associated with the emergence of drug cartels to provide novel evidence on the long-run effects of violence exposure throughout the life-course, on children’s educational attainment and academic achievement using administrative data. I find that, a higher homicide rate in early-childhood is associated with a higher probability of school dropout and conditional on completing high school, lower scores on a national end-of-high school exam. Results are robust to several falsification tests, and analyses of potential sources of selection bias. I provide supportive evidence that changes in fetal, child, and adolescent health outcomes are important potential mechanisms.


Introduction
With recent demonstrations of violence and terrorism across the world, more attention has been focused on the adverse effects of violence. Unicef (2016) reported that one out of nine children is born and raised in countries affected by repeated cycles of criminal violence. While there is growing evidence documenting the contemporaneous effects of violence, there is little empirical evidence on the long-term cumulative impacts of these shocks and how they affect different dimensions of human capital. This gap in the literature is particularly important given the large body of research showing that adult outcomes are shaped by conditions experienced over the life-cycle (Almond, Currie and Duque, 2018;Barker, 1998;Cunha and Heckman, 2007;Bharadwaj and Vogl, 2016).
This paper provides new evidence on the cumulative impacts of violence exposure over the lifecourse on long-term educational attainment and academic achievement using multiple sources of administrative data for the universe of students in public schools in Colombia. I exploit a large and unprecedented rise in homicide rates during the 1980s, which was associated with the surge in drug trafficking (see Figure 1) triggered by a shift in the international demand for cocaine. Over the course of one decade, the homicide rate went from 0.3 homicides per 1,000 inhabitants to more than 0.8. 1 The escalation of crime was accompanied by a sharp spatial concentration of violence, with the urban, wealthiest, and more economically developed regions experiencing a dramatic ten-fold increase in homicide rates and with little change in violence in other places (see Figures 2 and 3). I measure violence exposure using homicide rates, a commonly employed violence indicator that is highly correlated with the occurrence of other violent actions such as robberies, attacks, and explosions (Sánchez and Nunez, 2007), and is often more reliable than other measures of crime that are not always reported by the police (Levitt and Rubio, 2000;Soares, 2004).
Linking students in public schools to the universe of the end-of-high school exam takers and to the universe of poor households in Colombia, allows me to observe long-term outcomes linked to location and exact date of birth for almost 430,000 individuals born in Colombia in the 1980s.
The richness of the micro-and violence data enables me to measure violence exposures from in-utero up to adolescence. I focus on educational outcomes such as high school completion and end-of-high school national standardized test scores (as well as the separate math and language components of the exam) that broadly capture persistent health, cognitive, and non-cognitive impacts caused by early-life exposure to violence, as well as any potential compensatory and reinforcing parental or school investments.
To estimate the effects of violence on children's education, the paper exploits the temporal and geographic variation in homicide rates across municipalities-months-years using a difference-indifference framework that compares the outcomes of children born in different regions and years, thereby differentially exposed to the shock. The administrative data also allows me to identify a child's siblings, so as a robustness check, this paper also conducts a model that controls for a family fixed-effect that help disentangle the effect of violence from other types of socioeconomic disadvantage that may correlate with educational outcomes.
Results suggest that, an increase of a standard deviation in homicide rates in childhoodwhich is equivalent to an increase in violence from the 25th percentile to the 75th percentile in violence over the 1980s -is associated with an 15.0% increase in the probability of dropping out of high school and conditional on finishing high school, a 0.5 standard deviation (SD) decline in Colombia's end-of-high school exam (Icfes). Analyses by subject show that the math and language components of the Icfes exam fall by 0.20 SD and 0.25 SD, respectively. These findings are robust to an extensive set of falsification tests and analyses of potential sources of selection bias such as migration, fertility, and survival. 2 Using household survey data, I then study potential mechanisms. Motivated by a large literature on child development and human capital formation (for recent summaries on this research see Almond and Currie (2011) and Almond et al. (2018)), I focus on changes in child's health and nutrition, as well as on parental investments before and after birth. I find that higher homicide rates negatively affect health at birth and early childhood health outcomes: a one SD increase in homicide rates during pregnancy leads to a higher probability of low birth weight (a 60% with respect to the outcome mean) and a significant decline in child's height-for-age Z-scores (HAZ) (of 0.56 SD), a widely used indicator of child's nutritional and health status.
2 As falsification tests, I examine whether violence two and three years prior to conception have an effect on children's educational outcomes. As potential sources of selection bias, I conduct the following empirical tests (which I describe in more detail in Section 6): i) Selective fertility: I test the association between violence and a mother's total fertility after a (focal) child (i.e., the number of children she has after her first child born in the 1980s) and the association between violence and the timing of fertility (i.e., the time between two subsequent births); ii) Selective survival: I examine the proportion of movers (those who reside in a different municipality to where they were born) and estimate whether the effects of violence in the full sample differ to those in the sample of non-movers; iii) Selective survival: I analyze whether changes in violence are associated with changes in the sex ratio or in the cohort size. Several studies have shown that boys are biologically weaker and more susceptible to diseases and premature death than girls (Naeye et al., 1971;Waldron, 1985), and that they are more vulnerable to environmental factors in early-life (Pongou, 2013). Therefore, focusing on these two measures could be informative to understand potential demographic imbalances linked to violence.
The decline in child's height persists on to adulthood. Children exposed to high violence while in-utero and in childhood are significally shorter as adults, which is consistent with the idea that cognitive ability and height share similar inputs in the production function (Case et al., 2005;Case and Paxson, 2008;Currie and Vogl, 2013;Vogl, 2014) and with previous research showing that adults' height is mostly explained by conditions experienced in the first years of life (Victora et al., 2010). In terms of family inputs, results show some suggestive evidence that mothers are less likely to adopt compensatory behaviors such as breastfeed their child for at least 6-months conditional on in-utero violence exposure, and children are more likely to enter first grade "over-age" and to repeat a school grade. Results also show that children exposed to high homicide rates are more likely to live in households experiencing domestic violence and to witness this type of violence themselves. The findings show little evidence that violence affects the probability that a mother receives prenatal care or gets tetanus vaccination while pregnant; in fact, mothers are more likely to give birth at a clinic or hospital (as opposed to at home), suggesting that changes in access to health care might not be the primary mechanism explaining long-term declines in education.
This paper contributes to the body of research on the effects of violence on individual outcomes through the use of linked administrative data. This novel resource, which is usually unavailable in studies focusing on developing countries, allows me to explore the long-run effects of violence on students' learning outcomes as measured by achievement test scores. Focusing on test scores is important because these measures not only capture inputs into the production of human capital, but are also a proxy for an individual's future socioeconomic success (Currie and Thomas, 2001;Ebenstein et al., 2016). In that sense, my results complement a body of research on the short-and medium-term impacts of violence on students' learning (Bruck, Di Maio and Miaari, 2014;Monteiro and Rocha, 2017;Rodriguez and Sánchez, 2012;Haugan and Santos, 2016). This paper is also related to Leon (2012) and Galdo (2013) that provide evidence on the early-life exposures to Peru's civil conflict on education and wages, respectively. In addition to the use of richer data and more comprehensive outcomes, my paper adds to this literature by examining a context where the dynamics of violence was a more urban phenomenon resulting from the illegal activity of drug cartels rather than from guerrila groups, thus potentially informing future trends in countries experiencing similar types of shocks associated with drug trafficking and related organized crime (e.g., Mexico or countries in Central America). 3 Lastly, the findings of this paper also help inform the literature on civil wars or armed conflicts and its effects on school enrollment, educational attainment, and child's health, through changes in economic losses, (destruction of) infrastructure, teacher absence, and acess to health care (Akbulut-Yuksel, 2014;Akresh and De Walque, 2011;Chamarbagwala and Morán, 2011;Leon, 2012;Shemyakina, 2011;Galdo, 2013;Akresh, Lucchetti and Thirumurthy, 2012). Some additional differences between these studies and the present paper are that: i) these studies tend to focus on more disruptive forms of violence, which are often associated with substantial economic and political turmoil, and ii), they examine effects on a population that, even in the absence of violence, would start from a lower baseline (e.g., African children or indigenous Guatemalan children).
The remainder of the paper is organized as follows. Section 2 briefly describes the historical elements associated with the rise of violence in Colombia in the 1980s. Sections 3 and 4 present the different sources of microdata and the empirical strategy, and Section 5 shows the findings of the effects of violence on education using the difference-in-difference specication as well as the family fixed effects model, and provides some suggestive evidence on potential mechanisms.
Lastly, Section 6 performs some robustness tests to show that the results can be interpreted as lower-bound estimates of the impact of violence on education and Section 7 concludes.

