The impact of refugee experiences on education: evidence from Burundi

Disclaimer/Complaints regulations If you believe that digital publication of certain material infringes any of your rights or (privacy) interests, please let the Library know, stating your reasons. In case of a legitimate complaint, the Library will make the material inaccessible and/or remove it from the website. Please Ask the Library: https://uba.uva.nl/en/contact, or a letter to: Library of the University of Amsterdam, Secretariat, Singel 425, 1012 WP Amsterdam, The Netherlands. You will be contacted as soon as possible.


Introduction
The number of displaced persons worldwide is currently at its highest level since the Second World War. More than 65 million people around the globe were forcibly displaced in 2015, of which approximately a third (21 million) were displaced internationally (i.e. refugees). The vast majority of refugees reside in neighbouring developing countries (United High Commissioner for Refugees 2016a). The consequences of displacement for those affected are significant and frequently longlasting and affect multiple aspects of human life. One important aspect that displacement experiences can affect is access to education. Several studies suggest that forced displacement is one of the key channels through which conflict can have a detrimental impact on schooling outcomes (Chamarbagwala and Moran 2011;Justino et al. 2014;Verwimp and Van Bavel 2014). Most studies have, however, focused on internally displaced persons (IDPs) while there is scarce evidence on the impact of forced displacement on the education of refugees.
One of the main reasons for the scarcity of evidence on the impact of refugee experiences on schooling is the lack of datasets that include a large sample of individuals who experienced international displacement and their contemporaries who did not, so that the educational outcomes of both groups can be compared. This type IZA Journal of Development and Migration of analysis is only possible in countries which experienced a large outflow of refugees and a large inflow of returnees after the end of the conflict. This paper makes use of a nationally representative survey recently conducted by the authors in Burundi, a country which has experienced these two flows. The survey was conducted during early 2015 and involved interviews with 1500 households resident in 100 communities. The selected communities were distributed across all the provinces of the country according to the demographic weight of these provinces in the 2008 Census. A community representative was also interviewed in each of the communities. In addition, in households with returnee members, a randomly selected returnee was selected for an in-depth interview about experiences while in displacement and upon return.
In this paper, we explore differences in schooling outcomes between returnees, defined as individuals who were displaced to a neighbouring country and later returned home, and stayees, defined as individuals who never left the country during the conflict (i.e. those who were never displaced and those who were only displaced internally). Given the low levels of schooling in Burundi, we focus on the completion of primary education and explore differences in the impact of refugee and stayee experiences across different schooling cohorts.
Refugee experiences may have negative as well as positive impacts on schooling. For instance, when children are physically fleeing, which may be of greater or lesser duration, they do not have access to education. Before settling in a new location, children frequently end up in a transitory situation where they may not have access to schools. The family could also settle permanently in a remote area with no schools. Moreover, children in displacement are often more likely to become infected with certain diseases (Connolly et al. 2004), experience food shortages (Dharod et al. 2013) and rely on coping mechanisms such as early marriage, all of which may have a negative impact on education (Oh and van der Stouwe 2008). International displacement might also be related to greater loss of property and wealth during the conflict (Fransen et al. 2017) and a need for children to get involved in income-generating activities.
On the other hand, refugee experiences could lead to better schooling outcomes compared to those of children who, for various reasons, do not leave their country of origin when conflict erupts. Refugee children have a right to protection and assistance, including the right to a basic education, as stated by the 1951 UN Refugee Convention relating to the Status of Refugees (United High Commissioner for Refugees 1951). Many refugee camps or refugee hosting areas are therefore equipped with primary education facilities, often provided for and/or financed by NGOs or international agencies (United Nations Educational, Scientific and Cultural Organization 2011). Refugee children also frequently have more access to humanitarian assistance and other sources of support than stayees, particularly when they reside in camps. Children may end up being hosted in countries that have better education systems than those back in their country of origin as well. In contrast, their contemporaries in the home country need to rely on their national government to provide services such as education. Many conflict-affected states lack the capacity and/or willingness to provide these services (UNESCO 2011). Children who stay behind are also more likely to be conscripted and experience higher levels of insecurity, two factors that have been shown to have negative consequences for human capital acquisition (Blattman and Annan 2010).
