Armed Conflict and Adolescents’ Early Transition to Childbearing

Abstract

We estimate the effect of armed conflict on adolescents’ childbearing transition. Three types of models are estimated: models in origin, which determine the current effect of violence; context change models, which estimate the effect of decreased violence levels; and violence aftermath models, which indicate the groups that do not completely adapt to a peaceful context. Through multilevel logistic models, we find that the coefficient of violence depends on the municipality in where adolescents reside or resided. We conclude that violence promotes the transition to childbearing particularly among rural females. However, a reduction in violence levels reduces the probabilities of childbearing.

1. Introduction

This is the first research on Colombia to study the relationship between exposure to armed conflict and teenage motherhood, and worldwide it is the first to study the assimilation from violent to peaceful contexts regarding this transition. The causes of early motherhood have been studied widely. Several authors conclude that belonging to dysfunctional families or single parent households and growing up in socio-economic and cultural contexts that do not offer opportunities for personal accomplishment increase the risks of dropping out of school and early parenthood (Flórez et al., 2004; Jaffe, Moffitt, Caspi, & Taylor, 2003; Ribero, 2001). However, there are other aspects that affect adolescents’ decisions about prematurely facing motherhood that remain relatively understudied. One of them is the exposure to the contextual violence generated by an armed conflict.

We define contextual violence in the case of Colombia armed conflict as exposure to acts of violence committed by the guerrillas (The Colombian Revolutionary Armed Forces [FARC] and The National Liberation Army [ELN]) or paramilitary groups (Colombian United Self-defences [AUC]); hence, every person who lives in areas where these armed groups frequently commit violent acts is considered a victim of contextual violence. 1 The objectives of this research are: 1) to study the association between violent contexts with armed conflict and adolescents’ decisions about childbearing; 2) to determine whether or not a decrease in violence levels positively modifies the behaviour of adolescents in relation to making this transition prematurely; and 3) to identify if the relationship between contextual violence and adolescents’ behaviour differs according to adolescents’ characteristics.

Most studies focus on the effect of violent conflicts on the fertility behaviour of the entire population and not on adolescents’ behaviour. Verwimp and Bavel (2005) find that Rwandan refugee women have higher fertility rates than other population groups. Rutayisire, Hooimeijer, and Broekhuis (2014) show that fertility increased in Rwanda between 1992 and 2000 due to replacement fertility and an unmet need for family planning despite a decrease in the percentage of married women. In contrast, Shemyakina (2007) finds that in Tajikistan there is a negative effect of living in armed conflict areas on the age at first birth. She explains that the fertility rate might decrease during an armed conflict because families postpone having children, child mortality increases, spouses often separate during a war and contraception methods are used.

However, the sexual behaviour of adolescents tends to be riskier than that of adults, especially among migrants as internal displaced persons (IDPs). For adult migrants, Lindstrom (2003) and Chattopadhyay, White, and Debpuur (2006) find respectively that 1) rural women rapidly adopt urban fertility practices after migrating to urban areas and 2) delay the fertility timing of second and higher-order births. In contrast, Stack (1994), Baumer and South (2001), and Erulkar and Ferede (2009) find that migration and/or residential mobility increases the likelihood of adolescents engaging in premarital sex and of having multiple sexual partners. Therefore, it is possible that, because of their vulnerability and sense of responsibility, adolescents do not follow adults’ behaviour; instead, they could be more likely to exhibit higher fertility rates.

Some authors hypothesise about the channels through which violence may affect adolescents’ sexual behaviour. Upchurch, Aneshensel, Sucoff, and Levy-Storms (1999), Kirby and Lepore (2005), Harding (2009), and Uthman (2010) find that community social disorganisation (measured with the rates of substance abuse, violence, divorce, and so on) is an important predictor of high-risk sexual behaviour. Harding (2009) indicates that the probability of engaging in risky behaviours, like unprotected sex, increases when the environment does not offer a long life expectancy, as occurs during an armed conflict. In addition, armed conflict might cause an increase in cases of rape, sexual humiliation, forced prostitution, and forced pregnancy (Woroniuk, 2000). In the case of Colombia, displaced children and adolescents also mature earlier because of the need to find resources for their own livelihood or that of their families (Pfizenmaier, 2004). Sanchez-Cespedes (2017) finds that the odds of living without both parents for children and teenagers from violent areas is almost twice the odds for their peers from peaceful areas. Therefore, armed conflict promotes that adolescents become independent, taking up work, leaving their studies and forming their own families.

In addition to studying the effect of exposure to armed conflict on teenage childbearing, the main contribution of the study is the analysis of adolescents’ behaviour change when they stop experiencing contextual violence. The strategy followed to simulate the passage from a violent to a peaceful context is to use data concerning migrants. Portes and Zhou (1993) state that migrants’ assimilation to a new context depends on the characteristics of the native group that receives them in the destination, which can be considered a kind of peer effect. IDPs are commonly poor (94.4% of them live below the poverty line; Ibañez-Londoño, 2008); therefore, they settle in poor areas of receiving municipalities. Thus, the context after migration might encourage marriage and motherhood. The behaviour of poor and extremely poor adolescents in Bogota and Cali, the first and the third destination of IDPs, is described by Flórez et al. (2004). They find that adolescents from the lower strata start to engage in sexual intercourse and form families earlier than their peers from higher strata. They also find that the use of contraception in low strata starts after adolescents have begun their sexual life. They conclude that, in low socio-economic and cultural contexts, pregnancy enables young people to gain status, recognition, and social acceptance.

The study defines early transition to childbearing as a woman experiencing this transition before achieving sufficient physical and cognitive development. As it is difficult to establish an exact age when physical development and cognitive development are reached, two samples with different age ranges are considered: adolescents aged 12–20 and adolescents aged 12–24. According to Dixon-Muller (2008), the complete physical maturity of rural adolescents in Latin America is achieved between their 18^th and their 20^th birthday, about six years after their first period. This makes becoming a mother a premature event when it is experienced before the age of 20. In addition, some authors indicate that the problem of early sexual transition and parenthood is related more to adolescents’ cognitive immaturity than to their bodily immaturity (Barker, 2000). In fact, some authors claim that cognitive maturity is reached one decade after puberty or more, usually after the age of 24 (Breinbauer, Maddaleno, & American, 2005).

The analysis is based on 2006‒2009 data from Sisben, which is the official database used to target beneficiaries for social programmes in Colombia. Therefore, the study population includes persons aged 12–24 who have been classified as poor or extremely poor by the Colombian Government and who were living in municipalities with fewer than 100,000 inhabitants in the period 2006‒2009. These municipalities were chosen to guarantee a homogeneous impact of armed conflict on adolescents. Violence is measured with an index of violence and conflict (IVC) at the municipal level designed by the author. This index is explained in Section 2.