Background for Empirical Strategy
In this section, I briefly describe the historical elements associated with the rise of violence in Colombia in the 1980s, which was closely related to the rise in drug trafficking from Colombia to the U.S. (Castillo, 1987). 4 Following previous studies, police reports, and newspaper articles, I highlight three key factors that contributed to this rise. capital, and labor market outcomes, respectively. My paper also contributes to this research by showing the persistent effects of violence on important economic outcomes of young adults. 4 Of note is that this paper focuses on the spike in homicide rates in the 1980s and does not consider the intensification of Colombia's armed conflict that occurred in the second half of the 1990s. While the armed conflict was a major shock to the economy, the cohorts of interest in this paper were old enough not to be affected by this shock in their early stages. Recent studies have examined the impacts of the armed conflict in the late 1990s and early 2000s on human capital outcomes birth or in early childhood (e.g. Camacho, 2008;Duque, 2017) or in the primary and secondary school years (Rodriguez and Sánchez, 2012;Barrera and Ibanez, 2004;Ibanez and Moya, 2010;Calderon et al., 2011;Fernandez et al., 2014;Gerardino, 2014) and more recently, some studies have provided evidence on the effects of the origins of the armed conflict in the late 1940s, a period called La Violencia, on long-term outcomes (Fergusson et al., 2015). For instance, one of the most devastating consequence of the conflict was forced displacement. Appendix Figure 3 shows that the number of forced migrants in Colombigaldoa only emerged until the late 1990s when the guerrillas and paramilitary groups became powerful terrorist organizations.
First, there was a dramatic increase in the international demand for cocaine in the early-1980s (United Nations, 2012). This increase coincided with the introduction of crack cocaine in the U.S. markets -a cheap, addictive, and potent form of cocaine-, which represented an important technological change that lowered the price of illicit substances (i.e., cocaine) and expanded the market to a wider range of consumers (Fryer et al., 2013). 5 Second, following the shift in the demand for cocaine (mainly from the U.S.), there was a rapid response in the supply of cocaine. Colombian traffickers of cannabis who by the late 1970s and early 1980s were smuggling small quantities of marijuana to the U.S., found enormous economic incentives in the "cocaine boom" of the mid 1980s (United Nations, 2012;Castillo, 1987;Thoumi, 2002). Three factors were crucial to understand the rapid transition of Colombian marijuana traffickers to the cocaine industry: i) they already knew potential transportation routes to move drugs from South America to the U.S.; ii) they already had access to important distribution networks in some of the largest U.S. cities (e.g., Miami, New York, and Los Angeles); and iii) they were strategically close to coca-leaf producing markets such as Bolivia and Peru.
Colombia only became a major coca leaf producer by the end of the 1980s (Sánchez, 2007), so in the early 1980s, these trafficking organizations outsourced the coca paste from neighboring Andean countries that had historically produced it by its indigenous groups. The paste was later processed into pure cocaine in local laboratories and then transported to the U.S. (the main destination) and Europe (Castillo, 1987;Thoumi, 2002). The business grew so fast that by 1991, more than 80 percent of the cocaine that reached the United States had been produced in Colombia (Borrell, 1988;Gaviria, 2000;The Economist, 1994;El Espectador, 1986). Both the enormous profits derived from the drug industry and the attempts made by drug organizations to enforce their 'property rights', converted these small traffickers into powerful and violent drug cartels. One example was the infamous Medellin cartel, led by the drug lord Pablo Escobar.
A third factor associated with the rise of violence in the 1980s was the rapid transfer of criminal knowledge and criminal technology from drug traffickers to local criminals (Gaviria, 2000). 6 In sum, the combination of these factors with an increasing tension between drug cartels and the Colombian authorities 7 , resulted in the massive rise in homicide rates observed in the 1980s (see Figure 1). 8 9 One noteworthy characteristic of the rise of violence during this period was that not all regions in the country experienced the same intensity in violence (see Figure 2). For instance, while Medellin and Cali, the country's second and third most important cities, faced an increase in homicide rates of 700% (going from 0.5 to 4 homicides per 1,000 inhabitants in  and 300% (going from 0.3 to 1.2 homicides per 1,000 inhabitants in 1980-91), respectively, other important cities such as Tunja or Neiva experienced little or no change in violence. The concentration of homicide rates in large and wealthy areas was primarily explained by the fact that the cocaine industry was headquartered in these areas (Gaviria, 2000) (see Figure 2).
In this paper, I exploit the monthly-year-municipality level variation in homicide rates associated with the drug boom of the 1980s 10 to estimate the effects of violence on children's future education using a difference-in-difference framework, which utilizes the full sample of children in the administrative data. As an alternative specification, I exploit the fact that the data enables me to identify sets of siblings which I use to estimate a family fixed effects model, and thereby provide further evidence on the causal impacts of the shock.