Burundi experienced a civil war between 1993 and 2005. The conflict resulted in an estimated 300,000 casualties and an estimated 700,000 refugees (Ngaruku and Nkurunziza 2005). The majority of refugees settled in camps in Northwestern Tanzania (Fransen 2015;Ruiz and Vargas-Silva 2015, 2017Whitaker 2002). The United Nations High Commissioner for Refugees (UNHCR) supervised and sponsored the schools in refugee camps in Tanzania (Skonhoft 2010). The education system in Burundi was seriously affected as a result of the war, as national primary enrolment rates plummeted by close to 15% during the conflict (World Bank 2016). The extent to which refugee schooling outcomes differ from those who never left the country is unknown. Given that the large majority of Burundians displaced abroad by the 1993-2005 conflict had returned home in 2015 (Harild et al. 2015;Fransen 2015;Fransen et al. 2017), it is possible to compare the schooling outcomes of returnees with the outcomes of their contemporaries in Burundi with our data.
Our results suggest that, controlling for pre-war characteristics and cohort effects, returning refugees are 16 to 28 percentage points more likely to have finished primary school than their contemporaries who never left the country. The result is driven by individuals who were affected by displacement during their school age years. We also find that an additional year spent as a refugee while school aged increases the likelihood of finishing primary school by four to six percentage points. These findings correspond with reports which suggest that children who were of schooling age during the conflict and who were displaced internationally had better access to education facilities than those who stayed in Burundi (Integrated Regional Information Network 2002). We also provide a simple comparison of the schooling outcomes of returnees with those of residents of Kagera, a region of north-western Tanzania that borders Burundi, and there is suggestive evidence that returnees are better off than their hosts in Tanzania. While we cannot completely rule out that some of our findings are driven by pre-war differences between returnees and stayees (i.e. those who never left the country), we show that the results are robust to the inclusion of multiple controls for pre-war economic conditions. We also conduct a placebo test to support this conclusion.
The rest of the paper is structured as follows. The next section explains the rights of the displaced to primary school education and discusses the existing evidence. The third section presents the historical background. The fourth section presents the data and methodology. The fifth section presents the main results of the paper. Section 6 presents complementary evidence from the survey on experiences while in displacement. Section 7 presents a comparison of stayees with the outcomes of Tanzanians from the same schooling cohort. Section 8 presents a series of robustness tests, and the last section concludes.

Displacement and schooling outcomes: rights and previous evidence
The 1951 Convention Relating to the Status of Refugees establishes the right to primary education for refugees. In particular, the Convention states that host governments should ensure that refugees are given the "same treatment as is accorded to nationals with respect to elementary education" (UNHCR 1951). Moreover, UNHCR has a mandate to protect refugees, which includes the provision of education (Waters and LeBlanc 2005). However, in many cases there is a substantial gap between the legal right to education of refugees and the actual provision of such education (Dryden-Peterson 2015a). For instance, a study conducted on Syrian refugee children in Lebanon showed that 80% of them did not attend school in 2013. Similarly, 56% of Syrian school-age children did not attend school in Jordan in that year. School dropout rates and class failure rates were also significantly higher among refugee children as compared to local children (UNICEF 2015). Insufficient access to education is particularly likely in cases of urban displacement, as many urban schools are already stretched and lack space for new pupils. Often it is also unfeasible to build new schools or expand existing ones in urban areas. United High Commissioner for Refugees (2016a, b, c) estimates that only about half of refugee children worldwide had access to primary education in 2015.
In the estimations, we compare returnees with all those who remained in Burundi during the conflict. This includes those who were displaced internally (IDPs) as well as those who did not leave their communities of origin during the conflict. In the case of IDPs, the responsibility to provide primary education lies with national authorities. As explained by Justino (2011), educational facilities in IDP camps are not very common and the provision of this service "is typically disorganised, when it exists at all." National authorities are also responsible for providing education to those children who never leave their communities of origin. While these children are not affected by displacement, they suffer from other detrimental consequences of conflict for education, including the destruction of schools, killing and exodus of teachers, household income shocks and decreases in state investments on education (ibid.).