Multilevel logistic models are used to determine the effect of violence on female adolescents’ decisions concerning motherhood. To establish the effect of violence on the adolescents who are currently affected by it, we estimate models in origin. To identify the characteristics of the adolescents who change their behaviour because the level of violence decreases, the study estimates context change models. Finally, to identify the features of the adolescents who, after migrating from violent to peaceful destinations, do not completely adapt to the new context, the study estimates violence aftermath models. The three types of models are explained in Section 3.

The main conclusions of the study are as follows. (1) Exposure to armed conflict accelerates the transition to childbearing particularly among rural adolescents. (2) The context change models show that a reduction in violence levels decreases the probability of childbearing among those adolescents who experienced high levels of contextual violence in the past. (3) The violence aftermath models show that females who were poor before migration do not completely adapt to a peaceful context and even might exhibit probabilities of childbearing higher than extremely poor adolescents, who generally have higher probabilities in peaceful circumstances.

The remainder of the paper is structured as follows. Section 2 explains the concept of contextual violence and the estimation of the IVC. The third section explains the three types of models: models in origin, context change models, and violence aftermath models. Section 4 introduces the Sisben database. Section 5 presents the descriptive statistics of the dependent and independent variables. The sixth section discusses possible limitations of the study. Section 7 contains the results and their analysis. The final section concludes.

2. Contextual violence in the case of Colombia armed conflict

2.1. Definition

The definition of contextual violence is based on definitions of community violence given by other authors (for instance, the National Center for Children Exposed to Violence [NCCEV], 2003). ¹ We define contextual violence as exposure to acts of violence committed by the guerrillas (FARC and ELN) and paramilitary groups (AUC). Therefore, not only direct victims of these groups are considered victims of the contextual violence generated by them; but every person who lives in areas where armed groups frequently commit violent acts is also considered a victim of contextual violence. For example, it is common that entire families migrate to avoid forced recruitment of their children (Salazar, 2001). These families are victims of the contextual violence generated by armed groups, although none of their members have suffered a direct act.

2.2. Index of violence and conflict (IVC)

Contextual violence caused by armed conflict is measured through an index that we estimate and call the index of violence and conflict. This index is the predicted score of a logistic model, whose dependent variable is one if the municipal displacement rate in 2005 is higher than the average and zero otherwise. The advantages of logistic models over other methods to estimate indices are explained in Supplementary Materials 1. We calculate the municipal displacement rate with the following formula:

$displacementrat{e}_{i2005}=\frac{M_{i2005}}{p_i}\ast 100$

$p_i=\frac{p_{i2004}^{\hspace{0.17em}}+p_{i2005}^{\hspace{0.17em}}}{2}$

where $M_{i2005}$ refers to all outdisplaced migrants from municipality i in 2005. $p_{i2004}^{\hspace{0.17em}}$ and $p_{i2005}^{\hspace{0.17em}}$ are the populations of i in the years 2004 and 2005, respectively. $p_i$ is the average between these two populations. The independent variables of the logistic model describe the armed conflict in 2004; therefore, there is no double causality with the dependent variable. The variables of Table A1 in Supplementary Materials 1 included in the model are those that occur in at least 5 per cent of the municipalities. Table 1 presents the result of the logistic model or IVC. Figure A1 in Supplementary Materials 1 shows the descriptive statistics of the index and its distribution across the Colombian territory.

Table 1. Index of violence and conflict

$\mathrm{D}\mathrm{i}\mathrm{s}\mathrm{p}\mathrm{l}\mathrm{a}\mathrm{c}\mathrm{e}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{r}\mathrm{a}\mathrm{t}{\mathrm{e}}_{\mathrm{i}2005}$	Coef.		Std Err.	Z	P > z
Confrontations	0.605	***	0.216	2.800	0.005
Armed contact	0.274	***	0.040	6.900	0.000
Total terrorist acts	0.160	**	0.076	2.120	0.034
Homicides and kidnappings per 10,000 inhabitants	0.173	**	0.081	2.140	0.033
Assaults on private property per 10,000 inhabitants	0.580	***	0.212	2.740	0.006
Ambush	0.036		0.291	0.120	0.902
Harassment	0.099		0.121	0.820	0.410
Cons	−1.874	***	0.100	−18.760	0.000

Source: Universidad de los Andes – CEDE and Observatory of the Presidential Programme for Human Rights and International Humanitarian Law.

3. Method

The probability of premature transition to childbearing is modelled using multilevel logistic models because a female adolescent in an armed conflict area interact into a context in which other adolescents have also been influenced by the violence of armed groups. Thus, the effect of a violent context on the probability of childbearing of an adolescent depends on the effect of violence on her peers in her municipality. Hence, we have a hierarchical system with two levels: adolescent and the group of adolescents with whom she interacts. For feasibility reasons we define these two levels as: adolescents and municipalities; thus adolescents are nested within municipalities. In the empirical exercise we only include municipalities with fewer than 100,000 inhabitants, as this makes it more likely that people are affected similarly by the violence generated by armed conflict. Furthermore, there is no information about armed conflict for geographical units smaller than municipalities.

In hierarchical problems, like the one studied in this document, people from the same geographical area are more similar to one another than people from different geographical areas are; this leads to estimates with standard errors too small, obtaining spurious significant values. Multilevel models solve this problem, and additionally allow analysis of variables from different levels simultaneously considering the dependency between them (Hox, 2010). Another advantage of multilevel models is that we can assume that the intercept and the coefficients of the variables vary across groups to consider their heterogeneity, in this cases across municipalities. Hence, the intercept and coefficients have a distribution with mean and variance. We estimate the models using the GLIMMIX procedure of Statistical Analysis Software (SAS).

The models in origin are estimated for all female adolescents in 2009 who have not migrated in the last three years. The ideal population would be nonmigrant females since migration itself increases the likelihood of childbearing. This is confirmed in Supplementary Materials 2 with the total population of adolescents in Bogota in 2009; the odds for migrants is 24 times the odds for nonmigrants for the 12–20-age-group and 30 times for the 12–24-age-group. The context change and violence aftermath models do not have this problem, because their study population consists of migrants. To identify them, the Sisben databases are merged in two consecutive years and 2009, specifically 2006, 2007, and 2009 and 2007, 2008, and 2009. The details of this procedure are in the Supplementary Materials 3.