Microdata
The richness of the data is one of the major strengths of this study. I use several sources of large-scale administrative micro data for the analysis: the universe of students in Colombia's public-school system, the universe of Icfes exam takers, and the 2005 "universe of the poor". I describe each of these datasets below.
7 That included a bloody campaign of bombings and the killing of some of Colombia's leading figures (Gaviria, 2000). 8 Of note is that drug use rates in Colombia during the 1980s were relatively low to have contributed to the high homicide rates in this time period (Direccion Nacional de Estupefacientes 1993, 1996). 9 And while victims included many members of the police force who were combating the drug cartels, the vast majority were civilians not involved in the drug-illegal activity. 10 The identifying variation in the paper comes from the increase in homicide rates due to the spike in the cocaine demand during the 1980s. However, the main cocaine cartels were born out of former marihuana cartels, and this business had been increasing during the 1970s. Unfortunately, because the homicide data at the monthlyyear-municipality level is only available since 1979, I am not able to examine differences in violence in the years before.

The Universe of Students in Colombia's Public Schools
These data are the core database of the Ministry of Education that reports school progression records at the individual-year levels. This dataset began with the 'Resolution 166' of 2004 11 that mandated the Ministry to collect and report detailed information on the school progression of all students enrolled in the public-school system, starting in the first year a child entered a public school (e.g., first grade) up to high school graduation (or dropout). 12 In this paper, I use the universe of students in public schools from 2005 to 2012. A unique advantage of using these data, is that it is the only data source that includes information on the exact municipality of birth for each student.

The End-of-High School Exam: the Icfes
The Icfes 13 is the national high school exit exam provided by the Institute for the Promotion of Higher Education, the former acronym for the agency that administers the exam. It is taken by high school seniors regardless of whether they intend to apply to college, and it includes separate tests on math, Spanish, social studies, sciences, and an elective subject. 14 For those who transition onto college, the Icfes score determines college and major entrance. I use information on all students who took this exam from 2000 to 2009 (approximately one million observations).

The "Universe of the Poor": the Sisben
The Sisben is a proxy means test index used as a targeting mechanism for social programs in Colombia. It serves as an indicator of households economic well-being and is based on a cross-section sample of Colombian low-income households. The Sisben is usually referred to as the "census of the poor" or the "universe of the poor". This dataset include demographic and socioeconomic information on 33 million people, or 60% of the total population. 15 For the case of this study, the Sisben provides family background information on all students. While linking the administrative data to the Sisben data may restrict the analysis to two thirds of the 11 More information on this resolution is found here: http://www.mineducacion.gov.co/1759/w3-article-163147.html. 12 The school cycle is divided into three categories: basic primary (grades one through five), basic secondary (grades six through nine), and middle secondary (grades ten and eleven). After finishing 11 th th grade, students can enroll in college. Students usually start school at ages 5 to 7, and the minimum mandatory school level for students is 9 th grade, a period referred to as basic education.

