Risk factors associated with Rohingya refugee girls' education in Bangladesh: A multilevel analysis of survey data

In Bangladesh, the world's largest refugee settlement currently shelters approximately one million Rohingya refugees who fled Myanmar to escape military persecution. Educating a significant number of young Rohingya, roughly half of whom are female, presents a significant challenge. Despite the presence of learning centres (LCs) across refugee camps, Rohingya girls may encounter specific barriers to accessing education due to exposure to various risks, such as violence, child marriage, and trauma stemming from past military oppression. This paper investigates the association between these risk factors and Rohingya girls' likelihood of attending LCs, and how this association may vary across refugee camps. Using survey data and employing three ‐ level multilevel logistic regression models, I find that girls are less likely to attend LCs if they are at risk of encountering sexual abuse, child marriage, and psychological distress or trauma. These factors explain considerable variation in girls' LC attendance between camps and between households. In addition to providing more schooling opportunities to Rohingya children, priori-tising girls' safety, protecting them from forced and child marriage, and supporting their psychological well ‐ being require increased policy attention.


| INTRODUCTION
This study examines the barriers to educational access for young Rohingya refugee girls in Cox's Bazar, Bangladesh, home to nearly one million refugees in the world's largest refugee settlement (UNHCR [United Nations High Commissioner for Refugees], n.d.; Vince, 2020).Amid highly congested conditions, approximately 36% of these refugees, aged 5 to 17, fall within the primary and lower-secondary school age range (GoB [Government of Bangladesh] & UNHCR, 2023; IBE [International Bureau of Education], 2011).Educating this young population from the world's largest stateless minority (MacLean, 2019) remains a significant challenge.Traumatic experiences, including military persecution in Myanmar, exacerbate these challenges, with young girls particularly vulnerable to sexual violence, child marriage, and gender norms within refugee settlements (Haar et al., 2019;Nasar et al., 2022;Shohel et al., 2022).Using UNHCR (2021) survey data, this paper analyses factors hindering girls' access to education across the refugee camps in Cox's Bazar.
The study builds on previous research suggesting that while some low-and middle-income countries have favourable regulatory frameworks for refugees accessing education, others are less supportive or even prohibit access (Dupuy et al., 2022).Refugee education worldwide aims to provide quality education, a sense of belonging, and socioeconomic opportunities.Despite these noble goals, refugees often face exclusion from educational systems across different countries (Dryden-Peterson et al., 2019), exacerbating their existing vulnerabilities.I illustrate this issue by focusing on refugee girls.

| Rohingya educational facilities and barriers for girls
There are around 3400 temporary learning centres (LCs) providing education to Rohingya children in over 30 camps, with the United Nations Children's Fund (UNICEF) supporting 2800 of these centres (UNICEF, 2022).Despite these efforts, a significant disparity persists between Bangladeshi and Rohingya children in terms of being out of school due to various structural barriers such as limited funding, poor teaching quality, and the lack of credible curriculum certification (Rahman et al., 2022;Shohel, 2022).Figure 1 demonstrates that at the primary level, Rohingya children are about twice as likely to be out of school as their Bangladeshi peers, and around three times more likely at the lower-and upper-secondary levels. 1 To maintain a focused study on basic education, my analysis in this paper is limited to the primary and lower-secondary age groups.
Around 80% of Rohingya children aged 6 to 11 attend LCs with equal participation from girls and boys, although the gender gap widens as children approach ages 12-14 (UNICEF, 2022).Nonetheless, the Rohingya perceive these non-formal educational and skill-building initiatives as temporary (Habib et al., 2023;Hossain, 2023).
Refugee children also frequently drop out due to mistreatment from their teachers and peers, which includes severe physical assaults for minor errors (Rahman et al., 2023).Besides these issues, girls' education faces specific risks due to various factors.For instance, financial instability and uncertainty about settlement and citizenship contribute to a higher risk of prevalent child and forced marriage among girls (Melnikas et al., 2020;Uddin, 2021).Research shows widespread child marriage among Rohingya girls, influenced by perceptions of maturity, insecurity, family honour, preference for younger brides, and lax enforcement of the legal marriage age (Islam et al., 2021).