3.1. Models in origin

The purpose of the models in origin is to determine the effect of armed conflict on the transition to childbearing of adolescents who are currently exposed to it. The general two-level logistic regression model is the following:

$\begin{array}{rl}& Y_{ij}^{\hspace{0.17em}}=log\left(\frac{{\pi }_{ij}}{1-{\pi }_{ij}}\right)\\ & ={\beta }_{0j}^{\hspace{0.17em}}+{\beta }_{1j}^{\hspace{0.17em}}ag{e}_{ij}+{\beta }_{2j}^{\hspace{0.17em}}Bothparent{s}_{ij}\\ & +{\beta }_{3j}^{\hspace{0.17em}}Educationoftheheadofthehousehol{d}_{ij}+{\beta }_{4j}^{\hspace{0.17em}}Sisben{1}_{ij}\\ & +{\beta }_{5j}^{\hspace{0.17em}}Rura{l}_{ij}+{\epsilon }_{ij}\\ & {\beta }_{kj}^{\hspace{0.17em}}={\gamma }_{k0}^{\hspace{0.17em}}+{\gamma }_{k1}^{\hspace{0.17em}}Indexofviolenc{e}_j^{\hspace{0.17em}}+u_{kj}\end{array}$

where i represents the individual, j the municipality of residence, and k the individual variable. $Y_{ij}^{\hspace{0.17em}}$is one if adolescent i who lives in municipality j has assumed motherhood and it is zero otherwise. When each ${\beta }_{kj}^{\hspace{0.17em}}$ is replaced in the equation of $Y_{ij}^{\hspace{0.17em}}$, the final equation is:

(1) $\begin{array}{rl}& Y_{ij}^{\hspace{0.17em}}=log\left(\frac{{\pi }_{ij}}{1-{\pi }_{ij}}\right)={\gamma }_{00}+{\gamma }_{01}IV{C}_j+{\gamma }_{10}{a}_{ij}+{\gamma }_{20}b+{\gamma }_{30}{e}_{ij}+{\gamma }_{40}{s}_{ij}+{\gamma }_{50}{r}_{ij}\\ & +{\gamma }_{11}IV{C}_j{a}_{ij}+{\gamma }_{21}IV{C}_j{b}_{ij}+{\gamma }_{31}IV{C}_j{e}_{ij}+{\gamma }_{41}IV{C}_j{s}_{ij}+{\gamma }_{51}IV{C}_j{r}_{ij}+(u_{0j}+u_{1j}{a}_{ij}+u_{2j}b\\ & +u_{3j}{e}_{ij}+u_{4j}{s}_{ij}+u_{5j}{r}_{ij}+{\epsilon }_{ij})\end{array}$(1)

where $a_{ij}$ is the age of female i who lives in municipality j,$b_{ij}$ is one when she lives with both parents and zero otherwise, $e_{ij}$ is the number of years of education of the head of the household, $s_{ij}$ is one if she is classified as extremely poor by the Colombian government and zero if she is classified as poor (this is explained in more detail in Section 4), $r_{ij}$ is the degree of urbanicity of the area of residence (rural is one and urban is zero), and $IV{C}_j$ is the index of violence and conflict of municipality of residence j. The part outside the parentheses is the deterministic part, and the part inside the parentheses is the stochastic part. The interaction terms between the individual variables and the IVC express the moderator effect of violence on the relationship between the dependent variable $Y_{ij}$ and the individual predictors. Equation (1)(1) $\begin{array}{rl}& Y_{ij}^{\hspace{0.17em}}=log\left(\frac{{\pi }_{ij}}{1-{\pi }_{ij}}\right)={\gamma }_{00}+{\gamma }_{01}IV{C}_j+{\gamma }_{10}{a}_{ij}+{\gamma }_{20}b+{\gamma }_{30}{e}_{ij}+{\gamma }_{40}{s}_{ij}+{\gamma }_{50}{r}_{ij}\\ & +{\gamma }_{11}IV{C}_j{a}_{ij}+{\gamma }_{21}IV{C}_j{b}_{ij}+{\gamma }_{31}IV{C}_j{e}_{ij}+{\gamma }_{41}IV{C}_j{s}_{ij}+{\gamma }_{51}IV{C}_j{r}_{ij}+(u_{0j}+u_{1j}{a}_{ij}+u_{2j}b\\ & +u_{3j}{e}_{ij}+u_{4j}{s}_{ij}+u_{5j}{r}_{ij}+{\epsilon }_{ij})\end{array}$(1) can also be expressed in the following way:

(2) $Y_{ij}^{\hspace{0.17em}}={\gamma }_{00}^{\hspace{0.17em}}+{\gamma }_{01}IV{C}_j+\sum _{k=1}^K{\gamma }_{k0}{X}_{kij}+\sum _{k=1}^K{\gamma }_{k1}{X}_{kij}IV{C}_j+\sum _{k=1}^K{u}_{kj}{X}_{kij}+u_{0j}+{\epsilon }_{ij}$(2)

Where $IV{C}_j^{\hspace{0.17em}}$ represents the IVC in the municipality of residence j. $X_{kij}$ is the independent variable k for individual i who lives in municipality j. K is the total number of individual variables: five. $u_{0j}$ and $u_{kj}$ are the stochastic parts of the intercept and variable k, respectively.

3.2. Context change models

The context change models aid in determining the effect of a decrease in the intensity of violence on adolescents’ behaviour. The estimates only include migrant adolescents who come from municipalities with high levels of violence to ensure homogeneous conditions before migration. To identify these municipalities, all of the municipalities are classified into three levels of violence – low, medium, and high – by applying the method of K-means clustering to the IVC (see Figure A1 for the descriptive statistics of each level). The study assumes that all the adolescents from the municipalities classified in the high level (95 of 1102 municipalities) were exposed to similar violent conditions before migration. After migration these adolescents started experiencing different levels of violence, some of them continuing to live in violent municipalities and others starting to inhabit peaceful ones. To be included in the regression, each adolescent must not have experienced childbearing before arriving at the destination. The final equation for the two-level logistic regression model is:

$\begin{array}{rl}& Y_{ij}^D=log\left(\frac{{\pi }_{ij}}{1-{\pi }_{ij}}\right)=\\ & {\beta }_{0j}^D+{\beta }_{1j}^{D}ag{e}_{ij}+{\beta }_{2j}^{D}Bothparents{ }_{ij}+{\beta }_{3j}^{D}Educationoftheheadofthehousehol{d}_{ij}+\\ & {\beta }_4{ }_j^{D}Sisben{{11}_i}_j+{\beta }_5{ }_j^{D}Rura{l_i}_j+{\beta }_6{ }_j^{D}Exposuretim{e}_i{\hspace{0.17em}}_j+{{\beta }_i}_j,\mathrm{w}\mathrm{h}\mathrm{e}\mathrm{r}\mathrm{e}:\end{array}$

