Early-life conditions and child development: Evidence from a violent conflict

Abstract

This paper investigates how the exposure to violent conflicts in utero and in early and late childhood affect human capital formation. I focus on a wide range of child development outcomes, including novel cognitive and non-cognitive indicators. Using monthly and municipality-level variation in the timing and severity of massacres in Colombia from 1999 to 2007, I show that children exposed to terrorist attacks in utero and in childhood achieve lower height-for-age (0.09 SD) and cognitive outcomes (PPVT falls by 0.18SD and math reasoning and general knowledge fall by 0.16SD), and that these results are robust to controlling for mother fixed-effects. The timing of these exposures matters and differs by type of skill. In terms of parental investments, I find some evidence that parents reinforce the negative effects of violence by increasing their frequency of physical aggression.

Highlights

• •I investigate the effects of early-life violence exposure on human capital formation.
• •I focus on a wide range of outcomes, including cognitive and non-cognitive measures.
• •Violence is measured using sharp variation in the timing and severity of massacres.
• •These terrorist attacks reduce child height-for-age and cognitive test scores.
• •The timing of exposures matters and differs by type of skill.

1. . Introduction

It is well established that prenatal events have life-long consequences (Barker, 1992, Cunha and Heckman, 2007, Almond and Currie, 2011). Moreover, events that happen in the postnatal period also matter for future outcomes and its effects can often be large (Almond, Currie & Duque, 2016). This paper examines the effects of a world-wide public health concern that affects in-utero, early, and late childhood conditions on children's development: Exposure to violence (i.e., wars, armed conflicts, urban crime). More than 1.5 billion people live in countries affected by repeated cycles of violence (World Bank, 2013), many of which are children (Grantham-McGregor et al., 2007).

Recent research has shown the large damage on education and health outcomes from early-life violence (Camacho, 2008, Akresh et al., 2012, Brown, 2014, Valente, 2011, Leon, 2012). I contribute to this literature in two dimensions. First, how does violence affect other domains of human capital beside education and health (i.e., cognitive and non-cognitive skills)? Identifying such effects is important both because measures of human capital (physical, cognitive, and non-cognitive indicators) can explain a large percentage of the variation in later-life educational attainment and wages (Heckman, Stixrud, & Urzua, 2006) and to understand mechanisms behind previous effects found for educational attainment and health. Second, to what extent do the effects of violence at different developmental stages differ? In particular, I analyze the effects of violence in each trimester of pregnancy and in two distinct post-birth developmental stages, early (ages 0–3) and late (ages 3+) childhood, following the large evidence of critical and sensitive periods in human development (Gluckman and Hanson, 2005, Knudsen et al., 2009). Moreover, do impacts on the particular type of skill considered (e.g., health vs. cognitive outcomes) differ by the developmental timing of the shock? Identifying the timing of exposures is important to facilitate investigation of the mechanisms through which conditions affect later outcomes (Conti & Heckman, 2014).

To study these questions, I analyze survey microdata on 13,400 children collected in 2007 to evaluate a large social program in Colombia: a home-based childcare program called Hogares Comunitarios de Bienestar (HCB). HCB serves a million low-income children below age 7 with the goal to promote their health and cognitive and socio-emotional development by providing childcare, nutrition (50–70% of the daily allowance), and psychosocial stimulation. The HCB evaluation survey provides rich measures on child development that are not available in other national surveys, as well as it includes detailed information on parental investments that allow me to explore novel dimensions of the potential effects of violence. Most importantly, these data contain information on each child's year and month of birth, as well as household migration history, which allows me to identify with some precision a child's violence exposure in early-life. The data also include a subsample of siblings that I use to estimate models that account for a mother's time-invariant characteristic, which might be correlated with both the probability of residing in an area with high violence and with accumulating low levels of human capital. The mother fixed-effects models thus provide robust evidence on the effect of violence on children.

I measure violence shocks using the occurrence and severity of massacres at the monthly-municipality levels during Colombia's armed conflict, from 1999 to 2007 (period for which the microdata are available). Massacres are defined as the intentional killing of four or more people in a specific time and place, by another person or group. These events distinguish from other measures of violence (e.g., homicides), in their high levels of cruelty and visibility of violence, and in the huge levels of stress and anxiety that they cause. As discussed in more detail in the background section, these terrorist attacks were a common practice employed by violent groups during the most intense years of the conflict. From 1999–2007, more than 1000 massacres occurred in half of all municipalities, ranging from 4 to more than 25 victims killed in each episode. While the occurrence of massacres in a given municipality was not random –armed groups could have targeted specific populations or, in some cases, could have announced their terrorist plans–, the identification strategy relies on the fact that the timing within and across municipalities is uncorrelated with changes in other factors affecting human capital investments and with a child or family unobserved characteristics. I provide some evidence that supports this identifying assumption by showing that: (i) changes in the intensity of violence are not associated with changes in socioeconomic conditions at the local level measured by income or unemployment rate; and (ii), that the intensity of massacres at the municipality-monthly levels depicts little serial correlation.

Using models that control for a variety of individual-level characteristics, as well as geographic and temporal fixed-effects, I show that children exposed to violence in-utero and in childhood experience a decline in their physical and cognitive development. This finding is confirmed using the mother fixed-effects specification. In particular, a one standard deviation (SD) increase in violence in late pregnancy reduce height-for-age Z-scores (HAZ) by 0.09 SD and cognitive test scores by at least 0.16 SD. I also find negative but statistically insignificant effects on socio-emotional outcomes. Moreover, results show that violence is negatively associated with birth weight, an important input in the production of human capital. This impact is driven by exposure during the first trimester of pregnancy. Together, these results and the timing in which these exposures occur may suggest that stress is a relevant mechanism through which violence affects children.

Given the size and persistence of the effects of violence, I then ask an additional question: Do and how do parents respond to these shocks? To my knowledge, this is the first study that investigates the link between violence and parental responses. Family investments are important determinants of human capital (Cunha and Heckman, 2007, Aizer and Cunha, 2014) and parental responses can play a key role in compensating or reinforcing the effects of a shock (Almond and Currie, 2011, Almond and Mazumder, 2013). At present, well-identified empirical evidence on this question is scarce.