There is a substantial literature which has explored the overall impacts of conflict on schooling outcomes (Akresh and De Walque 2008;Chamarbagwala and Moran 2011;Di Maio and Nandi 2013;Ichino and Winter-Ebmer 2004;Lai and Thyne 2007;Leon 2012;Shemyakina 2011;Valente 2014), but just a few studies have explored the specific impact of forced displacement on schooling outcomes. This evidence mostly refers to internal displacement and suggests that these experiences have major negative consequences for schooling outcomes. For instance, Justino et al. (2014) estimated that, in Timor Leste, experiencing displacement decreased school attendance by 8.5 percentage points. For Burundi, Verwimp and Van Bavel (2014) estimated that the probability of completing primary schooling declined by 2 percentage points for every year spent in a camp. However, Verwimp and Van Bavel (2014) used data from 2002, before the large wave of refugee return to the country. That means that they are mostly measuring the negative impact of internal displacement experiences and are not capturing the impact of international displacement experiences (more on this in the next section).

Historical background of Burundi
Burundi is a small country in the African Great Lakes region that has been one of the poorest in the world for years. The country held the 184th place (out of 188) in the Human Development Index in 2014 (United Nations Development Programme 2015) and gross national income per capita was just USD 270 in 2014, which is substantially lower than the average for sub-Saharan Africa (UDS 1699). Burundi is also densely populated. It occupied the third place in population density in Africa in 2015 with 435 people per square kilometre of land area (World Bank 2016), which is much higher than the average for sub-Saharan Africa (42 people per square kilometre). Even though agricultural land is scarce, approximately 90% of Burundians depend on subsistence farming as their main source of income and nutrition (World Bank 2015).
Burundi's history is characterised by tensions between the country's two main ethnic groups: Hutus and Tutsis. These ethnic tensions are part of a complex and multifaceted power struggle that has led to large-scale conflict. In 1993, the events that led to the biggest conflict in Burundi's history started when Melchior Ndadadye became the first democratically elected Hutu president of the country. He was assassinated a few months later by Tutsi soldiers. The assassination led to a long civil war that lasted from 1993 to 2005 (Ngaruko and Nkurunziza 2005). Although there had been previous conflict episodes in Burundi, such as the one in 1972, the scale and intensity of the 1993-2005 conflict set it apart from earlier ones. Whereas previous violent episodes were limited to certain provinces, the 1990s war was a countrywide conflict.
Hundreds of thousands Burundians fled the country during the conflict. As shown in Fig. 1, while many Burundian refugees initially went to the Democratic Republic of Congo (DRC) and Rwanda, Tanzania, was by far the main host of refugees. 1 Burundians who fled to Tanzania were settled in refugee camps in the north-western part of the country (Fransen 2015;Ruiz and Vargas-Silva 2015, 2017. Refugee experiences in Tanzania typically lasted long, with an average duration of 10 years (Fransen et al. 2017).
Living conditions in camps in Tanzania differed across sites but were generally better than those in Burundi during the war. Still, many refugees experienced serious hardship. For example, Burundian refugees in Tanzania were not provided with agricultural land (Harild et al. 2015). This was in contrast to previous cohorts of refugees, such as those who fled the country in 1972. Mobility and economic activities of the refugees were also restricted. After arrival of the refugees, the Tanzanian government announced that refugees were not allowed to go further than 4 km from the camps and that they were not allowed to work outside the camps or engage in agricultural work in the camps (Millner 2013). The majority of the 1993 refugees consequently became entirely dependent on the support of international donors and NGOs during their stay in Tanzania (Harild et al. 2015).  Importantly for this study, primary schools in refugee camps in north-western Tanzania were funded by UNHCR, which paid for teacher salaries (Amnesty International 2005). It is estimated that around 90% of primary school-age children who arrived in Tanzania after 1993 were enrolled in school in 2000 (Jackson 2000). Qualitative studies suggest that Burundian refugees were highly motivated to send their children to the schools in camps, particularly the Hutus who felt they had been previously discriminated in the Burundian schooling system (Skonhoft 2010). Moreover, in the past, educated Hutus were one of the main targets of the Tutsi dominated Government and education was often seen as a liability (Nkurunziza and Ngaruko 2002;Skonhoft 2010;Verwimp and Van Babel 2014).