(3) ${\beta }_{kj}^D={\gamma }_{k0}^D+{\gamma }_{k1}^{D}Indexofviolenceandconflic{t}_j^D+u_{kj}$(3)

where i represents the individual, j the destination municipality, and k the individual variable. D indicates that in the context change model the IVC varies across the destination municipalities. In this case $Y_{ij}^D$ is one if female i who migrated to municipality j experienced motherhood in the destination and it is zero otherwise. $Y_{ij}^D$ is measured in the destination in March 2009. The context change models include the same individual variables as those considered for the models in origin plus the exposure time. The exposure time is the time in the municipality of destination measured in months. All of the individual variables are measured before migration except age, which is measured in March 2009. When each ${\beta }_{kj}^D$ is replaced in the equation of $Y_{ij}^D$, the final estimation is:

(4) $Y_{ij}^D={\gamma }_{00}^{\hspace{0.17em}}+{\gamma }_{01}IV{C}_j^D+\sum _{k=1}^K{\gamma }_{k0}{X}_{kij}+\sum _{k=1}^K{\gamma }_{k1}{X}_{kij}IV{C}_j^D+\sum _{k=1}^K{u}_{kj}{X}_{kij}+u_{0j}+{\epsilon }_{ij}$(4)

where $IV{C}_j^D$ represents the index in destination j. $X_{kij}$ is the independent variable k for individual i who migrated to municipality j. K is the total number of individual variables: six. $u_{0j}$ and $u_{kj}$ are the stochastic parts of the intercept and of the independent variable k.

3.3. Violence aftermath models

The purpose of the violence aftermath models is to evaluate the differences in the probability of childbearing between adolescents who have been exposed to violence but currently live in a peaceful context and those who have always lived in a peaceful context. Thus, the study population consists of migrant adolescents who come from municipalities with different levels of violence and migrate to a peaceful municipality. To guarantee homogeneous conditions in the destination, only migrants to Bogota are considered. This city is a peaceful municipality with a displacement rate of 0.015 per cent and an IVC equal to 0.22. Moreover, it is the most common destination for displaced and non-displaced people. To be included in the regression, each female adolescent must not have experienced motherhood before arriving in Bogota. Two levels are considered in this scenario: adolescent i and municipality of origin j. O in Equation (5)(5) $Y_{ij}^O={\gamma }_{00}^{\hspace{0.17em}}+{\gamma }_{01}IV{C}_j^O+\sum _{k=1}^K{\gamma }_{k0}{X}_{kij}+\sum _{k=1}^K{\gamma }_{k1}{X}_{kij}IV{C}_j^O+\sum _{k=1}^K{u}_{kj}{X}_{kij}+u_{0j}+{\epsilon }_{ij}$(5) indicates that in the violence aftermath model the IVC varies across the municipalities of origin. $Y_{ij}^o$ is one if female i who came from municipality j has borne a child in Bogota in March 2009 and it is zero otherwise. The individual variables are the same as those considered in the context change models. The exposure time corresponds to the time that the adolescent has lived in Bogota and is measured in months. Age corresponds to the adolescent’s age in March 2009. The rest of the individual independent variables are measured in the municipality of residence before migrating:

${\beta }_{kj}^O={\gamma }_{k0}^O+{\gamma }_{k1}^{O}Indexofviolenc{e}_j^O+u_{kj}$

and the multilevel model is:

$\begin{array}{rl}& Y_{ij}^O=log\left(\frac{{\pi }_{ij}}{1-{\pi }_{ij}}\right)\\ & ={\beta }_{0j}^O+{\beta }_{1j}^{O}ag{e}_{ij}+{\beta }_{2j}^{O}Bothparent{s}_{ij}\\ & +{\beta }_{3j}^{O}Educationoftheheadofthehousehol{d}_{ij}+{\beta }_{4j}^{O}Sisben{1}_{ij}\\ & +{\beta }_{5j}^{O}Rura{l}_{ij}+{\beta }_{6j}^{O}Exposuretim{e}_{ij}+{\epsilon }_{ij}\end{array}$

When each ${\beta }_{kj}^O$ is replaced in $Y_{ij}^O$, the final equation is:

(5) $Y_{ij}^O={\gamma }_{00}^{\hspace{0.17em}}+{\gamma }_{01}IV{C}_j^O+\sum _{k=1}^K{\gamma }_{k0}{X}_{kij}+\sum _{k=1}^K{\gamma }_{k1}{X}_{kij}IV{C}_j^O+\sum _{k=1}^K{u}_{kj}{X}_{kij}+u_{0j}+{\epsilon }_{ij}$(5)

where $IV{C}_j^O$ represents the index in origin j. $X_{kij}$ is independent variable k for individual i who came from municipality j. K is the total of individual variables: six. The stochastic parts of the intercept and of the independent variables are $u_{0j}$ and $u_{kj}$.

4. Data

Sisben is the System of Identification of Potential Beneficiaries of Social Programmes. The Sisben data do not contain a unique identification number for everybody because their objective is to list people, not to track them. The construction of the Sisben panel dataset 2006–2009 and its advantages over other datasets are discussed in Supplementary Materials 3. The Sisben data includes a wellbeing index. This index orders the population according to their living conditions, and higher scores imply better conditions. In the second version of this index, used from 2002 to 2010, the population is divided into six levels according to their scores. People classified in levels one and two, who live in extremely poor and poor conditions, are the potential beneficiaries of social programmes (Flórez, Sánchez, Espinosa, & Angulo, 2008). We focus on migrants who belong to social programmes, classified in Sisben levels one and two, because they must provide an update for Sisben every time they migrate.

5. Descriptive statistics

Table A5 in Supplementary Materials 4 presents the descriptive statistics of the independent variables. This table shows that the most vulnerable sample of migrants is the one of the context change models, 84 and 48 per cent of the adolescents in the 12–20 age-group are extremely poor and live without both parents before migration, respectively. In contrast, these percentages are 59 and 27 per cent for those who migrate to Bogota, who are the less vulnerable. Regarding head’s education, on average migrants to Bogota are the most educated (8.6 years), while those who do not migrate are the less educated (4.01 years). Figure A2 in Supplementary Materials 4 shows the probability of childbearing versus age. The graphs of the sample for the models in origin show that the higher adolescents’ vulnerability, the bigger the probability of childbearing. Regarding the IVC, the graphs for the models in origin and the context change models illustrate that adolescents with IVC > 0.5 exhibit higher probability than adolescents with IVC ≤ 0.5. The same occurs for the adolescents younger than 18 in the violence aftermath models.

Table 2 shows for both samples and the three models the distributions of the dependent variables for two IVC levels: values lower than 0.5 and values higher than or equal to 0.5. This table reports that in general the percentage of female adolescents who have assumed childbearing is higher among those who have lived in municipalities with IVC higher than or equal to 0.5 than among those who have lived in municipalities with IVC lower than 0.5. The difference between the two percentages can reach 3.5 percentage points.

