Household Composition and Civil War: Evidence from Burundi

Transcription

1 Household Composition and Civil War: Evidence from Burundi Olivia D Aoust October 31, 2013 Abstract In this paper, we propose to study household composition patterns following household s exposure to the civil conflict in Burundi. Our analysis relies on a panel dataset collected in rural Burundi in 2005 and Through cross-section and panel analysis, we show that households living in more violent-prone areas were smaller in size in 2005, and that there size has even decreased between 2005 and Looking at different age-sex groups within households, we find that civil conflict has had a negative impact on children and aged men headcounts. Finally, restricting our sample to the most affected villages, we find evidence of non-linear effects of violence, suggesting that civil conflict may have a positive impact on household size, in particular through an increase in young children members. Our results further highlight the complexities of demographic processes documented in the literature. Keywords: Civil war, Burundi, Household composition, Household size, Panel 1. Introduction In the last decade, civil conflicts have sparked off an increasing interest from social scientists. In particular, in 2003, a working group of the International Union for the Scientific Study of Population (IUSSP) was created to study the demography of civil conflict (Brunborg and Tabeau, 2005). Since then, the literature has particularly grown, and comprises studies on both the demographic causes and consequences of civil war. In this paper, we propose to study household composition patterns following household s exposure to the civil conflict in Burundi. Particularly, we have decomposed the sample in different age and sex group in order to look at the impact of the occurrence of violence since Université libre de Bruxelles (SBS-EM, ECARES), FNRS. odaoust@ulb.ac.be. I acknowledge financial support from the Fonds National de la Recherche Scientifique (FNRS). Data collection was financed by the MICROCON project (EU 6th Framework). I am grateful to An Ansoms, Philippe Bocquier, Bram De Rock, Olivier Sterck, Dominique Tabutin and Philip Verwimp for useful discussions and comments. All remaining errors are mine. Draft - Please do not distribute

2 1993 on particular groups, which should have been differently affected by the conflict. The analysis relies on a panel dataset collected in 2005 and 2010 in three provinces of rural Burundi: Bubanza, Bujumbura Rural and Cibitoke (Figure 1). These have been particularly affected by the conflict since its onset. Household composition is affected by well-identifiable factors : mortality, fertility, migration and marriage. While these have been studied in the context of civil war, household composition as such has not. This indicator however results from these phenomena and is the ultimate household demographic characteristic. To the best of our knowledge, only Winters et al. (2009) studyhouseholdcomposition,inthecontextofeconomicshocksin Nicaragua. In this study, we will analyze these dynamics in the context of civil war. In the absence of demographic measures on mortality, migration, fertility and marriage, we will derive different hypothesis explaining changes in household composition given the age and sex categories affected. Civil conflict has been shown to affect patterns in mortality, fertility, migration and the marriage market. Such analyses have long been cross-country studies - mostly studying the causes of civil war 1 -butmorerecentlyturnedtobecomemicro-level,orientedtowardsits consequences on households 2. In a study on the challenges faced by war-windows after the Rwandan genocide, Bruck et al. (2009) goesevenfurther,pointingoutanexistinggapinthe literature what concerned intra-household dynamics. This research therefore contributes at filling this gap, by proposing to study intra-household dynamics in terms of composition. It is also innovative in his research topic, which has rarely been studied, and never in the context of civil conflict, which has been shown to considerably affect household behavior and decisions (see Justino (2008) for a good review). The paper is structured as follows. The next section sets the stage, reviewing the major events that shaped Burundese history. The third section reviews the literature on the determinants of household composition in times of conflict. Section four consist of en empirical analysis of the impact of violence on household composition. There, we present the data, the empirical strategy and the results of the assessment. Section five concludes, and proposes some policy recommendations. 2. Historical perspective The Great Lake region, and more largely post-colonial Africa, are known for the succession of violence outbreaks that followed independence. These can be mostly associated to internal political stakes (Chrétien, 1981; Collier et al., 2003). Between 1960 and 2009, 1 As in Collier and Hoeffler (2004), Fearon and Laitin (2003) orurdal (2005) amongothers 2 See for example Verwimp and Van Bavel (2005) orde Walque and Verwimp (2010) 2