Dryden-Peterson (2015b) explains that Burundians who settled in Tanzania following the 1972 conflict were integrated into the national educational system, using a Tanzanian curriculum taught in Kiswahili and English, which are the main languages of Tanzania. The overall goal was to facilitate the integration of these refugees in the host country. However, there was no political will to integrate refugees from the 1993 conflict. Schools in camp areas shifted to a Burundian curriculum taught in Kirundi and French, which are the main languages of Burundi. Hence, while UNHCR supervised the schools, Burundian educators were in charge of developing the education system in the refugee camps (Skonhoft 2010). The overall goal was to facilitate the return of these refugees to Burundi.
It is estimated that the number of internally displaced reached 800,000 in 1999 (United Nations Office for the Coordination of Humanitarian Affairs 1999). Internal displacement experiences tended to be short and lasted approximately 1 year (Verwimp and Van Bavel 2014). Living conditions in the displacement camps within Burundi were generally poor. The majority of settlements lacked basic services such as clean drinking water and health care facilities (Zeender and McCallin 2002). Burundi's government was responsible for funding educational facilities in the camps, and reports suggest that at least 50% of school-age internally displaced children did not go to school (Integrated Regional Information Network 2002).
The Arusha Peace Agreement was signed in August 2000 and led to the end of the conflict a few years later. In 2005, Burundians elected Pierre Nkurunziza, a Hutu, as President of the country, reinforcing the conditions of the peace agreement. Following the end of the conflict, Burundi experienced a large wave of return of its displaced population. Estimates suggest that over 500,000 Burundians returned from Tanzania between 2000 and 2015 (Fransen 2015). This is a considerable number for a country that had a population of only 6.7 million in 2000. Moreover, during this period, Tanzania stopped the provision of education for Burundians in order to encourage refugees to return home (Dryden-Peterson 2015b).
In 2005, the Burundian Government announced that primary education in public schools would be provided for free from the following academic year. The gross primary enrolment rate increased from 82% in the 2004/2005 academic year to 101% in the 2005/2006 academic year (Sommeiller and Wodon 2014). There was a substantial increase in enrolment rates in all provinces of Burundi.

The survey
The survey was conducted across all provinces of Burundi during January to March 2015. 2 A total of 1500 households were interviewed in 100 communities. The primary unit of interest was the sous-colline ('sub-hill' in French), which is the smallest administrative unit in the country. There are an estimated 8000 to 9000 sous-collines in Burundi, but there were no national lists available. As such, the first step was to select 100 collines (the second smallest administrative unit), which were distributed over the 17 provinces of the country according to the demographic weight of these provinces in the 2008 Census. Each colline typically consists of between two and ten sous-collines. Within each colline, a sous-colline was randomly chosen to conduct the interviews in. Within each sous-colline, 15 households and one community representative were interviewed. Figure 2 shows the communities/sous-collines that were surveyed across Burundi.
Primary schooling age in Burundi ranges from 7 to 12 years of age. We limit the analysis to individuals who became of schooling age in 1973 or later and who were 12 years of age in or before 2014. 3 This means that the sample is limited to individuals who were between 13 and 49 years of age in 2015. As shown in Table 1, this results in a sample of 3712 individuals. Of those 3712 individuals, a total of 858 (23%) individuals belong to the pre-war schooling cohort (born between 1966 and 1980), 1887 (51%) belong to the war schooling cohort (born between 1981 and 1997) and 967 (26%) belong to the post-war schooling cohort (born between 1998 and 2002).
Refugee experiences were recorded at the individual level in the survey. A person was defined as a returnee if the person had moved internationally with the primary purpose of escaping conflict or political persecution and had resided in another country for a consecutive period of at least 3 months. As also shown in Table 1, 8% of those in the sample are returnees (308 individuals). This share is higher for the pre-war schooling cohort (14%) and smaller for the post-war schooling cohort (2%). Close to 9% of those in the war schooling cohort are returnees. Table 1 also indicates the share of individuals in each education cohort that finished primary school. First, note that only 15% of those in the pre-war cohort finished primary school. The share is similar for returnees and stayees, suggesting that there was not a strong selection into international displacement based on previous educational outcomes (more on this below). For the war generation, 35% finished primary school. The number is similar across returnees and stayees, although slightly higher for returnees. Finally, for the post war generation, returnees have a much higher primary school completion rate. However, it should be noted that the sample of returnees for the post-war generation is small.