13
Since 2009, the Icfes exam was renamed to Saber 11. 14 More than 90% of high school graduates take the Icfes exam. country's population, it is necessary to do so in order to be able to identify a student's family characteristics (e.g., mother's education). 16 17 I use restricted information on students' and their parents' full names (first and middle names and fathers' and mothers' maiden names), birth dates (day, month, year), and national identifications (type of ID document and number) to link individuals (and their families) across datasets.

Sample of Interest
I focus on individuals between 17 and 24 years of age (i.e., those who were born in the 1980s in Colombia). I restrict the sample to these ages because these are the oldest cohorts in my sample and for whom I can observe their high school completion as well as their test scores. 18 I also restrict the sample to individuals with exact date and municipality of birth. 19 The sample of interest in this paper includes 426,831 individuals (of which 262,123 are siblings).

Outcome Variables
The following list describes the outcomes of interest: (1) High school graduation, a dummy variable that takes the value of one when an individual has finished 11 th grade and zero otherwise.
(2) End of high school test score (Icfes score), a continuous variable that ranges from 0 to 100 indicating the Icfes test score. The icfes includes separate modules on math, language (Spanish), and other subjects.
Appendix Table 1 shows summary statistics on the sample of young adults and their families in the administrative data and as a point of reference, I also show means for cohorts born in the same years observed in the population Census of 2005. Mothers in my sample are on average 45 years of age and the vast majority has primary education or less. Almost 40% are married, families have 5.7 members, 75% live in urban areas, and 73% have access to water and sewage. Comparing these statistics to those drawn from the Census (column 2) reveal that families observed in the administrative data are a more disadvantaged group, which 16 The R-166 nor the Icfes data include information on family background.

17
Because the sample analyzed in the paper corresponds to the sample of people included in the SISBENthe 60% poorer population in Colombia-, I empirically test whether violence is related to a household's Sisben score. I find little evidence that increases in violence are associated with changes in a family's Sisben score. This result is robust to controlling by temporal and geographic characteristics as well as including with and not (including) family-specific controls. (Results not shown but available upon request).

18
While the administrative data include information on children and young adults up to 2012, the oldest individual observed in these data is 24 years of age.

19
I further restrict the sample by using the first and second years prior to conception (the homicide rate data is available at the municipality-year-month level since 1979) as falsification tests to examine the validity of my empirical strategy.
is unsurprising given the fact that the administrative sample is representative of Colombia's low-income population. The bottom of Appendix Table 1 shows outcome means. Only 42% graduate from high school (compared to 50% in the Census), and the average Icfes score is 43.0 with standard deviation 4.3 (the average math and language Icfes scores are 44.4 and 46.2, with standard deviations 7.3 and 7.0, repectively). 20

Data on Violence: Homicide Rates
Data on violence come from the National Police Department that provide total homicides that occurred in each of the 1,100 municipalities since 1979 at the monthly-year levels. 21 I construct measures of homicide rate that are defined as the total number homicides per thousand inhabitants in a given municipality and in a given year dividing this number by the total population and multiplying it by a factor of 1,000 inhabitants. These measures are linked to the administrative data at the birth municipality, birth month-year level.

Exposure to Violence During Critical Stages of Human Capital Formation
Following previous research in developmental psychology, epidemiology, and more recently in economics, focused on sensitive periods for skill formation (Gluckman, 2005;Heckman, 2007Heckman, , 2008Knudsen, Heckman, Cameron and Shonkoff, 2009;Thompson and Nelson, 2001), I focus on violence exposure over the life-course: from the prenatal period and up to age 15 (up until when basic education is mandatory in Colombia). Table 1 shows descriptive information on these violence measures. On average, individuals were exposed to 0.5 homicides per 1,000 inhabitants in each year from birth up to age 15 with a standard deviation of 0.7. It is noteworthy that the average child in the sample is exposed to more violence as they get older (see Figure   1).