3 Gleditsch et al. (2002) haverecorded73armedconflicts 3.Amongthese,75%areconsidered as civil wars, which occurred between the government of a state and one or more internal opposition group(s) without intervention from other states (Gleditsch et al., 2002). The State is therefore a main actor in the conflict. As it is a place of enrichment and power, it became the source of disputes between the elites, who fought to preserve their past glories, at any cost. Ethnic, religious and political propaganda kept being launched by the potential leaders looking for population support, in the mean time cultivating fear and hatred. Power therefore passed from hands to hands through violence, which was rapidly continuously triggered in a vicious cycle, in which each group wanted to pull the rug out from the other before he does (Chrétien, 2000). Ethnic discrimination started during the colonial era, during which both German and Belgian colonizers used ethnicity in their system of hierarchy. As they decentralized power, they offered the Tutsi the ruling positions and imposed a social hierarchy based on ethnic criteria in all spheres of the society. Such categorization led to a continuous opposition between Tutsi and Hutu in Burundi and Rwanda, and ultimately fueled the civil wars that devastated both countries during the 1990s. In Rwanda, ethnic massacres became virulent after independence. When the Hutu took the power, thousands of Tutsi fled to neighboring Uganda, DR Congo and Tanzania. Young exiled men later formed the Rwandese Patriotic Front (RPF) and participated actively to the conflicts in the Great Lakes (Lemarchand, 1997). The Rwandese syndrome rapidly struck neighboring Burundi, where authorities started attributing ethnic motives to all political tensions. Tutsi started to be persecuted, but severely responded to finally establish a military dictatorship, under the authority of general Micombero. The following years, Burundi was the scene of large scale massacres, particularly in 1972, after which all Hutu elite had either died or fled (Lemarchand, 2002). The first attempt to escape this tit-for-tat trap was initiated in Buyoya, who took power in 1987 through a coup d Etat, launched a process of political liberalization following pressures from the international community (Chrétien, 2000). This process led to elections, held in 1993, and won by a hutu political party. This was a turning point in Burundi as the power passed into the hands of Hutu, and not only escaped from the Tutsi elite, but also from a political and regional group of the southern province Bururi, very influential until then. The elections were perceived as unacceptable by most Tutsi, who felt threatened. When ethnic extremism spread in Rwanda at that time, it rapidly crossed the border and propaganda were largely diffused in Burundi. Ndadaye was assassinated in coup attempt in 1993, opening the way to more massacres on both sides. He was replaced by Ntaryamira, 3 defined as contested incompatibility that concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battlerelated deaths (Gleditsch et al., 2002). 3

4 who died five months later in a plane crash with his Rwandese counterpart, Habyarimana. This incident precipitated both countries into the turmoil of civil war, and constituted the starting point of the genocide in Rwanda, which killed around 800,000 people. Burundi was left without political foundations, and became the scene of ethnic killings, which killed an estimated 300,000 civilians and forced another million to cross the borders to seek refuge in Tanzania, DRC and Rwanda (Chrétien, 1997; UNHCR, 2012). In North Kivu, refugee camps became the basement of exiled Burundese hutu and Rwandese perpetrators. Figure 1 presents the major events that happened from independence until the genocides. Figure 1: Timeline : Independance 1962 First Republic 1966 Tutsi Massacres & Hutu repression 1972 Coup d État 1976 Coup d État 1987 New Constitution & Power-sharing 1992 Elections Plane crash June Oct. Coup d État 1996 Government by Tutsi from Bururi Lobby Hutu government Micombero Bagaza Buyoya Ndadaye Ntarymira Buyoya Ntibantunganya As violence increased in Burundi, the first provinces to be affected were Cibitoke, Bubanza, Bujumbura Rural and Ngozi, located close to the border with DRC, where hutu rebel groups had established. Our provinces of interest have therefore been exposed to the war since its onset (Chrétien and Mukuri, 2000; Bundervoet et al., 2009). In July 1996, the Tutsi, led by Buyoya, overthrow the government through a coup d Etat. Neighboring countries imposed an embargo to Burundi to condemn it. Three years later, the embargo was lifted but the country was at his lowest point (Chrétien, 2000; Prunier, 2009). Buyoya started the peace negotiations that led to the first peace accord, signed in Arusha in Unfortunately, the treaty was not signed by two main rebel groups, the CNDD-FDD and the FNL. As both parties were very active on the ground, the agreement was only superficial. The country therefore continued to be the scene of attacks until In the mean-time, the CNDD-FDD signed the peace agreement in 2004, and his leader, Pierre Nkurunziza was elected president in In 2009, a cease-fire has been signed with the FNL (Vandeginste, 2009). Combatants from both groups were demobilized or redirected to the government forces during the period Since then, the country is still affected by episodes of violence but events are mostly local, and politically driven. Civilians 4

5 are therefore less targeted. The tensions have evolved from Hutu vs Tutsi confrontations to Hutu against Hutu ones, their difference now lying in their political beliefs. The last winner of this opposition between the FNL and the CNDD-FDD was Pierre Nkurunziza, who was reelected in Figure 2 shows the key milestones of the peace process. Figure 2: Timeline : peace process Conflict Onset Coup d État Arusha Peace agreement Demobilization CNDD Demobilization FNL Ethnic motives Government (Tutsi) CNDD (Hutu) FNL (Hutu) Other groups Government (Transition) CNDD (Hutu) FNL (Hutu) Other groups Political motives CNDD Government (Hutu) FNL (Hutu) This sequence of events and the evolution of conflict stakes will be used to support the identification strategy of the empirical analysis. Before going to the core of the analysis of section 5, let us review the main channels through which household composition is affected, and how these evolved in times of conflict. 3. Channels Household composition evolved when its members die or migrate, or when new persons join, from extended family or foster children to newborns. In addition to mortality and migration, composition is therefore affected by fertility and marriage decisions. The literature has shown that civil war has an impact on the determinants of these demographic phenomena, such as education, health or income. We will therefore review how mortality, migration, fertility and marriage are affected by civil war, what are the implication of such changes on composition, and under which assumptions Mortality Everyone dies with certainty. Mortality has therefore been studied from the angle of its distribution, and given causes of death. In conflict settings, we expect mortality to increase. Both demographers and epidemiologists have developed tools to estimate excess mortality in such context, this is, mortality that would not have happen if there had not been a war. This counterfactual scenario has been subject to debate, as its estimation is a tedious exercise but is determinant for calculating excess deaths (see Checchi and Roberts (2007) for 5