Table 2. Distributions of the dependent variables by IVC level

	Adolescents aged 12–20				Adolescents aged 12–24
	IVC<0.5	IVC≥0.5	Diff.		IVC<0.5	IVC≥0.5	Diff.
Models in origin	0.082	0.097	0.015	*	0.177	0.202	0.025	*
			(0.0006)				(0.0007)
Context change models	0.258	0.292	0.034	*	0.361	0.398	0.0367	**
			(0.019)				(0.017)
Violence aftermath models	0.199	0.233	0.034		0.328	0.325	−0.002
			(0.025)				(0.023)

Source: Sisben data 2006–2009. Notes: Standard errors in parenthesis. Since the population for the models in origin is larger than 1 million, the coefficients are statistically significant (*) when the t values are greater than or equal to the Leamer–Schwarz critical value, $\sqrt{\mathrm{l}\mathrm{n}\left(N\right)}$. For the context change and violence aftermath models: *** significant at 99 per cent, ** significant at 95 per cent, and * significant at 90 per cent.

6. Sample selection, endogeneity and limitations

All the models considered adolescents whose places of residence, origin, and/or destination are municipalities with fewer than 100,000 inhabitants. The only exception is the destination of the females included in the violence aftermath models, which is Bogota. Considering migrants from municipalities with a population of fewer than 100,000 inhabitants allows us to assume that people from one of these municipalities have been exposed to similar levels of violence. This consideration drops 56 municipalities out of the 1102 in Colombia. This suggests that only big municipalities are discarded: municipalities where the main reasons for migrating are not related to armed conflict.

To take into account the degree of selectivity of migrants, we control by migrants’ characteristics, as Adsera and Ferrer (2014) suggest. These characteristics are named in Section 3 and include: age, heads’ education, level of poverty, municipality of origin or residence, and so forth. The Sisben survey does not allow us to control by people’s ambitions and motivations; however, it is likely that these features are related to the control variables considered.

Concerning the sample bias of the Sisben data, the baseline for the Sisben survey was established like a census in the socio-economic strata one, two, and three in all of Colombia. Therefore, the baseline of the Sisben data includes all the poor population in Colombia. Thus, for the first observation, before migration, the sample is not biased in the context change and violence aftermath models. However, the sample of migrants might be biased since only those who are interested in being beneficiaries of a social programme update their information after migration. Nonetheless, around 87 per cent of the people in the Sisben data must update their information after migration since they are beneficiaries of the Subsidised Health Regime (Living Conditions Survey LCS-2008). In addition, at least eight institutions and 22 welfare programmes use Sisben data either as the main condition for giving benefits or as a prioritisation criterion. Consequently, the sample selection bias is minimised by the incentive of receiving government aid.

For the violence aftermath models, the sample selection bias of migrants who do not update their information after migration must be very low, since the final destination is Bogota, which is the capital city and has no armed conflict. For the context change models, the probability of not updating the information might increase with the levels of violence in the destination. However, it is not possible to identify those who do not update their information. In this case, the effect of a change in context may be underestimated, because the beneficiaries who do not update their information after they have settled in a municipality might be the most vulnerable ones. Nevertheless, since the study population consists of those who stay for one year or more in a destination, it is very likely that they have decided to register in Sisben database to become beneficiaries of the Subsidised Health Regime.

Regarding endogeneity, the dependent variables of the contextual change and violence aftermath models are measured in 2009, after migration, and the independent ones, including the IVC, before migration; therefore, there is no double causality. Moreover, the probability of childbearing by adolescents does not explain any of the independent variables, so the models in origin do not involve endogeneity problems by simultaneity. However, the level of education of the head of the household is included in the control variables, and it might determine the poverty level. Nevertheless, in rural areas, where armed conflict mainly takes place, poverty is not highly associated with education. According to the LCS 2008, in rural areas 99 and 95 per cent of the heads of household in the poor (Sisben one and two) and the non-poor population (Sisben three and four), respectively, have the maximum basic schooling. Although the relationship between both variables may be a limitation to find a causal relationship, excluding any of them may cause endogeneity problems as a consequence of omitted variables, since both determine the probability of childbearing among adolescents (Flórez et al., 2004; National Research Council and Institute of Medicine, 2005).

7. Results

This section introduces and interprets the results of the three types of models: models in origin, context change models, and violence aftermath models. Only the interactions that are statistically significant at 90 per cent or above are considered in the final models. We centre the continuous independent variables at the first level considering the great mean. Thereby, the intercept is the value that corresponds to an adolescent of average age in 2009 from an urban area who was or is poor (depending on the model), who lived or lives with one or no parents, and whose head of household has or had an average educational level.

7.1. Models in origin

Table 3 reports the models in origin for the two age groups. The upper panel of this table shows the coefficients and odds ratios of the independent variables and the lower panel the variances of the coefficients across municipalities. The upper panel shows that on average the probability of childbearing increases with the adolescent’s age and poverty level and decreases with the level of urbanicity. Living with both parents instead of one or no parents also reduces this probability. All the previous results are expected. The only coefficient that is unexpected is that of the education of the head of the household. According to the estimates, the more educated s/he is, the higher the probability of teenage motherhood. This result might be explained by the low levels of education in small municipalities. For instance, for the 12–20 age group, only 20.64 and 1.16 per cent of the heads of household complete primary and secondary education, respectively.. Indeed, the average number of years of education of the heads of household is around four for both age groups. Therefore, it is possible that the importance of heads’ education for the reduction of the probability of childbearing is not evident among the residents of small municipalities.

Table 3. Models in origin – 2009

	Women aged 12–20			Women aged 12–24
	Coeff.		Odds ratio	Coeff.		Odds ratio
Age in 2009	0.552	*	1.737	0.398	*	1.489
	(0.003)			(0.001)
Living with both parents	−1.275	*	0.280	−1.482	*	0.227
	(0.023)			(0.022)
Education of the head	0.021	*	1.021	0.046	*	1.047
	(0.002)			(0.002)
Sisben 1	0.621	*	1.861	0.657	*	1.928
	(0.014)			(0.012)
From rural areas	0.124	*	1.132	0.251	*	1.285
	(0.020)			(0.017)
IVC	0.121		1.128	0.171		1.186
	(0.088)			(0.085)
Rural*IVC	0.253	*		0.185	*
	(0.054)			(0.048)
Intercept	−3.338	*		−2.330	*
	(0.031)			(0.029)
Var (intercept)	0.280	*		0.288	*
	(0.017)			(0.017)
Var (age)	0.003	*		0.001	*
	(0.000)			(0.000)
Var (both parents)	0.426	*		0.418	*
	(0.026)			(0.023)
Var (head’s education)	0.002	*		0.002	*
	(0.000)			(0.000)
Var (Sisben 1)	0.054	*		0.073	*
	(0.007)			(0.006)
Var (rural)	0.062	*		0.067	*
	(0.007)			(0.005)
Number of zeros	1,231,531			1,519,503
Number of ones	101,677			294,083
Leamer–Schwarz critical value	3.755			3.796

Source: Sisben data 2006–2009. Notes: The reference category for rural is urban; for living with both parents it is living with one or no parents; and for level 1 of Sisben it is level 2. Standard errors in parenthesis. The coefficients are statistically significant (*) when the t values are greater than or equal to the Leamer–Schwarz critical value, $\sqrt{\mathrm{l}\mathrm{n}\left(N\right)}$.