Becoming a returnee versus stayee
The literature on forced displacement suggests that in situations of conflict, those individuals from wealthier families can travel further and choose better destinations to settle in (Fransen et al. 2017;Van Hear 2006. This means that differences in educational outcomes between returnees and stayees could be driven by pre-conflict characteristics. In this sub-section, we explain why this is less likely to be the case for Burundi and introduce our approach to reduce this potential bias. First, all individuals in the sample were born in Burundi, and in 2012, Tanzania repatriated any remaining refugee camp residents to Burundi (Ruiz and Vargas-Silva 2016). As such, there is no selection issue in terms of returning home or staying in Tanzania, as the latter option was not available for those who fled the 1993 war. Any selection issue relates to being a refugee in the first place.
Second, it is important to note that substantial evidence indicates that exposure to conflict in Burundi was random (Uvin 1999). For instance, Voors et al. (2012) show that the type of violence experienced in Burundi was largely exogenous to household characteristics and local economic conditions. The authors, for example, test whether violence was affected by the likelihood of profit for the aggressors, measured as the   possibility of stealing assets (including livestock), and ethnic considerations such as the share of the local vote for the assassinated president. They find no support for these possibilities, which suggests that households had equal chances to be affected by the conflict in Burundi and that targeting of households based on certain household characteristics did not occur. Third, most refugee migration took place by foot. Because previous studies have shown that exposure to conflict was random, we assume that displaced individuals are more or less evenly distributed across Burundi, with proximity to the border being the most important determinant of international versus internal displacement. Fransen et al. (2017) show that this is actually the case and that distance to the border of Tanzania led to a higher probability of international displacement and a lower probability of internal displacement. The map in Fig. 3 also provides insights in this regard. The map depicts the number of Burundian refugees in Tanzania in the mid-2000s, per province of origin in Burundi (in brackets), based on UNHCR estimates. The second number in each province corresponds to the number of refugees from that province, as a share of the entire population of the province that was recorded in the 1990 Census (in parenthesis). Figure 3 therefore provides evidence that most of the refugees came from provinces in Burundi that border Tanzania.
We created several indicators in order to highlight that there were no major pre-war differences between stayees and returnees. First, we created a household level variable which reflects the average primary schooling outcome (i.e. primary schooling completion dummy) of those members of the household who are from the pre-war schooling cohort. As shown in Table 2, pre-war primary education was similar across the two groups (11% for returnees, 13% for stayees). This difference is not statistically significant.
We also explore possible differences in pre-war wealth using the information on prewar livestock and pre-war land holdings. 4 We use Tropical Livestock Units (TLUs) to standardise livestock ownership across individuals. Bundervoet (2009Bundervoet ( , 2010 conducted an exploration of the impact of conflict on livestock in Burundi. We replicate his analysis and use the following units as weights: 1 cow/ox = 1 TLU, 1 sheep = 0.17 TLU, 1 goat = 0.17 TLU, 1 pig = 0.25 TLU, and 1 fowl = 0.01 TLU. 5 As shown in Table 2, the difference in the average pre-war livestock and land holdings of returnees and stayees are not statistically significant. In Table 2, we also present results from a regression for the pre-war cohort in which the dependent variable is a dummy for finishing primary school and returnee status is one of the independent variables. The regression also controls for gender, age and province of birth. This regression provides a placebo test for the results regarding refugee experiences. This cohort was over primary school age by the start of the war, and we would not expect their schooling outcomes to be affected by future refugee experiences. The lack of an effect in this group would give additional support for the idea that there are no pre-existing trends that are accounting for differences between returnees and stayees. As suggested by Table 2, the coefficients is not statistically significant. In Table 2, we also show results from regressions in which pre-war education (i.e. household measure), pre-war livestock and pre-war land are the dependent variables. Again, returnee status is not significantly related to these variables.
The main estimations include controls for pre-war educational, livestock and land indicators in order to adjust for possible pre-war differences. While these estimations reduce the concerns about selection bias, we cannot fully rule this possibility out (e.g. those who become refugees could value human capital more, even if they have similar characteristics to others). Therefore, we avoid making any strong claims of causality in the discussion of the results.