6 methodological concerns, and Guha-Sapir and D Aoust (2011) forareviewoftheliterature). Two types of excess mortality need to be distinguish. First, we consider violent deaths as direct excess mortality. Violent deaths include death of the military during fightings, or civilians directly affected by bombings or gun shots. Second, a large part of excess mortality is also due to poor living conditions during conflict, following infrastructure breakdown, increased prevalence of infectious diseases, food insecurity etc. These deaths are considered as indirectly caused by the war. During civil conflict, indirect deaths tend to bear the highest burden. After a sudden and intense phase of violence at the onset of fightings, during which violent deaths are the main cause of mortality, indirect deaths increased following the destruction of buildings, fields and the breakdown of public services (Coghlan et al., 2006; Geneva Declaration Secretariat, 2008; Degomme and Guha-Sapir, 2010). Let us now go a step further in order to identify who will be more likely to be affected by one or the other type of mortality. In Rwanda, direct deaths resulted from targeted violence. The leaders and the young educated men were more likely to die during the 1990s (De Walque and Verwimp, 2010). While infrastructure breakdown affects the entire population, women and children are more vulnerable and are therefore more likely to bear the indirect cost of conflict (Guha-Sapir and Panhuis, 2004; Tabutin and Schoumaker, 2004; Guha-Sapir and D Aoust, 2011) Migration When studying migration, one has to distinguish between internal displacement, and international displacement, implying to have crossed the border. Moreover, two main motives have been highlighted : forced migration, mostly in situation of complex emergency ; or migration for economic motives which are essentially driven by better economic opportunities elsewhere. Finally, some migrations only involve one member of the household while others are undertaken by entire families. Timing additionally matters, seasonal migrations differ from longer term resettlements. Population movement have been shown to increase in times of civil war. Insecurity and political instability can become determinant factors in the decision to migrate. Violence and economicopportunitiesareoftenlinked. Lautze (1997)distinguishesspontaneousmigrations following military operations or sudden outbreak of extreme violence from the migrations for which potential benefits have been thought about, and for which daily insecurity has been weighted. Migration can therefore be a coping strategy amongst others. Asset, buildings, stocks, land owning or having a paid job therefore matter for households when deciding to migrate (Ibanez and Velez, 2008; Williams, 2008; Lindley, 2009). Another element forcing migration follows from the threat of persecution. The historical perspective has pointed out the fact that elites were particularly persecuted in the Great Lakes region (De Walque and Verwimp, 2010). In general, brain drain is a common phenomenon in conflict affected countries, where political instability has urged high skilled 6

7 workers to migrate. As for mortality, it is important to know whether migration is selective, and to distinguish household migration from individual migration, or from household dislocation in which different members leave to different destinations. Household composition will be affected differently according to these characteristics. Consequences can be economic, if the household productive capacity has been altered, as well as demographic if migration has affected fertility or marriage Fertility Household composition changes with new births, and therefore fertility levels. These differs from mortality in that they are partially determined by individual choices. Despite an average decrease in fertility in Sub-Saharan Africa in the last twenty years, trends are far from being linear across time and depend on biological, psychological and economic factors. In conflict settings, affected populations evolve to cope with adversity, and changes may translate to changes in fertility. The impact of insecurity on fertility is ambiguous. Some studies have shown that fertility tend to decrease following episode of violence, as stress increases and health deteriorates. Violence can additionally force migration, and hence lower the frequency of sexual intercourses. Moreover, as uncertainty increases, investments in education are dampened, and marriages postponed (Lindstrom and Berhanu, 1999; Blanc, 2004). A fertility boom in postconflict settings can also be explained by a reversal trend in these factors (Agadjanian and Prata, 2002). Inversely, violence can have a positive effect on fertility, through either an insurance effect or a replacement effect, both reinforced by insecurity. First, the insurance effect follows from uncertainty towards the future, and reflects the fact that having more children ensures a minimum income stream (Agadjanian and Prata, 2002; Verwimp and Van Bavel, 2005). Second, in contexts where child mortality peaks, high fertility compensates the risk of losing achild(kalemli-ozcan, 2003) Marriage When analyzing marriage decisions, we are mostly interested about their timing. For that matter, researchers look at the age of women at their first marriage. It becomes clear that both short term and longer term impact of violence should be disentangle. In the short run, violence can directly affect marriage decisions, but the direction of the effect has been shown to depend on the context and the determinants affected. Insecurity can lead to selective mortality and migration. As soon as there appears a differential in gender representation in the population, the marriage market will be affected. If the gap is important, such perturbation will last for future generations (Jayaraman et al., 2008). Shemyakina (2011) suggests that the deterioration of economic conditions has an 7