Table 3 shows that in peaceful circumstances the odds for adolescents aged 12–20 increases by 73.7 per cent for a one-year increase in age; this increase is 48.9 per cent when females aged 21–24 are considered. The odds of childbearing for adolescents who live with both parents are about one-quarter of the odds for those who live with one or no parents, whereas the odds for extremely poor females are almost twice (1.86 and 1.93) the odds for poor females. The odds for rural adolescents are 1.13 and 1.28 the odds for urban adolescents for the younger group and  for the group that includes females aged 21–24, respectively. In short, in peaceful situations the most vulnerable group is extremely poor rural adolescents who live with one or no parents.

Table 3 also shows that contextual violence regulates the effect of the degree of urbanisation. Specifically, living in violent contexts increases the likelihood of becoming a mother for rural females. For rural females aged 12–20, the odds of childbearing are 1.13 times larger than the odds for urban females in peaceful contexts, but in warlike circumstances the odds ratio between rural and urban females increases to 1.46 (exp (0.124 + 0.253)). The same occurs when 21–24-year-old females are considered in the sample. The coefficient of rural is given by the equation 0.251 + 0.185 IVC. This suggests that the current exposure to violence increases the likelihood of childbearing among rural females particularly. The reason behind this result is that Colombian armed conflict occurs mainly in rural areas, so the rural population is the most affected by it.

The lower panel of Table 3 shows that the intercept and the coefficients for all the independent variables of the first level vary across municipalities for both age groups. This means that there are municipalities with low intercepts, where the probability of childbearing among adolescents is lower than that in municipalities with high intercepts. Similarly, the coefficients of the independent variables are not the same for all the municipalities. For instance, the differences in the coefficient for Sisben one indicate that the relationship between adolescents’ poverty level and their predicted probability of childbearing is not the same for all the municipalities. In other words, some municipalities have a high value for the coefficient of Sisben one; in these municipalities, the differences between Sisben one or extremely poor adolescents and Sisben two or poor adolescents is relatively large, while, in municipalities with low values for this coefficient, the effect of the poverty level on the probability of childbearing is small. In short, the values of the coefficients depend on the municipality. This is exemplified in Table 4.

Table 4. Predicted probabilities for an adolescent aged 12–24 when the intercept and the coefficient of ‘from rural areas’ vary based on Table 3

Context:	1 As Table 3	2 Intercept+1 std dev.	3 Coeff. rural+1 std dev.	4 Intercept+1 std dev. and coeff. rural+1 std dev.
Intercept (IVC = 0)	0.089	0.143	0.089	0.143
Probabilities with the following changes:
Rural areas (IVC = 0)	0.111	0.176	0.140	0.217
IVC = 1	0.104	0.165	0.104	0.165
Rural areas and IVC = 1	0.152	0.234	0.188	0.284

Note: The intercept is the value for an adolescent of average age in 2009, from an urban area, who is poor, who lives with one or no parents, whose head of household has an average educational level and who resides in a peaceful area or IVC = 0.

Table 4 shows the predicted probabilities for the age group 12–24 when the intercept and coefficient of from rural areas change. In context 1 the intercept and the coefficients are those reported in the upper part of Table 3. For context 2 the intercept increases by one standard deviation, while for context 3 the coefficient of from rural areas increases by one standard deviation. The standard deviations are calculated based on the variances reported in the lower panel of Table 3. In context 4 both the intercept and the coefficient of from rural areas are those reported in Table 3 plus one standard deviation. Table 4 shows that in peaceful circumstances, IVC = 0, a one-standard-deviation increase in the intercept increases the predicted probability from 0.089 to 0.143, while a one-standard-deviation increase in the coefficient of from rural areas increases this probability from 0.111 to 0.140. This table also shows that the difference in the predicted probabilities between two adolescents who reside in a violent rural area, IVC = 1, but one in context 1 and the other in context 4, is 0.132 (this is 0.284–0.152). In brief, the change in the probability of childbearing induced by a change in any independent variable depends on the municipality of residence.

7.2. Context change models

The context change models are displayed in Table 5. The coefficients of the IVC shown in Table 5 report the effect of violence in the destination, because in the context change models all of the migrants come from violent areas and move to municipalities with different levels of violence. In this case no interaction is statistically significant, suggesting that the level of violence in the destination does not moderate the effect of any independent variable at the first level.

Table 5. Context change models

	Women aged 12–20			Women aged 12–24
	Coeff.		Odds ratio	Coeff.		Odds ratio
Age in 2009	1.035	***	2.815	0.609	***	1.839
	(0.088)			(0.040)
Living with both parents	−0.047		0.954	−0.011		0.989
	(0.202)			(0.148)
Education of the head	−0.035	**	0.966	−0.011		0.989
	(0.018)			(0.011)
Sisben 1	0.612	***	1.844	0.468	***	1.596
	(0.219)			(0.154)
From rural areas	0.056		1.057	0.263	*	1.301
	(0.236)			(0.152)
Exposure time	−0.072	***	0.930	−0.026	**	0.974
	(0.017)			(0.012)
IVC	0.977	**	2.656	0.907	***	2.476
	(0.454)			(0.328)
Intercept	−3.427	***		−1.912	***
	(0.354)			(0.202)
Var (intercept)	1.469	**		0.789	***
	(0.603)			(0.291)
Var (age)	0.304	***		0.078	***
	(0.101)			(0.024)
Var (both parents)	2.667	***		2.451	***
	(0.972)			(0.743)
Var (head’s education)	0.031	***		0.016	***
	(0.012)			(0.005)
Var (Sisben 1)	0.729			1.125	***
	(0.582)			(0.394)
Var (rural)	5.634	***		2.732	***
	(1.655)			(0.700)
Var (exposure time)	0.028	***		0.021	***
	(0.010)			(0.005)
Number of zeros	1964			2393
Number of ones	729			1431

Source: Sisben data 2006–2009. Notes: The reference category for rural is urban; for living with both parents it is living with one or no parents; and for level 1 of Sisben it is level 2. Standard error in parenthesis. *** Significant at 99 per cent, ** significant at 95 per cent, and * significant at 90 per cent.