Methods
I estimate the effects of violence over the life-course on educational outcomes at ages 17-24 using the following difference-in-difference framework: where the subscript i refers to an individual, j municipality of birth, and m and t to the month and year of birth. The variable Y denotes the educational outcome. V iolence jmt represent the level of violence (homicide rate) to which a child was exposed in each year of life, from age -3 up to age 15. 22 X includes a set of child characteristics such as sex, birth order, age dummies (17, 18-19, 20-22, 23+), 23 as well as dummies for mother's age in years (less than 35, 36-45, 44-55, 55+), education (no education, completed primary or less, HS or less, more than HS), marital status (married, cohabiting, single, other), whether the household lives in the urban sector, has access to piped water and sewage, and household size. Models on the Icfes exam also include dummies for the age and the calendar year at which a student took the test, and for his/her school schedule (full time, morning, evening, night, or weekend).
The terms α j , α t , and α m are fixed effects at the municipality, year, and month of a child's birth. The geographic fixed effects help absorb time-invariant differences at the municipality level while the time fixed effects absorb factors that vary over time but are invariant to the municipalities. For example, α j helps account for constant differences in poverty level across municipalities. The term θ j * t represents department 24 -specific linear time trends that help to control for differences in economic development across places that change over time (e.g., investments in health services) and that could potentially affect human capital investments.
These trends also allow me to account for differential trends in education across municipalities over the time period of analysis. ε is the random error term. Errors are clustered at the municipality level to account for within-municipality serial correlation in the observations. The key coefficients of interest are {β r } r∈{−3,−2,−1,...,15} , which describe the impacts of violence in each year of a child's life on his/her future education.
Alternative Identification Strategy. While the main identification strategy takes advantage of the large administrative sample and exploits the substantial geographic and temporal variation in violence, I also estimate a second regression model that controls for family fixed effects, leveraging variation in violence exposure across siblings. This model allows me to control for observed and unobserved time-invariant characteristics of the mother (and family), which may be correlated with both the probability of residing in a municipality with high violence and with accumulating low levels of education. For instance, if a family belongs to a demographic group that is likely to be particularly impacted by violence, the family may also be less likely to invest in their child's health and education. Hence, the identification in this framework, would 22 The homicide rate at ages -3 and -2 is used as a falsification test. 23 Including other potential individual controls such as birth order fixed-effects provides similar empirical findings. 24 Departments are much larger geographic units than municipalities -i.e., while there are around 1,100 municipalities in the country, there are only 33 departments. come from differences in violence exposure between siblings, from the prenatal period up to age 15.
Equation 2 describes the family fixed effects model. The only covariates included in this equation are child's sex and age in year dummies, birth order, and mother's age (in matrixX), municipality, year, and month of a child's birth dummies, and department-specific linear time trends as defined above. The term µ f indexes families.
Because the mother fixed-effects strategy focuses on the sample of families with more than one child born in the priod of interest, it is clear that the sample used here (60% of the overall sample) is much smaller than that used in the main difference-in-difference specification. For this reason, the mother fixed effect model serves as a robustness check for the main set of results.

Household Characteristics and Violence Exposure
Before estimating the long-run effects of violence on education, I examine changes in household sorting into municipalities more or less affected by homicide rates, which could bias the effects of violence on the outcome. I do so by relating a household's head observable characteristics and the shock, controlling for geographic and temporal fixed effects (as in Equation 1). A significant coefficient would imply that changes in violence are associated with chnages in family characteristics. Table 1 shows little evidence that families self-select into violence-affected regions based on their age, education, marital status, household size, or access to water or sewage conditional on accounting for fixed temporal and geographic characteristics. 25 If these associations are indicative of sorting across unobserved characteristics, the fact that there seems to be little evidence og geographic sorting helps validate the identification assumption, which is that changes in violence over the years-months and across municipalities are uncorrelated to changes in other inputs in the production of human capital. Of course, there may be changes in sorting within municipalities that I am unable to observe in the administrative data, for instance, that better off families are more likely to locate in less-violent neighborhoods. To 25 I select these observable characteristics given their importance and availability in the administrative data. I do not control for household income given that many households refuse to provide information on income. Race/ethnicity is not available in the data.
assess the extent to which this potential source of selection bias affects my estimates, in Section 5.4 I compare the estimates obtained from the difference-in-difference model to those in the family fixed effects specification, which by definition, help remove this potential sorting. Figure 4 shows the effects of violence on the probability of completing high-school. Recent research suggests that early exposure to adverse shocks is worse for human capital than later shocks (Heckman, 2007;Almond and Currie, 2011), however, neither economics nor medicine offers sharp definitions about when these early stages end, except for the specific case of the in utero period. I find that exposure to high homicide rates from birth up to age 6 significantly reduces high school (HS) graduation, while exposures at later stages do not seem to have an effect on the outcome. The magnitude of these early-life exposures could be interpreted as follows: an increase of a standard deviation (SD) in homicide rates at the year of birth-which is equivalent to an increase in violence from the 25th percentile to the 75th percentile during the 1980s -reduces the outcome by 2.5%, 26 and if we consider that violence systematically increased during the period of interest (as shown in Figure 1 and in Table 1), the average child observed in the data is likely to have been exposed to a persistent level of violence (i.e., for approximately 6 years) that could have lead to an overall decline in the probability of finishing high school of 15.0%.

High School Graduation
The figure also shows little evidence on differential trends prior to conception, which is given by the null effect of homicide rates two and three years prior to birth, providing support for the identification strategy. Figure 5 presents results on the Icfes test score. I find that exposure to high homicide rates from in-utero up to age 7 leads to significant declines on academic achievement at age 18. For each year of exposure to a one SD increase in homicide rates, the Icfes falls by 0.42 points (i.e., coeff*1SD in hom rates= -0.6*0.7), which is equivalent to a 0.10 SD decline.

Educational Achievement
for the case of Peru's civil conflict, provided the first evidence on long-term impacts of violence on education, which indicated that, individuals who were exposed since prior to birth, experienced a significant decline in years of education of at least 0.3. 27 A related study by Caudillo and Torche (2014) also showed that the increase in Mexico's crime rate was associated with an increase in the probability of children's school grade failure in the short term. 28 While Colombia has suffered from a long history of armed conflict, the intensification of the conflict is a relatively recent phenomenon (i.e., mid-1990s), which did not directly impact the early stages (the period from in-utero up to adolescence) of the cohorts studied here. For instance, Appendix Figure 3 shows that the number of forced migrants in Colombia, a particularly devastating consequence of the conflict, only emerged until the late 1990s when the guerrillas and paramilitary groups became powerful terrorist organizations.