8 ambiguous impact on age at marriage. On the first hand, if employment opportunities for women are poor, they tend to marry earlier. On the other hand, conflict, impoverishing families, prevent men to have enough funds to pay the bride price, and they will postpone marriage. There is a cultural component in marriage, and most developing countries, women have to be pure" when they marry. Valente (2011) shows that in Nepal, women living in exposed areas were more likely to be married before the age of 15, and explains this result by the fact that their physical integrity was being threatened following frequent abductions by the Maoists. In the longer run, violence can indirectly affect marriage through its long lasting effects of changes in mortality, migration and economic conditions, but also through its impact on education, which has been shown to influence the timing of marriages. Women s education offers them greater emancipation opportunities, and therefore postpone marriage. In civil conflict, schools may be destructed and teachers kidnapped, and hence force parents not to send their children to school. Return to education are less valuable in such uncertain context. Moreover, savings on schooling spendings, and additional revenues from an increased labor force can also be beneficial for parents. These factors are negatively associated to schooling, and hence may increase early marriages in the future (Jayaraman et al., 2008; Valente, 2011) Conceptual framework In the last sections, we have attempted to present the different channels of impact through which civil war can affect household composition. While these particular channels have been studied in the context of civil conflict, household size and composition have not. To the best of our knowledge, only one study by Winters et al. (2009) coversthistopic,buttheyassess the impact of economic shock on households, whereas we will study the effect of violence shocks. In this paper, we are looking at whether households affected by violence during the conflict in Burundi had an impact on household size and composition. Since the peace agreement of 2005 was the first that translated into changes on the ground, we will particularly look at two periods : before 2005 and after The occurrence of violent events in the hills sampled decreased steadily from 2005 onwards, as figure 2 shows. There are different measure of household composition. We first look at the size of households. Then, we decompose households in different groups according to age and sex. These are defined on the basis of the age of household members in 2005, and their sex. Given the age and sex of the persons affected, different hypotheses regarding the channels through which violence alters households composition can be made. For example, as underlined before, young men are more likely to have been targeted, and women and children to suffer from the deterioration of living conditions. The data does not allow us to identify precisely the channels at play, but we can point out the potential mechanisms driving changes in certain age-sex groups. Children will be more affected by changes in mortality and fertility patterns, while adults will be affected by migration, marriage and mortality. These are not 8

9 mutually exclusive, and there exist interaction between them, which we will discuss when commenting the results. The past and present trends of demographic indicators in Burundi indicates that there are substantial differences between rural and urban areas, age and education. In the estimations, we will account for differences in trends across provinces, and control for age and education of the head of household. In normal situations, these would be the major determinant of demographic processes. We are mostly interested in the impact of violence on household composition, holding the main determinants constant. This indicator has been constructed relying on data collected at the community level. It consists of the aggregated number of violent events declared by the village chief. It is a measure of occurrence and not of intensity, which we were not able to compute due to the incompleteness of such declarations. Let us now turn to the empirical analysis, starting with a description of the data and our identification strategy. 4. Data 4.1. Data sources The analysis draws on different types of data. The first dataset consists of a panel of households. The second dataset is a community questionnaire, collected at the village level. The first dataset constitutes a three-round household survey undertaken in Burundi. Figure 3 shows the timing of data collection. The first round is a Multiple Indicator and Cluster Survey (MICS) undertaken in September The second round, known as the Questionnaire des Indicateurs de Base du Bien-être (QUIBB), was collected in February It did not aim at building longitudinal data but nevertheless used the same sample as the MICS survey. The last round, undertaken in April 2010, only retained 3 provinces of the MICS/QUIBB sample: Bubanza, Bujumbura Rural and Cibitoke, located in the North-West of the country. The choice of these provinces is justified by the high level of violence in the region over the last years, as well as by budgetary constraints. The survey is characterized by a 2-stage cluster sampling. In the first stage, 88 hills were sampled and in the second stage, 15 households were interviewed in each primary unit. It resulted in 1320 households in the MICS survey in our three provinces. Attrition reduced this number to 1284 households in the QUIBB survey. For the 2010 survey, interviewers received instruction to track these 1320 households in the sampled hills and re-interview them 4. In addition, they also interviewed the newly formed households of the sons and 4 In order to maximize the reliability of the data, we trained interviewers for a week, not only on the questionnaire but also on the historical context of the conflict. After the training, interviewers were selected on the basis of an exam and simulated interviews. The questionnaire was tested during a pilot study in an out-of-sample hill. We assigned teams of five interviewers, each including a team leader and at least two women. Each interviewer did two interviews per day on average. The questionnaires were then controlled for accuracy and entered in a CS-PRO program by data entry agents. 9