I examine the association between violence and parenting, a key human capital investment. Results show little evidence that parents respond to the negative effects of violence on children by providing more nurturing care. In fact, I find that parental responses become harsher as violence increases (the frequency of personal care routines falls and physical aggression increases). As is discussed further below, these findings are consistent with the idea that parents reinforce the negative effects of the shock (Almond & Mazumder, 2013).

2. Background

2.1. The effects of violence on human capital

Previous research has shown the negative effects of early-life violence exposure on health, education, and labor-market prospects. Studies investigating prenatal violence exposure have found declines in birth weight (Brown, 2014, Camacho, 2008, Mansour and Rees, 2012, Valente, 2011) while those focusing on before-age-6 exposures have found a decrease in educational attainment, wages, and adult height (Chamarbagwala and Morán, 2011, Leon, 2012, Galdo, 2013). Relevant to my study, existing literature has found that war exposure during the first years of life reduces child HAZ by 0.2–0.5 SD (Bundervoet et al., 2009, Akresh et al., 2012).

To my knowledge, only two studies have explored how violence affects other dimensions of human capital. Rodriguez and Sánchez (2013) found that a 1 SD increase in the intensity of Colombia's conflict reduced test scores by 0.86 SD for children aged 11–18, a group much older than those analyzed here. Sharkey, Tirado-Strayer, Papachristos, and Raver (2012) found that being exposed to a neighborhood homicide in Chicago, worsened 4-year-old's behavior and academic skills, as well as attention, impulse control, and test scores by over a 0.3 SD. While these are large effects, the fact that it uses data from a developed country makes it difficult to compare estimates to those in developing countries. Moreover, these estimates represent short-term impacts and do not reflect lasting effects.

2.2. The timing of exposure and potential mechanisms

Violence is a major source of stress that can affect a family's resources and behavior (e.g., access to food). During pregnancy, stress and nutritional deprivation can affect fetal and newborn health and cognition through changes in the immune and behavioral systems (Barker, 1992, Denckel-Schetter, 2011, Gluckman and Hanson, 2005). The medical literature suggests that the timing of these alterations matter for a child's physical and cognitive development. Since the number of neurons is mostly determined by mid-gestation, both nutritional deprivation and stress in the first half of pregnancy may be particularly harmful for cognitive development. Fetal exposure to excess cortisol –the hormones responsible for regulating fetal maturation –may lead to impaired brain and spinal cord development, thereby diminishing the mental and motor skills of infants (Huizink, Robles de Medina, Mulder, Visser, & Buitelaar, 2003), and is associated with lower schooling and verbal IQ scores, and high chances of experiencing chronic health conditions at age 7 (Aizer, Stroud, & Buka, 2012) and later in life (Thompson, 2012). On the other hand, child height can be particularly sensitive to nutritional deprivation in the second half of pregnancy – the period in which the mother gains more weight (Stein & Lumey, 2000) –, as well as conditions during the first years of life (e.g., nutrition, infectious diseases) (Victora & de Onis, 2010).

During childhood, stress may compromise the family environment by affecting parental mental health and family relationships (Campbell, 1991, Repetti et al., 2002). Households can also modify their behavior in order to prevent victimization (e.g., mothers may refrain from letting their child leave the home or play outside). Sharkey et al. (2012) found that local violence is positively associated with higher parental distress, suggesting that parental responses may be a likely pathway by which violence affects children. Neurobiologists have shown that a strong and positive attachment in infancy promotes brain growth and healthy development (Schore, 2001). Thus, if violence disrupts the home environment, it may affect the child through changes in mother–child interaction.

Lastly, in terms of the supply-side mechanisms, Leon (2012) found that attacks against teachers decreased educational attainment; Rodriguez and Sánchez (2013) found that negative economic shocks and lower school quality due to violence, increased school dropout and child labor; Akbulut-Yuksel (2009) found that school-facility destruction and teacher absence lead to declines in education, and malnutrition and destruction of health services worsened adult's health; Minoiu and Shemyakina (2012) found that household economic losses helped explain declines in children's height and Akresh et al. (2012) showed that forced displacement was an important mechanism through which children's nutrition was affected.

2.3. Massacres

For more than 50 years, Colombia has faced an internal armed conflict, one of the longest in the world (for a detail description of the conflict see Grupo de Memoria Histórica (2013)). Massacres were a common practice employed by armed groups in the conflict, especially during the most intense years, 1996–2002. Fig. 1 depicts the monthly variation in massacres and victims of massacres in Colombia since 1993. As the figure shows, the occurrence of these terrorist attacks fluctuated between 5 and 15 massacres/month during the first half of the 1990s until it peak in the early 2000s when more than 25 monthly-massacres took place (with more than 150 monthly victims). Since 2004, the number declined due to the demobilization of paramilitary groups.

Fig. 1.

Monthly Number and Victims of Massacres in Colombia 1993–2009.

Source: University of Los Andes-CEDE violence dataset.

In terms of the spatial distribution, Fig. 2 shows that these attacks were particularly concentrated in certain areas of the country (i.e., the central and northern parts). These regions tend to be more developed, dense, more likely to be urban, and wealthier (see Table C9). Moreover, even within these more violent regions, the temporal variation was actually very large. Fig. 3 shows numbers on victims of massacres for the six largest (and wealthiest) cities, which together represent 40% of the country's population (Colombia's population is approx. 45 million). So, for instance, a child who was born in March 2003 in Bogota was exposed to a level of violence in-utero that was different to the in-utero violence experienced by a child born in November 2003 in Bogota. The high-frequency and large variation in the occurrence of these events by municipality-year constitutes a vital component of my identification strategy.

Fig. 2.

Victims of Massacres in Colombia in 2000 and 2005.

Source: University of Los Andes-CEDE violence dataset.

Fig. 3.

Monthly Victims of Massacres by City.

Source: University of Los Andes-CEDE violence dataset.