Table 5 shows that the effects of the individual variables have the same sign as those obtained in the models in origin, with the exception of the coefficients of the head’s education. For the context change models, this coefficient is negative, suggesting that the more educated the head of the household, the lower the probability of teenage childbearing. In fact, the odds for adolescents aged 12–20 decreases by 3.4 per cent for a one-year increase in the education of the head of the household. In this case 21.38 per cent of the heads of household have completed secondary education. This figure is 20 percentage points higher than that found for the sample used in the models in origin (1.16 per cent). In fact, the average years of education of the heads of household for both age groups is seven for the context change models, which is three years higher than the average found for the samples utilised in the models in origin. This suggests that migrants on average are more educated than non-migrants. Thus, it is likely that the behaviour of migrants is more affected by their education than that of non-migrants.

Another difference with the models in origin is that the coefficients of living with both parents in Table 5 are not statistically different from zero for either of the two groups. Traditionally, orphanhood, poor parents’ supervision and dysfunctional families have been identified as causes of teenage parenthood, because adolescents intentionally become pregnant to form a family (Flórez et al., 2004; Kidman & Anglewicz, 2004). It is possible that the result of the context change models is explained by its study population: migrants who come from violent municipalities. It is likely that being a migrant and coming from a violent context make an adolescent vulnerable regardless of the type of household she belongs to. Similar results are found by Palermo and Peterman (2009). They study the relationship between orphanhood status and teen pregnancy in 10 sub-Saharan African countries, finding little association between both variables.

Regarding the other variables, a one-year increase in age increases the odds of childbearing by 182 and 84 per cent for the age groups 12–20 and 12–24, respectively. The odds of childbearing for extremely poor females are 84.4 and 59.6 per cent higher than the odds for poor females for adolescents aged 12–20 and 12–24, respectively. Rural migrants are more vulnerable than urban migrants; their odds for the younger and the older group, in that order, are 5.7 and 30.1 per cent higher. A one-month increase in the exposure time in the destination decreases the odds of childbearing between 2.6 and 7 per cent. This may indicate that the effect of migration, which might increase the likelihood of childbearing among adolescents (Baumer & South, 2001; Erulkar & Ferede, 2009; Stack, 1994), decreases over time.

Concerning contextual violence, the estimates for both groups of adolescents show that adolescents who migrate to violent areas have a higher probability of childbearing than those who migrate to peaceful areas. This suggests that a decrease in the levels of violence improves the behaviour of the teenagers who have been exposed to violence. The odds of childbearing for an adolescent aged 12–20 who moves to a violent municipality are 2.66 times larger than the odds for an adolescent who moves to a peaceful context. This value is 2.48 when adolescents aged 21–24 are included in the sample.

As in the models in origin, the lower panel of Table 5 shows that the variances of the intercept and of the coefficients of the independent variables are all statistically significant. This suggests that their values depend on the destination municipality. To exemplify how different the probability of childbearing can be according to the destination, Table 6 shows the predicted probability for two contexts: for context 1 the intercept and the coefficients are those reported in the upper part of Table 5, and for context 2 the intercept is one standard deviation higher than the intercept of context 1. Table 6 shows that, for the 12–20 age group, the probability of childbearing in context 2 is about three times the probability in context 1: 9.84/3.15 comparing peaceful destinations, IVC = 0, and 22.48/7.94 for violent destinations, IVC = 1. For the 12–24 age group, these ratios are around double: 26.44/12.88 and 47.07/26.80 comparing peaceful and violent destinations, respectively.

Table 6. Predicted probabilities for an adolescent when the intercept and the coefficient of ‘from rural areas’ vary based on Table 5

	1		2
Destination:	As Table 5		Intercept+1 std dev.
Age group	12–20	12–24	12–20	12–24
Intercept (IVC = 0)	3.15	12.88	9.84	26.44
Probabilities with the following changes:
IVC = 1	7.94	26.80	22.48	47.08

Note: The intercept corresponds to an adolescent of average age in 2009, from an urban area, who was poor, who lived with one or no parents, whose head of household had an average educational level and who migrated to a peaceful area or IVC = 0.

7.3. Violence aftermath models

The violence aftermath models are reported in Table 7. As in the other two types of models, the variances of the intercept and of the coefficients are statistically different from zero for both age groups, which means that they vary across municipalities. In Table 7 IVC is the index of violence and conflict in the origin, because in the violence aftermath models the migrants come from municipalities with different levels of violence and all move to Bogota. According to this table, one interaction with the IVC is statistically significant at the 90 per cent level and above for each age group: for 12-to-20-year-old adolescents Sisben one and for 12-to-24-year-old adolescents coming from rural areas.

Table 7. Violence aftermath models

	Women aged 12–20			Women aged 12–24
	Coeff.		Odds ratio	Coeff.		Odds ratio
Age in 2009	1.418	***	4.127	0.587	***	1.798
	(0.189)			(0.041)
Living with both parents	−1.574	***	0.207	−0.125		0.883
	(0.557)			(0.130)
Education of the head	−0.002		0.998	0.013		1.013
	(0.046)			(0.012)
Sisben 1	1.628	*	5.092	0.351	**	1.420
	(0.840)			(0.146)
From rural areas	−1.773	**	0.170	−0.401	*	0.669
	(0.684)			(0.224)
Exposure time	−0.500	***	0.606	−0.166	***	0.847
	(0.082)			(0.018)
IVC	2.823		16.829	−0.519		0.595
	(1.861)			(0.453)
Sisben 1*IVC	−4.790	**
	(2.403)
Rural*IVC				1.142	*
				(0.671)
Intercept	−8.552	***		−1.660	***
	(1.187)			(0.189)
Var (intercept)	16.263	***		0.473	*
	(6.095)			(0.266)
Var (age)	1.818	**		0.159	***
	(0.796)			(0.039)
Var (both parents)	18.056	**		0.728	**
	(7.946)			(0.311)
Var (head’s education)	0.399	***		0.029	***
	(0.155)			(0.006)
Var (Sisben 1)	32.736	***		2.490	***
	(12.541)			(0.611)
Var (rural)	30.990	**		2.028	***
	(12.159)			(0.598)
Var (exposure time)	1.216	***		0.040	***
	(0.426)			(0.009)
Number of zeros	2000			2683
Number of ones	513			1305

Source: Sisben data 2006–2009. Notes: The reference category for rural is urban; for living with both parents it is living with one or no parents; and for level 1 of Sisben it is level 2. Standard errors in parenthesis. *** Significant at 99 per cent, ** significant at 95 per cent, and * significant at 90 per cent.