10 Figure 3: Timeline : Conflict onset Arusha Peace Agreement CNDD demobilization FNL demobilization MICS QUIBB PANEL daughters of the head of household (the split-offs), as it is current practice in panel surveys in Africa (Verwimp and Bundervoet, 2009; Beegle et al., 2008). During this third-round survey, 1222 households were interviewed in 85 hills 5. Those include not only the one sampled in but also 158 split-offs, who account for 13% of our sample. Compared to the 2005 survey, we traced 80.6% of the households that were interviewed in the first round. During the 2010 survey, the interviewers also undertook a community survey in each hill, in which they collected information on population and violence. This community data provides contextual information and controls for each hill surveyed. In particular, the data on violent events will be the variable of interest of our analysis. For the purpose of the analysis, we will only retain the 2005 and the 2010 surveys, as the MICS and the QUIBB were collected at very close intervals. This choice follows the fact that the MICS data were collected in the end of 2005, right before the second phase of the peace process, which began in early 2006 when the FNL joined the negotiations. Additionally, the 2010 data used the 2005 sampling frame. We will nonetheless rely on the QUIBB survey to run quality checks on the data. In the analysis, we will also exclude the newly formed households, which have particular characteristics. Headcount according to the sex of the head of household, his education and his matrimonial status are presented in table Variable of interest Let us review the variable used in the analysis. We will first go thought household composition indicators, before turning to conflict proxies and finally to review the set of control to be included in the estimations. Household composition In Figure 3, wepresenttheage-sexdistributionofthe2010 survey, along with the corresponding masculinity ratio. In general, age declarations are 5 There were three hills in which we could not track households, all located in Bujumbura Rural. In two hills, the villagers reported not to know the tracked households, either because they had migrated or were invented by 2005 interviewers. The remaining hill was not secure enough to conduct the survey. 10

11 questionable and this may explain some irregularities in the distribution, particularly above 75 for which masculinity ratio peaks. Some observations are worth making before turning to multivariate analysis. In figure 4(b), we observe a surplus of women in the sample,as the ratio is mostly below unity. In figure 4(a), there is a slight gap in children between 10 and 14, for both sex. These were born between 1996 and 2000, period in which the conflict was quite intense. In order to check whether the gap is not mostly due to wrong age reporting, we plotted the number of individuals between 10 and 20. We do not see a peak at age 15, which would point out some displacement between the 14 and 15. On the contrary, headcount at 14 are larger than at 15. We could expect a peak at age 18, which stems from the fact that some pretend to have reached majority. This would not affect the category. What could have happen is some displacement from 19 to 20, which would imply an underestimation of the 15-19, which would then underestimate the gap between the and the We will therefore take a closer look at this dip. We will further look at different age-sex categories, determined according to the timing of the conflict. We have defined six categories, based on population age in First, the sample will be divided into two : the individual less than 15 (children) and above 15 (adults). Second, following our observation on the population pyramid, we will take a closer look at the 5-9 category (who were in 2010). They were born between 1996 and 2000, which was a period of intense violence. Then, we regrouped the young adults (30-54), and further disaggregated them according to sex. They had 18 to 42 in Descriptive statistics of these indicators in 2005 and 2010 are shown in table 3. The table also includes averages of the attrited households. Civil conflict proxies In the first place, we rely on the declarations of the village chief, to whom we asked, year by year, how many violent event happened in his village since We then aggregated these to come up with an index measuring the occurrence of violence in the hill. On average, communities living in the hills sampled faced 26 events per year. The frequency of events decreased between the two periods considered ( and ), following the first cease fire. During the first period, we recorded 36 events per year, compared to 11 for the next period. Recorded events for each province are shown in table 2. Figure5 shows the distribution of violence across hills. On average, 5.5 events took place during the period Over the period , the average occurence is In the analysis, we will use this threshold to divide our sample into two, according to the degree of affectedness.we then constructed a second proxy of civil conflict using the Armed Conflict Location and Event Dataset [in progress]. Controls We will control for the characteristics of the head of household, mainly his sex, age, matrimonial status and education attainment. Descriptive statistics of these indicators are shown in table 3. 11