Massacres were not random terrorist acts; instead, they were a deliberate strategy of illegal armed groups to expand their territorial, economic, and social control (Duncan, 2006). While the decision to commit a massacre in a given region was not random, I argue that my identification strategy – to exploit the timing of massacres at the monthly level both within and across municipalities –, allows to overcome this selection problem and helps provide evidence on the impact of violence on individual outcomes. In particular, I provide some evidence that supports this identifying assumption by showing that changes in the intensity of violence are not associated with changes in economic conditions at the local level (i.e., income, investments in public goods, or unemployment), which could potentially affect the outcome, and that the intensity of violence depicts little serial correlation over time (conditional on controlling for municipality, year, and month fixed-effects). Additionally, this paper focuses on the period from 1999 to 2007, which exploits the huge variation from both the most intense years of the conflict, 1996–2002, and from a period of very low violence (see Fig. 1, Fig. 2, Fig. 3). Lastly, while massacres could have been announced in some cases, this was typically done within days prior to the episode, which limited household coping strategies and still caused significant levels of stress and losses on the local population (Appendix A presents anecdotal evidence of victims of massacres and newspaper articles that support these claims).

3. Data

3.1. Data on children: HCB

To investigate the effects of violence on children, I use the household survey collected in 2007 to evaluate Colombia's home-based childcare program, Hogares Comunitarios de Bienestar. Because HCB had already been implemented for several decades before it was formally evaluated in 2007, scientists used a non-experimental approach to estimate the treatment effects of the program (see Bernal et al. (2009) for more details). Data for the evaluation were collected from a random sample of HCB centers across 69 municipalities (out of the 1100), in 29 departments (out of the 33), and included exceptional measures on child outcomes for 21,000 children aged 1–7, as well as information on the exact date and place of birth for a large number of children. A particular advantage of these data is that it include a subsample of siblings that I use to estimate mother fixed-effects models. The reason for the small sibling sample is that, the HCB evaluation survey was not originally designed to sample all members within a family – only the focal child and parents were interviewed –, and so the sibling cases were families in which both siblings were HCB participants, hence, this subsample tends to be small and more disadvantaged than those in the full sample. I carefully analyze the contribution of the mother fixed-effects in the results section.

Two possible limitations of using these data are: (i) since this program is exclusively for the poorest households in the country, the survey only samples families in the lowest income quartile and so it is not representative of the Colombian population; and (ii), given the selection criteria of participants in the HCB data, the municipalities included in the sample are mostly urban.

3.1.1. Outcome variables

I investigate three sets of outcomes for children 3–7 years of age (all outcomes are rescaled with 0 mean and unit SD): Physical health (nutritional status) is measured using a child's HAZ. Cognitive development is measured using the Peabody Picture Vocabulary Test (PPVT) that captures a child's receptive language and the Spanish version of the Woodcock–Johnson (WJ) battery III, which measures mathematical reasoning and general knowledge about the world. Socio-emotional development is measured using the Penn Interactive Peer Play Scale (PIPPS), which capture a child's aggression, isolation, and adequate interaction during peer-play.

3.1.2. Sample of interest

While the HCB data include children between 1 and 7 years of age, I focus on children who were 3 or older in 2007 (i.e., who were born between 2000 and 2004). I restrict the sample to these older children because it allows me to measure their complete early-childhood period (ages 0–3). Moreover, I restrict the sample to children in non-migrant families since the HCB data do not include information on a child's municipality of birth. Based on a mother's reports on household migration history, I identify those children who were in-utero and in early childhood in the municipality of interview. Almost 90% are classified as non-migrants (of a total of 15,379 children). The remaining sample is therefore excluded from the analysis. I come back to this point in Section 7 were I carefully analyze the problem of endogenous migration. Comparing the observable characteristics of migrants and non-migrants, I find that the non-migrants tend to be a more advantaged group in terms of mother's age, education, and marital status, and that children in these families are significantly more likely to have better outcomes in all dimensions (see Table B3). Thus, excluding children of migrant families from my analyses may bias the estimates of violence towards zero.

3.2. Data on violence: massacres

Data on massacres comes from the CEDE research center at University of Los Andes (Bogota) and it includes information on the number of victims killed in each massacre by municipality-year-month in Colombia since 1993. I merge the violence data with the HCB data based on a child's municipality, month, and year of birth.

I construct five measures of violence exposure: three for the prenatal period (one for each trimester) and two for the postnatal period (one for early childhood, ages 0–3, and one for late childhood, ages 3+). To identify each trimester of pregnancy I count backwards nine months from the date of birth of the child. Since I do not have information on the duration of pregnancy (i.e., gestation weeks), I also test my results assuming that the child was born prior to completing his/her nine months of pregnancy (i.e., 8 months). Similarly, I construct measures for early and late childhood by counting 36 months since the month and year of birth, and 37 months or more, respectively.

4. Methods

I estimate the effects of violence on children using two models, one that controls for a rich set of covariates, and one that accounts for time-invariant mother characteristics. Eq. (1) describes the first model:

$$Y_{\mathit{ijmt}}={\beta }_0+\sum _{q=1}^3{\beta }_q^u{\mathit{Violence}}_{\mathit{jmt}}^{{\mathit{trim}}_q}+\sum _{r=1}^2{\beta }_r^c{\mathit{Violence}}_{\mathit{jmt}}^{{\mathit{childhood}}_r}+\gamma {\mathbf{X}}_i+{\alpha }_j+{\alpha }_m+{\alpha }_t+{\theta }_j(t)+{ϵ}_{\mathit{ijmt}},$$ (1)

where the variable $\mathit{Y}$ denotes child i's outcome and the subscript j refers to the municipality, m the month, and t the year of birth. Violence^trim₁, Violence^trim₂, and Violence^trim₃ represent the level of violence (victims of massacres) to which a child was exposed to during the first, second, and third trimesters of pregnancy, and ${\mathit{Violence}}^{{\mathit{childhood}}_1}$ and ${\mathit{Violence}}^{{\mathit{childhood}}_2}$ represent the level of violence during early (ages 0–3) and late (ages 3+) childhood. $\mathit{X}$ includes a set of child characteristics such as gender and age in month dummies (36–48, 49–60, 60–72, 73+) and an indicator for HCB participation, as well as dummies for mother's age in years (23, 23 to −26, 27 to −33, 33+), education (completed primary or less, less than high school, high school or more, and unknown), and marital status (married, cohabiting, single, other). The terms α_j, α_t, and α_m are fixed effects at the municipality, year, and month of the 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 ${\theta }_j(t)$ represents municipality linear time trends that control for differences in economic development across municipalities that change linearly over time (e.g., investments in health services) and that could potentially affect a child's development. These trends also allow me to account for differential linear trends in child development 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 ${\beta }_1^u-{\beta }_3^u$ and ${\beta }_1^c-{\beta }_2^c$, as they describe the impacts of violence in each trimester of pregnancy and in early and late childhood.