Mother Fixed-Effects
In this section I present estimates obtained using the mother fixed effects model. Figures   8-12 show that controlling for a mother's observed and unobserved time-invariant characteristic leads to substantially similar coefficients on children's educational outcomes to those presented earlier. For instance, while an increase of a SD in homicide rates at the year of birth reduces high school completion by 2.5% in the main DD specification (discussed above in Section 5.2), in the mother FE model I find that this effect is 3.3%. A similar pattern is observed on the Icfes score (and on the math and language tests): in the DD model, exposure to violence at age 2 is associated with a 0.42 points decline in the Icfes, while in the mother FE this effect is 0.39 points. Both estimates are statistically significant at the 0.95 level. The mother FE results also show that violence has little effects on children's outcomes after age 8.
I conclude that, although the mother fixed effects framework might raise some concerns in terms of sample selection due to only considering families with at least two siblings born in the period of interest, the effects of violence obtained from this model are consistent to those obtained in the DD model using the full sample of children, thereby providing evidence that the violence shock is orthogonal to many observed and unobserved family characteristics (or characteristics of the environment), conditional on accounting for all the controls described in equation 1.

Potential Mechanisms
In this section, I discuss some of the potential pathways through which violence could affect child development. First, violence is a major source of stress and as such, it could affect mothers and children's health and health behaviors through chronic stress, anxiety, or changes in nutrition. Declines in nutrition and chronic stress in the first years of a child's life can be particularly detrimental for fetal and newborn health and cognitive outcomes through changes in the immune and behavioral systems that may lead to permanent alterations in the body's systems (Dunckel-Schetter, 2011;Gluckman, 2005) or through impaired development of the brain, thereby diminishing the mental skills of infants (Huizink, Robles de Medina, Mulder, Visser and Buitelaar, 2003). Some studies have associated stress in early life with lower schooling attainment and verbal IQ scores, and with higher incidence of chronic health conditions at age 7 (Aizer, Stroud and Buka, 2016) and later in life (Thompson, 2012). Researchers have also identified that child's height (a key measure of child development in developing countries) can be particularly sensitive to shocks occurring during the first 1000 days of life (Stein and Lumey, 2000;Victora, de Onis, Hallal, Blossner and Shrimpton, 2010), of which violence is a particularly harmful one (Grantham-McGregor et al., 2007;Akresh et al., 2012;Duque, 2017).
Second, stress (due to violence) may compromise the family environment by affecting parental mental health and family relationships, weakening parenting quality that in turn may hinder human capital development (Campbell, 1991;Repetti, Taylor and Seeman, 2002). Sharkey et al. (2012), for example, found that local violence is positively associated with higher parental distress, suggesting that parental responses may be a likely pathway by which local violence affects young children. Lastly, to the extent that violence stems from terrorist activity it could also be related to the amount and quality of resources in the local community (supply-side mechanisms). 29 29 A number of studies have provided evidence on potential supply-side mechanisms. Leon (2012) showed that attacks against teachers in conflict-affected areas during Peru's political conflict decreased educational attainment; Rodriguez and Sánchez (2012) showed that negative economic shocks and lower school quality due to violence increased school dropout and child labor in Colombia; Akbulut-Yuksel (2014) also found that school-facility destruction and teacher absence in WWII Germany accounted for declines in education, and malnutrition and destruction of health services worsened the health outcomes of cohorts exposed to the war. Minoiu and Shemyakina (2014) and Justino and Verwimp (2013) found that household economic losses helped explain declines in children's height during Cote d'Ivoire's civil conflict and in Rwanda's genocide, respectively. Moreover, in a related literature, Hernández (2016) found that violence acted as a key mechanism limiting the positive spillovers of economic booms on educational attainment in Colombia.
I attempt to provide suggestive empirical evidence on some of these potential mechanisms using auxiliary data. 30 I use the National Health and Nutrition Survey (ENSIN) 2010 31 that includes extensive information on health and nutritional outcomes for 25,544 individuals aged 18-30 and the Demography and Health Survey (DHS) years 1986 and 1990, which provide rich indicators on fetal and child health. I limit the sample to families with children born in the 1980s. As outcomes of interest in the DHS, I consider birth weight, low birth weight, whether a birth is premature, and HAZ up to age 5; measures of maternal investments during pregnancy include whether a mother received prenatal care, whether the birth occurred at a hospital, breastfeeding duration (months), and whether a mother received tetanus vaccination in pregnancy; and in the ENSIN, I focus on young adults' height using the World Health Organization (WHO) child growth charts for appropriate reference groups. Table 3 shows that prenatal exposure to violence is negatively associated with child's health.

Fetal, Child, and Adolescent Health
A one SD increase in homicide rates is associated with a decline in birth weight (although not statistically significant) and a significant increase in the likelihood of low birth weight of 5 percentage points (or a 60% rise with respect to the outcome mean). I also find that changes in violence are positively associated with the probability that a child is born at a clinic or hospital (as opposed to at home) but little change on prenatal care use, which suggest that perhaps access to care might not be the main channel through which violence affects health. I do find, however, a lower probability that a mother breastfeeds her child for at least 6 months (the minimum recommended duration according to the WHO), which could be in part explained by changes in mothers' nutrition, health, or stress. Column 8 shows that children experience a substantial decline in HAZ, a widely used indicator of child's nutrition and health status. Results suggest that prenatal exposure to violence reduces HAZ by 0.56 of a SD, an effect that is consistent to that found in previous studies, ranging between 0.1 and 1.0 SD (Akresh, Lucchetti and Thirumurthy, 2012;Bundervoet, Verwimp and Akresh, 2009;Duque, 2017;Minoiu and Shemyakina, 2014).
Using an independent indicator of long-run health, height, affirm that childhood exposure to violence negatively affects young adults' health. Figure 8 plot estimates for the standardized height and results show that, exposure to violence from in-utero up to age 6 significantly reduce 30 While the administrative data provides a number of empirical advantages with respect to other datasets, it lacks information on health outcomes or parental behavior. 31 Or the Encuesta Nacional de Salud y Nutricion in Spanish.
height by a large 0.1 SD per each year of high violence exposure. Linear pre-trends are again flat. These findings are consistent with the impacts of violence shown in Table 3 column 8 and suggest that gaps in child's height emerge early in life.