12 5. Empirical analysis 5.1. Identification strategy The identification strategy is based on econometric methods for panels of two periods. Before turning to such models, we will first present a cross-sectional analysis of the 2005 data. For such analysis, we assume that households had similar demographic trends before the onset of the war. If this is true, households have evolved in an homogeneous environment until It seems reasonable that in these three provinces of rural Burundi, demographic characteristics in 1993 were similar, as the entire country had been affected by small-scale, sporadic events before the war. These three provinces are located close to each other, and have very similar profiles in the DHS 1987 data. The 2005 cross-sectional model to estimate is the following: Y i = + 1 V l, V 2 l, X i + i, (1) with the dependent variable, Y i is the number of member in household i belonging to an age-sex category. Such categories are household size, members below 15, members above 15, children between 5 and 9, and adults between 30 and 54, the latter for each sex. The variable of interest is V l, ,theviolenteventsthathappenedbetween1993and2005inhilll, where household i lives. We include a quadratic term to capture the possible non-linear effect of violence. X i regroups a set of socio-demographic controls. This regression will be estimated with the TOBIT model, which is appropriate for dependent variable that takes only positive values. Our household composition indicators are indeed truncated at zero or even one (for household size), and their distribution is continuous for positive values. The TOBIT model will restrict prediction to positive values, which makes more sense than the approximations of linear models. This model is estimated by maximum likelihood (Wooldridge, 2002). Then, in a second step, we will use the panel dimension of the data to assess the impact of violence on the evolution of household size. The following model will be estimated by OLS: Y i = + 1 V l, V l, X i, i, (2) where the dependent variable is now the difference, between 2005 and 2010, in household members belonging to one age-sex category, and we added a variable V l, capturing violence between 2005 and X i,2005 is a set of control at the initial time period. We had underline the presence of attrition when describing the data, which could lead to bias in our estimations. This is a common issue in civil conflict settings, as people 12

13 or households are more likely to die or migrate. Let us discuss the problems that can arise following the presence of attrition. Attrition could bias the estimations through a selection bias, calling into question the randomness of the sample. In our case, attrition is aproblemonlyifitisdifferential, this is, if an effect could be attributed to an unequal loss of households across the hills with more or less demobilized ex-rebels. In order to test if attrition is differential, we constructed a dummy measuring the probability of being sampled in both rounds. The dummy takes the value 1 if the household was interviewed in 2005 and 2010, and 0 if the household was interviewed in 2005 only. For the dependent variables considered in this paper, we regressed this dummy on the violence and the control variables, as well as on the interactions between them, according to equation (3). P il ( ) = Z i, V l, Z i,2005 V l, i (3) where P il ( ) takes the value 1 for household i in hill l if he was sampled during both years, and 0 otherwise ; Z i,2005 is a vector comprising the set of controls in 2005 (further denoted X i,2005 )andthehouseholdcompositionindicatorsin2005(furtherdenotedy i,2005 ), and V l, is the number of violent events that occurred in hill l between 1993 and Attrition is expected to bias the OLS/TOBIT estimations of our models if the coefficients associated with the interaction terms are significantly different from zero. In particular, 3, the coefficient associated to the dependent variable of our structural equation interacted with violent events, should not be significantly different from zero. The results of this regression are presented in table 4. The coefficients associated with the interaction terms are not significant, which supports the use of OLS estimators Results Cross-section analysis The estimation of equation (1) arepresentedintables5 to 8. In table 5, weestimated the impact of violent events on household size. In column (1), we estimated equation (1) for the entire sample, in column (2), we restrict ourselves to less affected households ; in column (3), only highly affected households are considered and finally column (4) adds a quadratic term for the latter households. The threshold defining which household have more or less affected is 5 events, as discussed in section 4. Intables6, 7 and 8, weestimatedtheimpact of conflict on the size of the age-sex groups for the whole sample, the less affected and the more affected respectively. The coefficients of TOBIT models cannot be directly interpreted as in the linear model. To obtain the partial effect of a one unit increase in x j,theyhavetobeadjustedbyafactor 6 We presented the results for household size. The other age-sex categories were also tested for signs of differential attritions, and proved to be robust to attrition (results not shown). 13

14 x. This j = j x The estimations reported are the ones adjusted to obtain average partial effects. Coming back to table 5, wefindthathouseholdsthathavebeenaffected by one additional event are smaller on average, by member. This means that on average, households living in hills where there has been 35 events would be smaller by one member than households living in peaceful hills 7. If we restrict our sample to the households that were less affected, the coefficient stays negative but becomes non-significant (p-value 0.186). However, turning to the most affected households, the impact of event seems to take an inverted u-shaped trend. The results suggest that for an increase of violent events up to 12, household size increase on average. When the occurrence of event goes beyond this threshold, the impact becomes negative. Figure 6 presents the linear regression for the whole sample (column (1)), and the quadratic estimation for the most affected (column (4)), together with confidence intervals. In the literature, the overall impact of civil conflict on household size should be negative, except in the case where the insurance effect observed for fertility patterns is strong. Our results could reflect such phenomena. It would imply that the number children in households would be on average positively affected by violence. Let us turn to estimations of household composition to assess whether we indeed find such an effect for the young categories. In table 6, welookattheimpactofviolenceonthenumberofmembersindifferent age-sex categories. For these regressions, we restricted our sample to households for whom age declaration were coherent across surveys 8. These estimations first shows that for both the sub-sample of children (< 15) and adults (>15), violence has had a significant negative impact on the headcount in these categories. In the < 15 group, the impact is a slightly higher. For children from 5 to 9, the impact of conflict is negative but not significant. Turning to the individuals that were young during the conflict intense phase, we find no significant effect for the whole sample. This may partially be explained by lack of power coming from the low variability of smaller age-sex classes. When we observe only the hills that were less affected, all effects are negative but none is significant (Table 7). On the contrary, when turning to the most affected hills (Table 8), we find that men aged are less numerous in 2010 in hills where battles happened more 7 Note that the most affected hill declared 19 events 8 For example, a child between 5 and 9 years old in 2005 could not have 35 in The entire household will be excluded from the regression related to this category. For some categories, such as the adults, the exclusion criterion is necessarily less strict. 14