The second model controls for mother fixed-effects and is estimated using Eq. (2). The only covariates included in this model are child's gender and age in month dummies (in matrix $\tilde{\mathit{X}}$), year and month of child's birth dummies (vectors α_t and α_m, respectively), and municipality linear time trends. The term μ_f indexes families.

$$Y_{\mathit{ijmt}}=\sum _{q=1}^3{\beta }_q^u{\mathit{Violence}}_{\mathit{jmt}}^{{\mathit{trim}}_q}+\sum _{r=1}^2{\beta }_r^c{\mathit{Violence}}_{\mathit{jmt}}^{{\mathit{childhood}}_r}+\gamma {\tilde{\mathbf{X}}}_i+{\alpha }_m+{\alpha }_t+{\mu }_f+{\theta }_j(t)+{ϵ}_{\mathit{ijmt}}.$$ (2)

This model exploits the sample of siblings in the HCB evaluation (2182 siblings out of the 13,444 children) 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 experiencing “worse” developmental outcomes. 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 children's health and education.

5. Results

5.1. Descriptive statistics

Table 1 reports descriptive characteristics for the sample of children and mothers and by violence exposure. Low (high) violence is defined as whether a child was exposed to an overall violence level, from in-utero through late childhood, that is equal to or below (above) the median violence (25 victims of massacres). Results show little differences across groups; however, I do find that mothers of children who grew up in violent contexts tend to be slightly more educated and are more likely to be single, compared to other mothers. Children in these environments also tend to be older; by construction, older children have had a longer exposure to violence than younger children. Thus, I include controls for child's age in month-group-dummies in all specifications to account for this mechanical correlation (using child's age in month fixed-effects instead of the age group dummies, provides substantially similar estimates of the effects of violence). In terms of child outcomes, I find that violence-exposed children tend to have lower socio-emotional outcomes; however, these differences are not statistically different.

Table 1.

Descriptive statistics by violence exposure.

	Full sample	No violence	Violence
			Low	High
Mother characteristics:
Age	25.48	25.70	25.54	25.30
	[6.67]	[6.61]	[6.72]	[6.69]
Primary or less***	0.33	0.35	0.34	0.30
Less than HS***	0.29	0.26	0.28	0.31
HS or more**	0.32	0.32	0.30	0.33
Unknown	0.07	0.07	0.07	0.06
Married	0.18	0.19	0.17	0.17
Cohabiting	0.55	0.56	0.55	0.55
Single***	0.10	0.09	0.09	0.11
Divorced/widow**	0.17	0.16	0.19	0.17
Child characteristics:
Female	0.48	0.49	0.48	0.48
Age (months)***	49.6	46.62	49.79	51.12
	[9.47]	[8.88]	[9.04]	[9.78]
Participates in HCB	0.50	0.53	0.49	0.50
Child outcomes:
HAZ (Z-scores)	−0.98	−1.05	−0.95	−0.96
	[1.02]	[1.02]	[1.02]	[1.02]
PPVT	0.00	0.00	0.02	0.00
Math ability	0.00	−0.08	0.05	0.05
General knowledge	0.00	−0.03	0.01	0.03
Aggression	0.00	−0.08	0.01	0.04
Isolation	0.00	0.01	−0.01	0.00
Adequate interaction	0.00	−0.01	0.05	0.02
Violence (massacre victims):
Trimester 1	4.92	0.00	1.21	11.48
	[10.28]	–	[2.45]	[14.01]
Trimester 2	4.85	0.00	1.34	11.17
	[10.33]	–	[2.59]	[14.27]
Trimester 3	4.42	0.00	1.19	10.05
	[9.79]	–	[3.11]	[13.58]
Childhood 0–3	97.26	0.00	23.58	227.23
	[156.28]	–	[22.83]	[186.67]
Childhood 3+	29.69	0.00	4.07	72.59
	[86.57]	–	[7.56]	[127.71]
N	13,344	2888	5115	5341

Note: Sample includes children 3–7 years of age. Please refer to Section 5.1 for details. ***p<0.01, **p<0.05, *p<0.1.

The bottom of Table 1 includes descriptive statistics on the level of violence in-utero and in childhood, measured in victims of massacres. On average, children were exposed to 5 victims of massacres in each trimester of pregnancy (with a standard deviation of 10 victims), almost 100 victims in early-childhood (ages 0–3) (with a SD of 150 victims) and 30 victims in late-childhood (ages 3+) (with a SD of 90 victims).

5.2. Effects of violence on physical health

The first set of results is shown in Table 2. I only report the coefficients of interest, but the models include all covariates as described in Eq. (1).

Table 2.

The effect of violence on children's Health.

	Height-for-Age (Z-scores)
	(1)	(2)	(3)	(4)
2 Years before conception				−0.0003
				[0.0021]
1 year before conception				0.0001
				[0.0010]
In-Utero	−0.0016**
	[0.0006]
Trimester 1		0.0003	−0.0006	−0.0007
		[0.0017]	[0.0015]	[0.0023]
Trimester 2		−0.0027***	−0.0038***	−0.0038***
		[0.0005]	[0.0006]	[0.0010]
Trimester 3		−0.0022	−0.0032*	−0.0032*
		[0.0014]	[0.0017]	[0.0017]
Childhood 0–3			−0.0006***	−0.0006***
			[0.0002]	[0.0002]
Childhood 3+			−0.0002	−0.0002
			[0.0002]	[0.0003]
N	13,344	13,344	13,344	13,344

Note: Sample includes children 3–7. Please refer to Section 4 for details.

^***$p<0.01$.

^**$p<0.05$.

^*$p<0.1$.

I start by estimating the effect of violence during the in-utero period, as shown in column 1. Then, I disaggregate the effect by each trimester of pregnancy (column 2), and in column 3, I include the violence exposures in early and in late childhood. Results show that violence is particularly harmful in the second and third trimesters and during the first 3 years of a child's life. A one standard deviation increase in violence is associated with a 0.04 SD decline in HAZ if the increase in violence occurs in the second trimester, with a 0.03 SD decline in HAZ if the increase in violence occurs in the third trimester, or with a 0.09 SD decline in HAZ if the increase in violence occurs in a child's early childhood. In column 4, I include controls for exposure to violence in the two years before conception, which helps inform whether preexisting trends in violence could be driving the estimates. I find little evidence that these prior exposures have an impact on child's height.