Parental Investments
Lastly, I examine whether violence affects other inputs in the production of human capital using the DHS and administrative data, and the population Census of 1993. In the DHS data, I test whether violence is associated with changes in the home environment as measured by domestic violence. Although domestic violence is self-reported and households are likely to misreport it, results show that children exposed to high homicide rates in early-life are significantly more likely to live in families experiencing domestic violence and to be a witness of domestic violence themselves as shown in columns 1 and 2 in Table 4. Column 3 shows that violence increases the probability that a child enters "over-age" (i.e., after age 7) for first grade.
Lastly, conditional on school enrollment, Appendix Figure 3 shows that they are more likely to repeat a school grade with violence exposure prior to school entry.
Overall, these results provide some supportive evidence that declines in child's health and nutrition and household resources may be some of the potential channels through which violence affects future educational outcomes.

Robustness Checks
In this section, I analyze potential sources of selection bias such as fertility, mobility, and survival.

Fertility
Violence could affect a woman's fertility decisions by either influencing her desired number of children or by delaying her decision to become pregnant. If violence affects some women more than others based on their observable characteristics, this could potentially lead to a biased estimate of violence on education.
To test for the presence of selective fertility, I examine the association between violence and a mother's total fertility after a (focal) child (i.e., the number of children she has after her first child born in the 1980s) and the association between violence and the timing of fertility (i.e., the time between two subsequent births). Results shown in Appendix Table 2 suggest little evidence on this potential source of bias; most of the coefficients on violence are small and statistically insignificant suggesting that higher homicide rates during the 1980s do not seem to be related to changes in fertility.

Mobility
Endogenous migration in response to (or in expectation of) high violence could also induce (an upward) bias in my estimates of violence on education, if households who migrate due to an increase in homicide rates are different to those that don't migrate, in dimensions that could affect a child's education (e.g., if migrant families are wealthier or more educated than those who stay).
To asses the importance of selective migration, I first examine the proportion of migrants in the data and I find that 25% of the sample reports having been born in a municipality that is different from where they were interviewed in the Sisben data. 32 Second, to empirically examine how endogenous migration affects my estimates on education, one possible way is to estimate whether the effects of violence in the full sample differ to those in the sample of nonmovers. Appendix Figure 1 shows estimates of violence on children's education for these two groups. Results indicate that the negative effects of violence are always larger in the sample of non-migrants than in the full sample, which could suggest that, the non-migrants tend to be a more-negatively selected sample that even in the absence of violence would invest less than other households in their children's education or that migrants could have invested more in their education in expectation of high violence. The fact that this paper uses the municipality of birth to assign violence exposure (instead of the violence experienced at the place of residence) suggests that the true impacts could be larger than those found here.
A third step in the analysis of endogenous migration focuses on the potential effects of the internal forced displacement (IFD) on my estimates. The IFD has been one of the most dramatic consequences of the recent armed conflict in Colombia, where more than 3.5 million people have been forced to migrate since 1997 (around 8% of the total population) (United Nations, 2010). 33 Appendix Figure 2 shows the number of forced migrants since 1979 in Colombia. As the figure illustrates, forced displacement is a relatively recent phenomenon in the country (i.e., late 1990s and early 2000s) and as such, I assume that it did not directly affect the early lives of the cohorts who were born in the 1980s in the country. 34 32 In the population Census of 2005, the fraction of migrants is 30% 33 The IFD groups have very low socioeconomic indicators, including educational attainment and health status. Identifying internally displaced populations (IDP) in the Census 2005 is only possible for those who have recently been forced to move (in the last 5 years). 34 Furthermore, while I cannot identify IDPs prior to year 2000, the number of IDPs in the Census is small (they represent 5% of the recent migrants and are approximately 2% of the total sample), providing some additional evidence that internal migration in the sample is not particularly driven by violence.
I conclude that there is little evidence that could suggest that endogenous migration due to violence could be driving the estimates of homicide rates on long-term educational outcomes.
The fact that I obtain comparable estimates in the family fixed effects model provides further evidence pointing in this direction.