15 frequently. They had between 18 and 42 during the conflict, and were more likely to die or migrate. This effect translate into the effects on the > 15 and on the in general. Moreover, what children are concerned, coefficient associated to violence are positive and close to significance (p-value = 0.11 for the 5-9 years old). This may corroborates the insurance effect explanation Longitudinal analysis The estimations of equation (2) arepresentedintables9 et 10. Table 9 looks at the impact of violence on the change in headcount between 2005 and Let us start with household size (column (1)), before decomposing. The recent events did not significantly affect the number of members in the households. Coefficients related to battles between 2005 and 2010 do not show a clear pattern. However, past events affected significantly household size change, which decreased between both surveys. Intuitively, if one household of average size (5 members) would have witnessed one additional event per year between 1993 and 2004, his size would have decreased by 4% between 2005 and While the cross-section analysis indicated that more affected household were smaller, the panel analysis indicates that their size decreased more if hills more frequently affected. If we focus on children, the past events ( ) do not influence headcount (not for < 15 nor for the 5-9). The transversal analysis had shown that the numbers of 5-9 were higher following hostilities. The panel analysis shows a stable trend between 2005 and The differential observed in 2005 stayed between the two waves. Following recent events, the 5-9 have not significantly changed. Violence however seems to have impacted the < 15 headcount which has decreased on average. The negative impact grows as we go further in the frequencies, compared the peaceful hills. The impact attains -0.4 members in the most affected hills (more than two events between 2005 and 2010). In other words, if we consider a household composed of 2.5 young individuals (which is the average in 2005), it would mean a 15% decrease of the < 15 headcount. Unfortunately, decomposing further this age class increases considerably standard errors, which translates in loss of precision in our estimates. This may explain why we do not find any impact in 5 years interval age class but well in the < 15 class. If we turn to adults between 30 and 54, we observe a negative impact associated to recent events in hills affected by more than two events between 2005 and Compared to households living in peaceful hills, the affected households have decreased by males between 30 and 54. We cannot determine the cause of such fall, but it may be that they have migrated or been killed following hostilities. We had underlined a differential in 2005, which is growing in time. For adults aged > 15, wefindapositiveimpactofrecentviolenceonheadcountinhills affected by more than two events. We think that the increase in such members may be due to the demobilization program that started in 2005 in Burundi. The program has led 15

16 to the return of ex-rebels in their households during these years. Following this intuition, we have reproduced estimations of equation 9 for young between 15 to 24 years old in These estimates, presented in table 10, indicatethathouseholdslivinginhillshaving witnessed two or more battles in the last five years had 0.2 to 0.3 more members of than the households from peaceful hills. This effect is significant only for males. Related to the literature on civil conflict, violent hills could have been populated by more youth, which would render our coefficient endogenous. An exogeneity test does not confirm this problem. Another possible explanation is that males that joined the rebel forces came from the affected hills. Indeed, recruitment is known to be more important where rebel group are active, and hence where hostilities are (Collier and Hoeffler, 2004). They came back in these hills from 2005 onwards, when the first peace agreement was signed. The hills that were affected the most between 1993 and 2004 are also the ones whose chief declared the highest number of events. Ex-combatants would have come back to their village of origin, which were the most affected, explaining the positive relationship between violence and headcount of Unfortunately, we do not have valid instruments that could help us to provide substantial estimates of the causal impact of violence on the youth. 6. Conclusion In this paper, we provide a first attempt to study the impact of civil conflict on household size and composition in rural Burundi. Despite the fact that the structure of our data does not allow us to assess the impact of war on demographic phenomenon such as mortality or fertility, we proposed some thoughts on channels of impact of civil conflict on households evolution. Through a cross-section analysis, we first showed that households living in most affected villages are smaller in size. This corroborates with the literature on civil conflict, which predicts high mortality and migration and low fertility, all leading to smaller household size. Then, relying on the panel dimension of the data, we suggest that household size decreased more between 2005 and 2010 if they have witnessed more hostilities in the past. Some categories of individuals seem to be more affected than others. This is for example the case of children, women or young men. Looking at household composition, we find that headcount aged < 15 in 2005 were less represented in more affected hills. We also find such a negative effect on adults between 30 and 54 years old. Restricting the sample to only most frequently affected villages, the effect is stronger. More recent violence additionally led to a decrease in males aged in Such an effect is not significant among women, suggesting that males were more targeted. As decomposing may lead to small numbers and lack of precision, we were not able to provide more refined argument what the channels are concerned. A larger sample or more precise questions would be needed in order to study channels of impact for each age-sex 9 Rebels age generally falls in this range. 16