One potential threat to my empirical specification occurs if mothers of certain characteristics who are more likely to be impacted by violence have children with “worse” developmental outcomes. If this were the case, the estimates of violence from Eq. (1) could be overestimating the true impact. Accounting for a mother fixed-effect allows to control for this potential bias.

Table 5 columns 1 and 2 show estimates on the effects of violence on HAZ for the sample of siblings, using both the OLS specification (Eq. (1)) and the mother fixed-effects model (Eq. (2)). The OLS specification conducted on the sibling sample serves as a way to test for the fact that the sibling sample may be a non-representative sample of the full sample of children (i.e., of the 13,144 children in the full sample with complete information on HAZ, only 2182 were siblings). If the results from this analysis are similar to those from the full sample OLS, it provides some evidence that the smaller sibling sample is a random draw of the larger sample.

Table 5.

Mother fixed-effects estimates of violence on children's health and cognitive outcomes.

	HAZ		PPVT		Math reasoning		General knowledge
	OLS sibling sample	Mother FE	OLS sibling sample	Mother FE	OLS sibling sample	Mother FE	OLS sibling sample	Mother FE
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
TRIM 1	0.0029	−0.0006	−0.0055***	−0.0044	−0.0136**	−0.0164**	−0.0084	−0.0005
	[0.0032]	[0.0056]	[0.0022]	[0.0079]	[0.0052]	[0.0081]	[0.0057]	[0.0079]
TRIM 2	−0.0033*	0.0034	−0.0052	0.0046	−0.0068	−0.0286***	−0.007	−0.0169*
	[0.0018]	[0.0058]	[0.0037]	[0.0103]	[0.0053]	[0.0098]	[0.0042]	[0.0094]
TRIM 3	−0.0044**	−0.0099*	−0.0035	−0.0179**	0.0014	−0.0081	−0.0069	−0.0160*
	[0.0017]	[0.0057]	[0.0041]	[0.0082]	[0.0043]	[0.0085]	[0.0044]	[0.0082]
0–3	−0.0014***	−0.0003	−0.0007	−0.0008	−0.0012***	−0.0037***	−0.0014*	−0.0005
	[0.0005]	[0.0005]	[0.0005]	[0.0008]	[0.0004]	[0.0008]	[0.0008]	[0.0008]
3+	−0.0005	−0.0001	0.0006	−0.0003	−0.0004	−0.0064***	−0.0011	−0.0018
	[0.0007]	[0.0007]	[0.0014]	[0.0021]	[0.0010]	[0.0020]	[0.0010]	[0.0019]
N	2182	2182	741	741	754	754	766	766

Note: Sample includes all children 3–7. Please refer to Section 4 for details.

^***$p<0.01$.

^**$p<0.05$.

^*$p<0.1$.

Comparing the full sample estimates in Table 2 and the sibling sample results in Table 5 suggest that, while the smaller sample size causes less precision in the estimates, the two sets of results are very similar in magnitude and qualitatively tell the same story. For the most part, the coefficients in the full sample OLS and sibling sample OLS are not statistically distinguishable from each other, which suggests that the sibling sample is representative of the full sample. Therefore, the maternal fixed effect estimates for HAZ are useful in terms of examining the potential biasing effects of fixed unobserved heterogeneity across mothers.

Column 2 shows mother fixed-effects estimates and these results suggest that, changes in violence occurring during the third trimester significantly reduce HAZ. This finding contrasts with that found in Table 2 in which violence in the second and third trimesters as well as in the first 3 years of life, adversely impact height. Moreover, the magnitude of the effect of violence in the third trimester has actually tripled with respect to the OLS full sample and doubled with respect to that in the OLS sibling sample. The mother fixed-effect result implies that children who are exposed to a 1 SD increase in violence in the third trimester experience a 0.1 SD decline in HAZ.

How do these estimates compare to those in the literature? Previous studies have identified that children exposed to war have 0.2–0.5 SD lower HAZ (Akresh et al., 2012, Bundervoet et al., 2009). Two factors could help explain the difference in magnitude between my estimates and those in the literature. First, previous studies have focused on massive violence episodes such as civil wars or genocides that usually last a few years and are highly disruptive. The Colombian conflict distinguishes from the rest by its long duration and low intensity, in which civilians have, to some extent, learned how to live under the threat of armed groups. Second, most previous research has analyzed these impacts in African countries, in which children, even in the absence of wars or high violence, start from a lower nutritional baseline (Victora & de Onis, 2010), making them more vulnerable to adverse environmental conditions. Compared to studies on the effects of adverse shocks in early life, Rosales (2013) found that children in Ecuador exposed to the 1998 El Niño weather shock experienced an average decline in HAZ of 0.09 SD and the negative effect came from exposure during the third trimester in-utero and in the first year of life.

Using previous estimates on the relationship between height and labor market productivity (Vogl, 2014), I find that the decline in child's height could be translated into a future 0.6% to 1.6% decrease in hourly wages.

5.3. Effects of violence on cognitive outcomes

Table 3 shows estimates on cognitive test scores using Eq. (1). Results indicate that the negative effect of terrorist attacks are particularly concentrated during the first and third trimesters of pregnancy and in the first years of life. As violence increases by a SD, PPVT falls by 0.07 SD if the increase in violence occurs in the first trimester. Similarly, a 1 SD rise in massacres in early childhood leads to a 0.19 SD decline. Consistent with these results, I find that math reasoning and general knowledge are particularly affected due to exposures in the first trimester and in early childhood, and, as in the case of child's health, violence does not seem to play a major role on cognitive outcomes after age 3. The decline in math reasoning ranges between 0.04 and 0.25 SD and in general knowledge ranges between 0.03 and 0.25 SD.

Table 3.

The effect of violence on children's cognitive development.