Estimation
In order to study the impact of refugee experiences on schooling outcomes, we start by estimating the following model: Where S i is a dummy indicating the person completed primary school, α p are dummies for province of birth, W_PW i is a dummy indicating that the individual is from the war or post-war cohort (i.e. pre-war cohort is the control category), X i are controls for gender, age and controls for pre-conflict characteristics (i.e. livestock, education and land). In the Note: data reflects the situation in mid-2000. Source of information on refugees in Tanzania is UNHCR. Data on share of the province population constructed with information from the 1990 Census. Note that Muramvya and Mwaro were one province during the period of interest. baseline estimations, R i is a dummy which indicates that the individual is a returnee. The coefficient of interest in this case is δ, which is similar to a differencein-difference estimator.
In (1), the effects for the war and post-war cohort are combined. In a second step, we separate these effects. That is, we replace W_PW i by W i , which is a dummy indicating that the individual belongs to the war cohort, and run the estimation including only individuals from the pre-war and war cohort. Likewise, we also replace W_PW i by PW i , which is a dummy indicating that the individual belongs to the post-war cohort, and run the estimation only including individuals from the pre-war and post-war cohort.
Finally, we include a variable (Y i ) in the estimation which indicates the numbers of school years for which the individual was displaced and estimate: Given that we are controlling for returnee status and schooling cohort, the parameter ρ provides something comparable (although not equal) to a differences-in-differences estimator. Please note that all variables included in the estimations refer to predisplacement factors and are not affected by refugee experiences. Finally, all estimations are presented with clustered standard errors at the sous-colline level. Table 3 provides descriptive statistics of the independent variables. The sample is slightly more female (53%), a fact that also holds for each of the cohorts, and the average age varies per cohort by construction.  We also make use of the information from the in-depth interviews with the returnees to put the results in the context of experiences abroad and use data from Tanzania to compare the schooling outcomes of returnees to those of Tanzanians from the same schooling cohort. Finally, we show a series of robustness tests, paying particular attention to the control for conflict in the estimation. Table 4 provides the baseline results of differences between returnees and stayees on the likelihood of finishing primary school. The main variable of interest is the interaction of returnee status with the schooling cohort variable. In the estimation without controls for pre-war characteristics, the coefficient of the interaction term is positive but not statistically significant (panel A, column 1). This coefficient becomes significant when we include the pre-war controls. For instance, the estimation with the pre-war education levels suggests that returnees from the war and post-war cohorts are 25 percentage points more likely to have finished primary school than their non-returnee counterparts (panel A, column 2). The estimated gaps are smaller in the estimations which control for pre-war livestock and pre-war land (22 and 16 percentages points, respectively) and higher for the estimation which includes all pre-war controls (28 percentage points) (see columns 3 to 5).

The effect of returnee status
As explained above, we would expect the main effect of returnee status to be on those in the war cohort, as this group was more likely to have attended schools in camps. For those in the post-war group, the effect is more indirect. In order to explore this in Table 4, we also present results comparing each group directly to the pre-war cohort (see panel B and panel C). The results suggest that, as expected, the results regarding schooling and returnee status are driven by those in the war cohort.

Number of years displaced while of school age
In Table 5, we focus on the number of years for which the individual was displaced while of schooling age. This is a better measure of exposure to schools while in displacement than the simple cohort dummy variable. Given that we are also controlling for school cohort and returnee status, the coefficient on the number of years displaced while of school age also provides some intuition similar to that of a differences-indifferences estimator. As suggested by Table 5, an additional year spent as a refugee while school aged is associated with an increase in the likelihood of finishing primary school of four percentage points. Including pre-war controls increases the estimated impact slightly to six percentage points. Table 6 provides information on the schooling experiences of refugees while in displacement. This particular information is only available for one randomly selected returnee per household. While this is a small sample, we still get a good indication of the experiences of different cohorts. First, note that those who were above schooling age when displaced and those who returned before schooling age did not accumulate much schooling while abroad (an average of 0.06 and 0.23 years, respectively). On the other hand, those who were of schooling age accumulated about 1.72 years of education abroad.