17 category. Our results however follow the prediction found in the literature, and underline differential in composition. These particularly appear at vulnerable ages, according to whether households have been more or less afflicted. The literature also identifies complex processes emerging during civil conflict, such as the insurance effect when studying fertility patterns or more generally the heterogeneous impact of violence. If we only consider the most affected villages, we find non-linear effect related to past violent events. Our estimations suggest that civil conflict may have had a positive impact on the children headcount in households living in more violent-prone areas. Finally, we have underlined a positive relationship between the number of males aged and hostilities since This may be explained by the return of ex-combatant in their households, following the launch of the demobilization program. This result should be further tested in an instrumental variable approached to correct for endogeneity bias. From then on, the interdependency between all demographic phenomena deserves to be further investigated, on the basis of specific data related to each question raised. 17

18 References Agadjanian, V., Prata, N., War, Peace, and Fertility in Angola. Demography 39, Beegle, K., De Weerdt, J., Dercon, S., Migration and economic mobility in Tanzania: Evidence from a tracking survey. The Review of Economics and Statistics 93, Blanc, A.K., The Role of Conflict in the Rapid Fertility Decline in Eritrea and Prospects for the Future. Studies in Family Planning 35, Bruck, T., Schindler, K., Brück, T., The Impact of Violent Conflicts on Households: What Do We Know and What Should We Know about War Widows? Oxford Development Studies 37, Brunborg, H., Tabeau, E., Demography of Conflict and Violence: An Emerging Field. European Journal of Population/Revue européenne de Démographie 21, Bundervoet, T., Verwimp, P., Akresh, R., Health and civil war in rural Burundi. Journal of Human Resources. 44, Checchi, F., Roberts, L., Interpreting and using mortality data in humanitarian emergencies - A primer for non-epidemiologists. Humanitarian Practice Network 44. Chrétien, J.P., Les racines de la violence contemporaine en Afrique. Politique Africaine 43, Chrétien, J.P., Le Défi de l ethnisme: Rwanda et Burundi, Karthala, Paris. Chrétien, J.P., L Afrique des Grands Lacs. Deux mille ans d histoire. Technical Report. Paris. Chrétien, J.P., Mukuri, M., Burundi, la fracture identitaire: logiques de violence et certitudes ethniques, Karthala, Paris. Coghlan, B., Brennan, R.J., Ngoy, P., Dofara, D., Otto, B., Clements, M., Stewart, T., Mortality in the Democratic Republic of Congo: a nationwide survey. Lancet 367, Collier, P., Hegre, H., Hoeffler, A., Elliot, V.L., Reynal-Querol, M., Sambanis, N., Breaking the conflict trap: civil war and development policy. World Bank and Oxford University Press, Washington DC, USA. Collier, P., Hoeffler, A., Greed and grievance in civil war. Oxford Economic Papers 56, De Walque, D., Verwimp, P., The Demographic and Socio-Economic Distribution of Excess mortality during the 1994 Genocide in Rwanda. Journal of African Economies 19, Degomme, O., Guha-Sapir, D., Patterns of mortality rates in Darfur conflict. The Lancet 375, Fearon, J.D., Laitin, D.D., Ethnicity, Insurgency, and Civil War. American Political Science Review 97, Geneva Declaration Secretariat, Global Burden of Armed Violence, in: Geneva: GD Secretariat, Geneva: GD Secretariat. Geneva: GD Secretariat, Geneva, Switzerland. Gleditsch, N.P., Wallensteen, P., Eriksson, M., Sollenberg, M., Strand, H., Armed Conflict : ANewDataset.JournalofPeaceResearch39, Guha-Sapir, D., D Aoust, O., Demographic and Health Consequences of Civil Conflict. Background paper for the World Development Report on conflict, security and development. Guha-Sapir, D., Panhuis, W.G., Conflict-related Mortality: An Analysis of 37 Datasets. Disasters 28, Ibanez, A., Velez, C., Civil Conflict and Forced Migration: The Micro Determinants and Welfare Losses of Displacement in Colombia. World Development 36, Jayaraman, A., Gebreselassie, T., Chandrasekhar, S., Effect of Conflict on Age at Marriage and Age at First Birth in Rwanda. Population Research and Policy Review 28, Justino, P., The Impact of Armed Civil Conflict on Household Welfare and Policy. IDS Working Papers, Kalemli-Ozcan, S., A stochastic model of mortality, fertility, and human capital investment. Journal of Development Economics 70, Lautze, S., Saving Lives and Livelihoods: The Fundamentals of a Livelihoods Strategy. Technical Report. Feinstein International Famine Center. Boston. Lemarchand, R., Patterns of State Collapse and Reconstruction in Central Africa: Reflections on the Crisis in the Great Lakes. Africa Spectrum 32,