	PPVT	Math reasoning	General knowledge
	(1)	(2)	(3)
2 Years before conception	0.0025	0.0001	0.0023
	[0.0018]	[0.0018]	[0.0019]
1 Year before conception	−0.0024	−0.0011	0.0016
	[0.0020]	[0.0015]	[0.0013]
Trimester 1	−0.0068***	−0.0040***	−0.0059***
	[0.0022]	[0.0014]	[0.0008]
Trimester 2	−0.0034	−0.0016	−0.0029*
	[0.0024]	[0.0022]	[0.0016]
Trimester 3	−0.0046*	−0.0009	−0.0035***
	[0.0024]	[0.0012]	[0.0010]
Childhood 0–3	−0.0012**	−0.0016***	−0.0016***
	[0.0005]	[0.0004]	[0.0005]
Childhood 3+	0.0001	−0.0003	0.0001
	[0.0006]	[0.0006]	[0.0007]
N	4666	4645	4700

Note: Sample includes all children 3–7. Please refer to Section 4 for details.

^***$p<0.01$.

^**$p<0.05$.

^*$p<0.1$.

Columns 3–8 in Table 5 show estimates of violence on the sample of siblings using Eqs. (1), (2). As in the case of child's nutrition, results show a consistent story between the OLS sibling and full samples and the mother fixed-effects model, and for some cases, the magnitude of the coefficients becomes larger. For instance, the findings on math reasoning (columns 3 and 4) suggest that this outcome significantly declines with exposure to massacres during the first and second trimesters, as well as with exposure to violence during the whole childhood period, an impact that was only observed in the first trimester and during the first three years of life in the full and sibling OLS samples. For PPVT and General knowledge, while I find some effects during the in-utero period that are similar in magnitude to those shown in Table 3, the impacts of violence in the first trimester and in early childhood are no longer statistically significant (although the effect on PPVT early in pregnancy is similar in size in both specifications). Interestingly, the magnitude of the effect of massacre exposure in the third trimester is reinforced once I control for mother fixed-effects. This pattern is also observed in columns 1 and 2. I come back to this point when I analyze potential sources of selection bias.

While it is difficult to compare my estimates to those found in previous research given the very few number of papers exploring this relationship and the fact that these studies have mostly focused on older individuals or on children in wealthier countries, it seems that my coefficients lie in the lower range of those in the literature. For instance, Sharkey et al. (2012) showed impacts of 0.3 SD in PPVT among 4-year-olds in Chicago and Rodriguez and Sánchez (2013) found a decline of almost 1 SD on high-school-graduates test takers in Colombia.

Following the discussion in Alderman, Behrman, Ross, and Sabot (1996), I find that the decline in child's cognitive test scores due to violence would lead to a 0.7–5% decrease in future hourly wages.

5.4. Effects of violence on socio-emotional outcomes

I find little effects of violence on socio-emotional outcomes using OLS or mother fixed-effects models as shown in Table 4, Table 6; however, the sign of the coefficients tend to indicate “worse” outcomes as violence increases. For instance, it is worth noting the increase in magnitude of the coefficients during the childhood years once models account for mother fixed-effects. I am not aware of any study exploring the effects of violence on non-cognitive outcomes in a developing country. In the Chicago paper cited above, Sharkey et al. (2012) also explored the effects of homicides on children's non-cognitive abilities, and found negative effects on attention and impulse control of a third of a SD. It is possible that the smaller (and not statistically significant) effects that I find here, may be due to the fact that I consider a more “indirect violence exposure” than that observed by the authors. Moreover, the socio-emotional measures in the HCB data – indicators of a child's behavior during peer-play –, may not necessarily capture certain aspects of a child's socio-emotional potential that could be affected by violence.

Table 4.

The effect of violence on children's socio-emotional development.

	Aggression	Isolation	Adequate interaction
	(1)	(2)	(3)
2 Years before conception	−0.0026	0.0021	0.0027
	[0.0016]	[0.0026]	[0.0022]
1 Year before conception	−0.0014	−0.0026	0.0044
	[0.0014]	[0.0022]	[0.0029]
Trimester 1	−0.0003	0.0027*	−0.0031
	[0.0021]	[0.0016]	[0.0029]
Trimester 2	−0.0005	0.0021	0.0007
	[0.0023]	[0.0018]	[0.0015]
Trimester 3	0.0025	0.0017	−0.0022
	[0.0022]	[0.0012]	[0.0020]
Childhood 0–3	−0.0003	0.0003	−0.0007
	[0.0006]	[0.0003]	[0.0004]
Childhood 3+	0.0008	0.0001	−0.0006
	[0.0005]	[0.0008]	[0.0008]
N	4613	4613	4613

Note: Sample includes all children 3–7. Please refer to Section 4 for details. ^***p<0.01, **p<0.05, *p<0.1.

Table 6.

Mother fixed-effects estimates of violence on children's socio-emotional development dataset.

	Aggression		Isolation		Adequate Interaction
	OLS sibling sample	Mother FE	OLS sibling sample	Mother FE	OLS sibling sample	Mother FE
	(1)	(2)	(3)	(4)	(5)	(6)
TRIM 1	−0.0033	−0.0063	−0.0005	0.007	−0.0057	−0.0117
	[0.0038]	[0.0185]	[0.0052]	[0.0154]	[0.0059]	[0.0167]
TRIM 2	−0.0100**	−0.0177	0.0055	−0.0062	−0.0025	−0.0072
	[0.0037]	[0.0215]	[0.0060]	[0.0180]	[0.0058]	[0.0195]
TRIM 3	0.0100*	−0.0428	−0.0031	−0.0265	−0.0047	0.0089
	[0.0059]	[0.0270]	[0.0054]	[0.0224]	[0.0058]	[0.0243]
0–3	−0.0004	0.0032	0.0004	0.0072*	−0.0006	−0.0046
	[0.0008]	[0.0047]	[0.0009]	[0.0040]	[0.0015]	[0.0044]
3+	0.0007	−0.0013	−0.0011	0.0044	−0.0016	−0.0034
	[0.0013]	[0.0064]	[0.0018]	[0.0053]	[0.0021]	[0.0058]
N	723	723	723	723	723	723

Note: Sample includes all siblings 3–7 years of age. Please refer to Section 4 for details. ^***p<0.01, **p<0.05, *p<0.1.