Experiences while in displacement
Also, and perhaps more importantly, the survey collected information for returnees and former IDPs on whether there was a primary school in the community of displacement. The information for IDPs was only collected for those who were adults before displacement (that is, the pre-war generation). As shown in Table 6, 71% of the returnees stated that there was a primary school in their community of residence  Pre-war education X X Pre-war livestock X X

Pre-war land X X
Pre-war education is only available for those households with members from the pre-war generation. t statistics are presented in parenthesis *Indicates that the coefficient is significant at the 10% level; **indicates that the coefficient is significant at the 5% level; ***indicates that the coefficient is significant at the 1% level abroad (mostly camps). As such, we can corroborate the availability of educational facilities for many refugees while in displacement. On the other hand, this was only the case for 54% of the IDPs.

Comparison with hosts
In this section, we use data from the Kagera Health and Development Survey (KHDS) to explore if the primary school completion rate of Burundian returnees is very different from that of their hosts in Tanzania. Kagera is the most north-western region of Tanzania (see Fig. 4). It borders Burundi and was one of the main destinations of Burundian refugees in Tanzania. The KHDS is representative of the population of the region and has been used by multiple papers to explore the consequences of hosting refugees (see Baez 2011;Maystadt and Verwimp 2014;Ruiz and Vargas-Silva 2015, 2017. The last round of the KHDS was conducted in 2010. As such, we can compare the outcomes of residents of Kagera in the same schooling cohort as our "war" generation in Burundi (i.e. who were of primary school age from 1988 to 2009).
The KHDS data suggests that only 28% of Kagera residents in that cohort finished primary school, a smaller proportion than the one for the returnees in our sample (37%). Pre-war education is only available for those households with members from the pre-war generation. t statistics are presented in parenthesis *Indicates that the coefficient is significant at the 10% level; **indicates that the coefficient is significant at the 5% level; ***indicates that the coefficient is significant at the 1% level While it is not possible to make strong conclusions from this comparison, the finding suggests that returnees could be better off in terms of schooling than both Burundian stayees and residents of north-western Tanzania. It is important to highlight that Kagera is one of the poorest and most remote regions of Tanzania and that primary school completion rates are much higher in other parts of the country.

Robustness
One of the main concerns about the estimation is whether we are controlling adequately for conflict experiences. If this is not the case, it is possible to argue that the refugee experience indicator reflects the impact of several other factors related to conflict exposure. Columns 1 to 5 in panel A of Table 7 present the estimations if we control for the number of years in which the individual was of schooling age during the conflict, instead of simply including a dummy for schooling cohort. This change does not affect the main conclusions from the analysis. It is still the case that spending longer as a refugee while school aged is associated with better schooling outcomes. 6 In panel B of Table 7, we account for the fact that the conflict did not affect all provinces at the same time or for the same length of time. We follow the same approach of Verwimp and Van Bavel (2014) to create a variable in which exposure to conflict varies by province and age cohort. These authors argue that the spatial spread of the conflict was determined by geography and natural endowments. In this case, an individual is assumed to be exposed to conflict during school age if he/she had resided in a province that was affected by conflict and was of school age when the province was affected by conflict. Following Verwimp and Van Bavel (2014), we construct the conflict variable using the estimates from Bundervoet (2009) on the percentage of people whose fathers were killed during the initial stage of the conflict (i.e. above and below the median) 7 and Chrétien and Mukuri's (2000) account of the spread of the conflict for the later stages. The results do not change if we use this alternative way of controlling for conflict exposure.
In Table 8, we test the robustness of the results by employing propensity score matching (PSM) techniques in order to match returnee individuals with a comparable group of stayees. In this case, the treatment (T) is being a returnee. As we explained above, the large majority of refugees from the 1993-2005 conflict returned home Page 15 of 20 before our data collection. Hence, the treatment is essentially being a refugee in the first place, a factor that was largely determined by distance from the border of Tanzania.
We start by estimating a probit model to predict the likelihood of being a returnee based on age, gender and province of birth, and then we match individuals based on treatment status. Once we check for the balancing properties and common support across the treatment and comparison group, we proceed to use the nearest neighbour estimation matching procedure. With the matching at hand, the difference in the outcome variable is calculated to estimate the average treatment effect of the treated. As shown in Table 8, the results support the idea that returnees are more likely to have finished primary school than stayees.