Survival
Lastly, I analyze how violence affects the probability of survival. Selective survival can occur if, for instance, violence increases the probability of mortality disproportionately among the frailer fetuses or children, leading to healthier babies or children surviving. In this case, the effects of violence on education would result in a lower-bound estimate of the true impact.
One way to test for selective survival is to explore whether changes in violence are associated with changes in the sex ratio or in the cohort size. Several studies have shown that boys are biologically weaker and more susceptible to diseases and premature death than girls (Naeye et al., 1971;Waldron, 1985), and that they are more vulnerable to environmental factors in early-life (Pongou, 2013). Therefore, focusing on these two measures could be informative to understand potential demographic imbalances linked to violence.
I assess the importance of selective survival by regressing, at the municipality-year level, the sex ratio and cohort size of the cohorts of interest (i.e., individuals born in the 1980s), on the violence that they experienced over the life course. Results shown in Appendix Table 3 show that although coefficients associated with violence are negative, overall these are small in magnitude and not statistically significant suggesting little evidence that changes in homicide rates are associated with changes in the ratio of males versus females observed in the Census in 2005. Column 2 shows associates between violence and the cohort size, which again, do not seem to suggest a significant change in population size due to higher homicide rates. Now, I examine the link between violence and child's sex and child mortality at the household level using the DHS data. 35 I estimate the association between early-life violence exposure and the probability that a child is female and the probability that a child dies before age 1 and before age 3, respectively. 36 Appendix Table 4 shows no significant association between changes in prenatal violence and child's sex (consistent with the results shown in Appendix Table 3), but the findings do suggest that child mortality increases with an increase in violence. This change represents an increase of a third with respect to the outcome mean. While these are not negligible effect sizes, this type of selection bias will likely result in estimates of violence that 35 I use DHS data because the Census do not provide information on child mortality. 36 Unfortunately, due to data limitations I am not able to examine child mortality at later ages.
represent a lower bound if one way that violence affects human capital is through an increased in the probability of child mortality. 37

Conclusions
This paper contributes to the literature on the effects of violence on human capital by providing new evidence on the long reach of early-life violence exposure on human capital outcomes.
It does so by using a large and representative data set that link school records with census data and with information on the place and date of birth for the universe of low income students in Colombia. The paper exploits the temporal and geographic variation in crime and violence associated with the proliferation of drug cartels in the 1980s, which was greatly influenced by the huge increase in the international demand for cocaine in the 1980s.
Results show that early-life exposure to homicide rate increases the probability of HS dropout by 15% and reduces the end-of-HS exam score by 0.5 SD. The estimates of violence found in this paper can be viewed as lower bounds on the true value. First, by using homicide rates at the municipality level, I am unable to point out whether, within a particular municipality there is actual sorting into who is being more or less exposed to violence. If families adopt preventive measures that would reduce their own exposure to violence or if they make additional investments on their violence-exposed child, it is likely that the impacts that I estimate represent a lower bound of the true effect. Second, I measure early-life violence exposure based on a child's municipality of birth. Since migration is always a challenge, if a family moves to a different municipality right after their child's birth, I will not be able to identify his/her 'current' homicide-exposure in that new location. Although using the baseline (place of birth) municipality is more exogenous than the current place of residence, it may introduce an additional attenuation bias in the estimates. Despite these sources of measurement error, it is clear that the results presented in this paper are economically meaningful. For instance, the effects on test scores are within the range of the effects of many educational interventions (Hanushek, 2006). Moreover, the fact that the results obtained in the main difference-in-difference specification are similar to those in the family fixed effects model, is reassuring that the effects of violence do not seem to be driven by potential sources of selection bias and rather reflect a scarring effect on children's future outcomes.
I conclude that violence had persistent effects on the long-term outcomes of disadvantaged cohorts born in Colombia during the 1980s, a result that may have disturbing implications for their future labor market prospects as well as for their own children's outcomes, considering the strong intergenerational transmission of human capital (Aizer and Currie, 2014;Black et al., 2005

Figure 8. Mother Fixed-Effects Estimates of Violence on High School Completion
Note: The family fixed effects include controls for sex, birth order, age, mother's age, fixed effects at the child's year, month, and municipality of birth, and department linear time trends. Errors are clustered at the municipality level. Dashed lines represent confidence intervals at the 95% level.    Note: Sample includes individuals 17 to 24 (born in the 1980s) observed in ENSIN data. All models include controls for sex, age, education, urban/rural status, fixed effects at the child's year, month, and municipality of birth, and department linear time trends. The outcome is measured as a height-for-age indicator in standard deviations using the WHO reference groups. Errors are clustered at the municipality level. Dashed lines represent confidence intervals at the 95% level.  Note: sample includes DHS data for 1986 and 1990. All models include individual controls for child's gender and age, mother's age, marital status, education, whether the household is urban, household size, and fixed effects at the individual's year, month, and municipality of birth. Low is defined as whether a mother's education is less than HS and high means HS plus. Errors are clustered at the municipality level. *** p<0.01, ** p<0.05, * p<0.1. Note: sample includes DHS data for 1986 and 1990. All models include individual controls for child's gender and age, mother's age, marital status, education, whether the household is urban, household size, and fixed effects at the individual's year, month, and municipality of birth. Errors are clustered at the municipality level. *** p<0.01, ** p<0.05, * p<0.1.

A. Appendix: Figures and Tables
Appendix Table 1 Note: the sample is a panel at the municipality-month-year levels. The sex ratio is defined as the number of boys per number of girls in a given year-municipality associated with the cohort of interest (those born in the 1980s). The cohort size is calculated as the number of individuals born in a given municipality-year associated during the 1980s. All models include fixed effects at the year, month, and municipality levels, department-specific time trends, and population weights. Errors are clustered at the municipality level. *** p<0.01, ** p<0.05, * p<0.1.
Appendix Note: the sample includes families with a child below age 5 in the DHS data in waves 1986, 1990, and 1995. The outcomes of interest are dummies for whether a child died before age 1 or before age 3, and 0 otherwise. Models include individual and mother covariates as well as year, month, and municipality of birth fixed effects, wave fixed effects, and department-specific time trends. Errors are clustered at the municipality of birth. *** p<0.01, ** p<0.05, * p<0.1.