5.5. Further evidence: effects of violence on fetal health using natality data

I examine whether in-utero massacre exposure provides estimates consistent with those in the literature. I use data from Vital Statistics Birth Records since HCB does not provide information on birth outcomes. Table 7 shows a small but negative effect of violence on birth weight and an increase in the probability of being low birth weight, both effects occurring in the first trimester of pregnancy. These results provide some evidence consistent with the idea that maternal stress could be a relevant channel through which violence impacts children (Aizer et al., 2012, Denckel-Schetter, 2011). Lastly, I do not find evidence that violence is associated with changes in the probability of being preterm.

Table 7.

The effect of violence on health at birth using birth records.

	Birth weight	Low birth weight	Preterm
	(1)	(2)	(3)
TRIMESTER 1	−0.1435**	0.0001***	−0.0002
	[0.0651]	[0.00002]	[0.0001]
TRIMESTER 2	0.0816	0.00003	0.0003
	[0.1219]	[0.00002]	[0.0002]
TRIMESTER 3	−0.0335	0.0001	0.0000
	[0.1124]	[0.00004]	[0.0001]
N	3,873,065	3,289,309	3,784,338

Note: Pooled data from Vital Statistics, years 1998–2001 and 2005–2006. Mean birth weight, low birth weight ($<2500\mathrm{g}$), and preterm are 3155 g, 8%, 13%, respectively. In addition to basic covariates (see Section 4), models include indicators for multiple births parity, urban household, baby delivered at a hospital, and mother had medical insurance (public, private, other). ^*$p<0.1$.

^***$p<0.01$.

^**$p<0.05$.

6. Robustness

6.1. Potential sources of selection bias

A complicating factor in the study of the impacts of violence on child outcomes is that violence may not only have a scarring effect on affected cohorts, but may also induce selection through sorting, migration, fertility, or mortality (Almond, 2006, Bozzoli et al., 2009). In this section, I carefully explore these potential sources of selection bias.

6.1.1. Geographic sorting

Families living in conflict-prone municipalities may also be disadvantaged in other dimensions (e.g., they may be less educated). To test for selective sorting within Colombian municipalities, I explore the association between mother observable characteristics and the level of violence that they experienced during pregnancy and when their child was in childhood. In the presence of selective sorting, this association should remain even after accounting for municipality fixed effects. Results are shown in Table B1.

The first panel of Table B1 shows that there are some observable differences between mothers who are exposed to different levels of violence in their child's early life. For example, column 3 suggests that as violence increases in a child's first trimester while in-utero, the proportion of mothers with less than high school education increases. Column 3 also shows that violence in the third trimester is associated with a higher proportion of educated mothers (i.e., with high school or more). After controlling for municipality fixed-effects, the bottom panel shows that some of these significant differences are removed (e.g., the coefficient on the third trimester is no longer statistically significant); however, the size of the associations have actually increased. For example, the coefficient on the first trimester has more than quadrupled and the third trimester effect size has more than doubled. If this relationship is indicative and consistent with the type of selection that exists on unobserved characteristics (e.g., more educated mothers self-selecting in the third trimester as violence increases), a maternal fixed effect model should cause the third trimester effect to become larger in the direction of an adverse relationship. In Table 5, that is precisely what is observed for HAZ, PPVT, Math reasoning, and General knowledge. Hence, in order to appropriately identify the effect of violence on children, it is important to control for maternal fixed-effects that eliminates the time-invariant component of the unobserved heterogeneity that exists between mothers that are facing differing levels of violence during pregnancy.

6.1.2. Migration

Migration can be an important concern for my empirical analysis if households who migrate in relation to massacres differ from those who do not, in dimensions that affect child development. I perform three analyses on endogenous migration. First, I explore the factors that motivate families to migrate. Table B2 shows that (12%) of households migrate due to family reasons followed by violence (1.8%) and looking for better conditions (0.26%). The small number of cases reporting violence provides the first piece of evidence that selective mobility is likely to be low in the HCB sample. Second, I compare the characteristics of migrant and non-migrant families. Table B3 shows that those who move are actually more disadvantaged than those who stay. For example, children in migrant families have mothers who are younger, less educated, and less likely to be married, and they have “worse” developmental outcomes themselves. Third, I examine the effects of violence on outcomes across movers and non-movers. If selective migration is likely to be an important issue in my analyses, I should find that the impacts of violence on child outcomes should differ across subsamples. Table B4 shows this is actually not the case; the estimates of violence are almost identical in the two cases. (Table 6).

6.1.3. Fertility

Violence may also affect fertility decisions in terms of family size or timing of pregnancy. To test for selective fertility, I examine whether massacres are associated with a woman's fertility decisions across her observable characteristics. I employ Demographic and Health Survey (DHS) data instead of the HCB data to perform these tests, since HCB do not include information on fertility. In particular, I focus on the number of children that have been born after a given child, and the timing of fertility is measured using both the succeeding birth interval (in months) after a given child is born and the preceding birth interval (in months) before a given child is born.

Results in Table B5 report that the associations between violence in-utero and up to age three years, interacted with mother characteristics, are weakly associated with total fertility and timing of fertility, offering little evidence on selective fertility.

6.1.4. Survival

The estimates of early-life shocks may also be affected by selection on survival both at birth and during childhood: Violence is likely to increase the chances of dying for those with weaker health endowment (see, for example, (Almond, 2006)). To test how massacres affect survival, I provide evidence on how changes in violence affect a child's mortality in the first month, first year, and first three years of life using DHS data.

Table B6 shows that neither the associations between violence and child's mortality nor the interactions between violence and mothers' characteristics are statistically significant. Consistent with this finding that there is little evidence on selective survival, results in Table 7 showed that violence is not associated with the probability of being preterm.

6.2. Additional robustness checks

I perform additional robustness checks that are shown in the Online Appendix. In particular, I examine potential threats to the identifying assumption such as serial correlation in massacres and potential confounders of violence.

7. Parental investments

7.1. Background

After documenting the negative effects of violence on children, I explore changes in parental investments. A priori, it is ambiguous whether parents respond to violence by increasing, decreasing, or keeping their investments on children constant. Economic theory provides competing hypotheses on how parents make investments based on a child's endowment (e.g., birth weight). If parents are motivated by maximizing the returns of their investments, resources are invested in their high-ability child (Becker & Tomes, 1986), whereas if parents seek to equalize outcomes across their children, resources are directed to their less-able child (Behrman, Pollak, & Taubman, 1982). Furthermore, parents’ behavior could be affected by violence per se, inducing high levels of stress and imposing a number of constraints on the family. Studies have found evidence consistent with parents trying to equalize (Breining, Daysal, Simonsen, & Trandafir, 2015), reinforce (Almond and Mazumder, 2013, Advharyu and Nydshadham, 2014), or compensate investments across children (Aizer & Cunha, 2014). In this study, I empirically test the link between violence and parenting. While some studies have found an association between local violence and parental distress, and have suggested that parental responses may be a likely pathway by which local violence affects young children (Sharkey et al., 2012), the link between violence and measures of parenting has not been directly tested in the literature.