Conclusions
In this paper, we studied the effects of refugee experiences on primary education and explored if the educational outcomes of individuals with refugee experiences in Burundi differed from the outcomes of those who did not leave the country during the 1993-2005 civil war. Despite the increasing academic interest on the well-being of displaced populations worldwide, the relationship between displacement experiences and  Pre-war education X X Pre-war livestock X X Pre-war land X X Notes: Pre-war education is only available for those households with members from the pre-war generation. t statistics are presented in parenthesis ***Indicates that the coefficient is significant at the 1% level educational outcomes is a relatively underexplored topic and even fewer studies have focused on the consequences of international displacement on education. Our survey was conducted 15 years after the signing of the peace agreement in Burundi and after the return of most former refugees to the country, which enables a long-term perspective on the impacts of displacement on education, including the role of education in displacement camps abroad. Our findings show that, once we controlled for pre-war characteristics of the households, former refugees who returned to Burundi had better schooling outcomes than their contemporaries who never left the country. This finding most likely reflects the varying levels of access that children had to education during the war. While children who stayed home were likely to be affected by the negative impacts of conflict on schooling (e.g. destruction of schools, killing and exodus of teachers, child soldiering, household income shocks, higher levels insecurity and decreases in state investments on education), those in neighbouring countries, and particularly those who resided in camps in Tanzania, had access to UNHCR-funded schools. We also provide a simple comparison of the schooling outcomes of returnees with those of Tanzanians, and there is suggestive evidence that returnees were better off than their hosts in Tanzania, again probably because of the specific schools that they had access to by virtue of being refugees.
Although the higher likelihood of completing primary school can be seen as a positive side effect of the refugee experience, the reality is that the primary school completion rate for returned refugees was still low (37%). These findings align with current concerns about the access to education of displaced populations during conflict times (United High Commissioner for Refugees 2016b). With the number of displaced populations in the world on the rise, an increasing number of children do not have access to education. The impact of refugee experiences on education is likely to have implications for future labour market outcomes and, more generally, durable peace after the end of conflict. More emphasis is therefore needed on providing primary education to refugees. However, our findings highlight that particularly the children who stay behind when conflict erupts suffer serious gaps in their education, which indicates that there is an additional need for educational support programmes that allow these children to catch-up with those who were not as affected by the war.
Endnotes 1 Note that there were a considerable number of refugees from Burundi in Tanzania before the events of 1993. These refugees fled the violence in 1972. These refugees were given land for cultivation and, by all accounts, were largely self-sufficient (Thomson 2009). 2 In 2015, over 200,000 people were displaced from Burundi to neighbouring countries (United High Commissioner for Refugees 2016c). This is the first episode of large displacement in the country in over a decade. The displacement is the result of increasing tensions and violence in response to the April 2015 announcement that the president of Burundi was running for a third term in office. Many interpreted a third term in office as a violation of the Arusha peace agreements. The data collection for this article was finalised approximately 6 weeks before the president's announcement and before this new wave of tensions and displacement. 3 As explained above, there was a conflict in Burundi during 1972. This conflict was of a smaller scale compared to the conflict from 1993 onwards but could have nonetheless affected the outcomes of those who were of schooling age at the time. Therefore, we limit the sample to those who became of schooling age in 1973 or afterwards. 4 There is no information on the current status of the pre-war livestock and land (i.e. whether it was lost, sold or still in possession). 5 TLUs allow animal species of different average size to be compared by a common unit. These measures are based on the typical weight of the animal raised to the power of 0.75 (also known as the metabolic body weight), compared with the equivalent figure for an animal of 250 kg. Please note that this measure relies on both species being under the same feeding system (which is a reasonable assumption in our case) but does not account for the possibility of different breeds of the same species. 6 Please note that the variable measuring exposure to conflict measures the number of years of schooling age that the individual was exposed to conflict (i.e. based on timing of the conflict), while the displacement variable indicates the number of years of schooling age that the individual was displaced abroad. 7 Bundervoet (2009) adjusts the estimates for bias related to households in which all members were killed.