Table 7 shows that a one-year increase in age increases the odds of childbearing by 312.7 per cent for the younger group and 79.8 per cent when 12–24-year-old adolescents are included. Living with both parents instead of one or no parents decreases the odds by 79.3 per cent for the former group. The coefficient of the education of the head of the household is not statistically significant in both models. A one-month increase in the exposure time in the destination decreases the odds of childbearing by 40 per cent for the younger group and by 15 per cent for the older. It is likely that this result is related to the assimilation process in Bogota; the longer the exposure time, the more adapted and less vulnerable the adolescents are. Contrary to the results of the two previous models, the coefficient of rural is negative; this means that the likelihood of childbearing is higher among urban migrants than among rural migrants. According to the estimates, the odds for rural adolescents are 17 per cent of the odds for urban adolescents considering only young adolescents and 67 including adolescents aged 21–24.

Table 7 shows that the level of violence in the origin moderates the relationship between the likelihood of childbearing and the level of poverty following the equation: log of the odds equals 1.628–4.790 IVC. Thereby, when adolescents come from peaceful municipalities, the odds of childbearing for extremely poor adolescents are five times (exp(1.628)) the odds for poor ones. However, when they come from municipalities where the IVC equals one, the odds ratio between extremely poor and poor adolescents is 0.042 (exp(−3.16)), which is statistically significant at 90 per cent. This suggests that a poor adolescent has a higher probability of childbearing than an extremely poor one when both come from a very violent area. Hence, an increase in the level of violence in the origin decreases the differences in behaviour between poor and extremely poor adolescents.

The estimate that includes females aged 21–24 reports that the odds ratio between a rural female and an urban female is 0.67 when both come from a municipality with IVC = 0. If both migrate from municipalities with a higher IVC value, the coefficient of coming from rural areas is given by the equation −0.401 + 1.142 IVC. This means that if IVC = 1, the odds of childbearing for rural adolescents are twice the odds for urban adolescents.

Using the results of Table 7, the upper part of Table 8 shows four contexts for adolescents aged 12–20: context 1 uses the reported coefficients in Table 7, contexts 2 and 3 increase the intercept and the coefficient of Sisben one by one standard deviation, respectively, and context 4 increases both the intercept and the coefficient of Sisben one by one standard deviation. The predicted probability of childbearing in Bogota for an extremely poor adolescent from a violent municipality with context 1 is 0.01 per cent; in contrast, if the adolescent comes from a violent municipality with context 4, the probability is 70.33 per cent. The lower part of Table 8 shows the same exercise for the age group 12–24. In this case the coefficient of ‘from rural areas’ varies, instead of that of Sisben one. The difference in the predicted probability between a rural adolescent from a violent municipality with context 8, when both the intercept and the coefficient increase, and context 5, which uses the coefficient reported in Table 7, is 50 percentage points (69.59–19.18).

Table 8. Predicted probabilities for an adolescent when the intercept and the coefficient of Sisben and ‘from rural areas’ vary based on Table 7

Context	1 As Table 7	2 Intercept+1 std dev.	3 Coeff. Sisben l + 1 std dev.	4 Intercept+1 std dev. and coeff. Sisben 1 + 1 std dev.
Age range	12–20	12–20	12–20	12–20
Intercept	0.02%	1.08%	0.02%	1.08%
Probabilities with the following changes:
Sisben 1	0.10%	5.26%	23.10%	94.43%
IVC = 1	0.32%	15.50%	0.32%	15.50%
Sisben 1 and IVC = 1	0.01%	0.77%	4.03%	70.33%
Context	5 As Table 7	6 Intercept+1 std dev.	7 Coeff. rural l + 1 std dev.	8 Intercept+1 std dev. and coeff. rural+1 std dev.
Age range	12–24	12–24	12–24	12–24
Intercept	15.98%	27.45%	15.98%	27.45%
Probabilities with the following changes:
Rural areas	11.29%	20.21%	38.14%	55.10%
IVC = 1	10.16%	18.37%	10.16%	18.37%
Rural areas and IVC = 1	19.18%	32.08%	53.49%	69.59%

Notes: The intercept corresponds to an adolescent of average age in 2009, from an urban area, who was poor, who lived with one or no parents, whose head of household had an average educational level and who migrated to a peaceful area or IVC = 0.

8. Conclusions

Some studies (such as Rutayisire et al., 2014; Verwimp & Bavel, 2005) find that exposure to armed conflict violence modifies women’s fertility behaviour. This article contributes to the literature by confirming that female adolescents’ fertility behaviour is also modified by their exposure to armed conflict. In addition, it not only studies current exposure to violence but is the first to analyse adolescents’ behaviour after they have migrated to peaceful municipalities. Through the use of migrants, this study determines how the change in context from violent to peaceful and the pre-exposure to armed conflict violence affect the probabilities of commencing childbearing. The study also identifies the characteristics of the adolescents who continue to behave as if they were still in a violent municipality despite moving to a peaceful destination.

The models in origin, which study the current exposure to armed conflict, indicate that violence generally accelerates the transition to childbearing; however, its effects differ according to the degree of urbanisation of the area of residence. The influence of living in rural areas on the likelihood of childbearing increases with the level of violence.

The context change models determine the impact of a reduction in the levels of violence on the probabilities of childbearing. The study population consists of adolescents who were living in areas with high levels of violence and migrated to areas with different levels. These models find that a decrease in the exposure to violence reduces the likelihood of childbearing. The odds of females who migrate to violent areas are more than twice the odds of those who migrate to peaceful areas.

The violence aftermath models compare migrant adolescents who have been exposed to violence with those who have never been exposed, identifying the features of the adolescents who, after migrating from violent to peaceful destinations, do not completely adapt to the new context. According to the results, rural adolescents and those who were poor (Sisben two) before migration are those who, after residing in a peaceful destination, still experience the aftermath of their previous exposure to violence. In fact, the odds of poor adolescents might be higher than those of extremely poor adolescents, who exhibit a higher probability of childbearing in peaceful situations.

For the three types of models, we find that the intercept and the coefficients of most individual independent variables vary across municipalities. As a result, their effects and the effect of contextual violence might change considerably between the municipalities of residence, destination and origin.

Acknowledgments

Author offers her sincerest gratitude to her PhD supervisor, Birgitta Rabe; She also thanks Patricia Justino, Reene Luthra and two unknown referees for their comments. She thanks Roberto Angulo, who was the coordinator of the quality-of-life division of the National Planning Department in 2009, for allowing him to use the Sisben database. She also thanks the engineers of this division for helping him with the construction of the Sisben panel database.

Disclosure statement

No potential conflict of interest was reported by the author.

Supplemental Material

Supplemental data for this article can be accessed on the http://dx.doi.org/10.1080/00220388.2017.1342816

Notes

1. NCCEV (2003) defines community violence as exposure to acts of interpersonal violence committed by individuals who are not intimately related to the victim.