7.2. The effects of violence on parenting

I focus on the quantity and quality of parenting (Table 8 shows descriptive statistics): Parenting quantity is a mother self-report on the amount of time (hours) she spends with her child in a given day. Parenting quantity is a mother self-report on how often she spends time with her child in the following activities: (i) Personal care routines: keeping the child safe, fed, clothed, and sheltered; (ii) Active stimulation routines: reading books to the child, talking with the child, etc.; and (iii), Physical and psychological aggression: adopting physical (hitting or spanking, etc.) or psychological (shouting, scaring, etc.) actions against her child. All these measures have been rescaled to have mean zero and SD one.

Table 8.

Descriptive statistics for parental investments by violence exposure.

	Full sample	No violence	Violence
			Low	High
	(1)	(2)	(3)	(4)
Quantity of maternal time [hrs/wk]***	33.47	34.77	34.90	31.23
	[18.62]	[18.09]	[18.01]	[19.31]
Quality of maternal time:
Personal care routines	0.00	0.03	−0.08	0.06
Active stimulation routines***	0.00	−0.04	−0.04	0.07
Physical aggression***	0.00	−0.05	−0.04	−0.02
Psychological aggression***	0.00	−0.02	−0.05	0.04
N	13,444	2888	5115	5341

Note: Sample includes all children 3–7. Please see 5.1, 7.2 for details. ***p<0.01, **p<0.05, *p<0.1.

Table 9 shows the effects of violence on parenting using the OLS and maternal fixed-effects models. Results in the top panel (OLS), suggest little evidence that changes in violence are associated with changes in parenting except for a small decline in the frequency of personal care routines in early childhood. A potential explanation for this result could be that, changes in parental responses are not necessarily persistent over time. Parents may respond right after the shock occurs but after a certain time, the quantity and quality of their investments may return back to their initial level. Unfortunately the lack of longitudinal data limits the scope of the paper to analyze this issue in more detail.

Table 9.

The effect of violence on parenting.

	Time use mother w/child (hrs)	Personal care routines	Active stimulation routines	Physical aggression	Psychological aggression
	(1)	(2)	(3)	(4)	(5)
Without mother fixed-effects
IN-UTERO	0.0042	0.0001	−0.0004	−0.0000	0.0001
	[0.0103]	[0.0010]	[0.0010]	[0.0007]	[0.0005]
CHILDHOOD 0–3	0.0009	−0.0006*	−0.0002	−0.0001	0.0001
	[0.0052]	[0.0003]	[0.0003]	[0.0002]	[0.0001]
CHILDHOOD 3+	0.0006	0.0002	0.0001	0.0001	−0.0000
	[0.0039]	[0.0002]	[0.0002]	[0.0002]	[0.0002]
With mother fixed-effects
IN-UTERO	0.0024	0.0014	−0.0027	0.0052*	0.0029
	[0.0237]	[0.0023]	[0.0021]	[0.0029]	[0.0026]
CHILDHOOD 0–3	−0.0032	−0.0007*	−0.0002	−0.0002	0.0002
	[0.0043]	[0.0004]	[0.0004]	[0.0005]	[0.0005]
CHILDHOOD 3+	−0.0024	0.0004	−0.0001	0.0011**	0.0002
	[0.0036]	[0.0004]	[0.0003]	[0.0004]	[0.0004]
N	13,077	13,344	13,344	13,344	13,344

Note: Sample includes all children 3–7. Please see 4, 7.2 for details. ^***$p<0.01$.

^**$p<0.05$.

^*$p<0.1$.

The second panel shows estimates after controlling for a mother's fixed-effect. I find that the incidence of physical aggression increases and the decline on personal care routines remains. Moreover, as in the case of child outcomes, the size of the coefficients increase (becomes more negative) for almost all measures of parenting, suggesting that as violence increases the relationship mother-child becomes harsher. These results provide some suggestive evidence that is consistent with the hypothesis that parents reinforce the negative violence shock by providing less nurturing parenting or less time investments to their child. This result is in line with recent empirical evidence on how parental investments respond to health endowment at birth, at least in developing countries (Almond & Mazumder, 2013).

8. Concluding remarks

This study contributes to a growing body of research on the effects of early life conditions on human capital, by providing new evidence on the effects of a particular pervasive shock affecting millions across the world: terrorist attacks.

Results show that the occurrence of violent episodes in “sensitive periods” in a child's life can lead to health and cognitive declines. The paper showed that estimating mother fixed effects models that account for time-fixed characteristics of mothers that tend to conceive in different moments, provides robust evidence on the effect of violence on child outcomes. In particular, the findings from the mother fixed-effects models suggest that as violence increases by a standard deviation, HAZ and cognitive tests scores decline by at least 0.1 SD when the shock occurred in pregnancy. Moreover, while this paper does not directly measure stress, the results are consistent with stress being an important mechanism driving these effects. In terms of parenting, my findings show some evidence that parents may be reinforcing the negative effects of the shock by being more aggressive with their child. However, the lack of longitudinal data limits the paper's capacity to provide more conclusive evidence on this dimension. To my knowledge, little research has examined parenting in the context of community violence.

Given the long duration of the Colombian conflict, many cohorts have to some extent been exposed to high levels of stress at some point in their lives. One might therefore expect that, as the estimates found in the paper are quantitatively relevant, the costs of violence on human capital may go far beyond those effects.

My results suggest important implications for public policy that would complement government's efforts in reducing violence. For instance, social programs targeting vulnerable children should screen on violence exposure as it represents a relevant marker of risk and inequality. Lastly, an important question for future research is whether there is potential for remediation: could social policies help mitigate the negative effects of these types of adverse environments? While providing causal evidence in this respect is challenging, exploiting different sources of exogenous variation in how programs are allocated to certain groups could be a promising start.