RESEARCH NOTE
Girls are often excluded from school as they are expected to prioritise household chores, reflecting societal norms (UNICEF, 2022).This expectation is reinforced by the fear and trauma associated with facing sexual violence outside the home while residing in Rakhine State, Myanmar (Haar et al., 2019;Rahman et al., 2023).Furthermore, Rohingya women and girls reportedly face sexual harassment from employers, violence at home from various sources including family members, community members, and camp authorities, and are sometimes coerced into prostitution due to economic hardship (Akhter & Kusakabe, 2014;Guglielmi et al., 2020).Reflecting on the above discussion, I investigate three research questions in the study.
1. How far are different risk factors associated with the attendance of Rohingya girls at LCs? 2. To what extent do these risk factors interact with each other in influencing girls' attendance at LCs?
The second question explores how the combined effect of risk factors, such as sexual violence and child marriage, influences girls' participation.
3. How do the risk factors contribute to differences in attendance at LCs between camps and households?
The third question reflects that with over 30 Rohingya refugee camps in Cox's Bazar district (UNOCHA [United Nations Office for the Coordination of Humanitarian Affairs], 2022), it is likely that educational resources and risk factors vary across these sites.Moreover, refugee camps are often viewed through a unidimensional lens.To better understand the microelements, it is essential to conduct a more thorough analysis of the different facilities and community characteristics within and between camps.

F I G U R E 1
Out-of-school children among Bangladeshis and Rohingyas.Note: I estimate out-of-school children based on the definition of the UNESCO Institute for Statistics (UIS) (n.d.).This means the proportion of children or young people in the official age range for a given level of education (e.g., primary) who are enrolled in school irrespective of the level.In Bangladesh, the official age range for education spans from 6 to 10 for primary, 11 to 15 for lower-secondary, and 16 to 17 for upper-secondary levels.However, the UNHCR data used for Rohingya children do not have a numeric age variable, but age categories.Hence, I used the available age range of five to nine for primary, 10-14 for lower-secondary, and 15-19 for upper-secondary education.While the age range slightly varies, the statistics overall do not deviate from the narratives in other sources (e.g., UNICEF, 2022).UNICEF (2019) and UNHCR (2021).UNESCO, United Nations Educational, Scientific and Cultural Organization; UNHCR, United Nations High Commissioner for Refugees.In this study, I utilise the UNHCR (2021) survey collected from 15,935 individuals in 3165 households and 33 refugee camps, which was made available upon my request.The sample was designed to be representative of the Rohingya population at the camp level (UNHCR, 2021). 2 For this study, I narrowed the analysis to two age groups: (1) 5-9 years old and (2) 10-14 years old.The dataset only includes discrete categories for age and these two age groups approximately correspond to the primary and lower-secondary education age ranges in Bangladesh.Primary education in the country spans ages 6 to 10, while lower-secondary education covers ages 11 to 15 (with junior secondary spanning ages 11 to 13 and secondary covering ages 14 to 15) (IBE, 2011).Hence, my analysis predominantly includes children in the primary and lower-secondary education age brackets.While I mainly focus on girls, I also conduct analyses on boys to compare results (presented in the online supplement).The small number of missing observations for girls constitutes less than 0.1% of the total.

| Attendance at LCs
The dependent variable in the study is whether primary and lower-secondary school-age girls attended LCs or not during the survey conducted in January 2019 (UNHCR, 2021).I create this variable by combining two binary indicators in the dataset: attendance at (1) a non-religious LC run by a nongovernment organisation (NGO) or the government, or (2) a religious learning space such as a madrasah (Islamic school).These facilities are provided by UNICEF, the Government of Bangladesh, and other NGOs (UNICEF, 2022).Table 1 illustrates that boys are more likely to attend LCs than girls.

| Risk factors
I include nine indicators of risk factors as the main independent variables in the study.These are binary variables indicating whether a girl is at risk of experiencing (1) sexual abuse and violence, (2) child marriage, (3) psychological distress or trauma, (4) child labour, (5) recruitment by armed groups, (6) being kidnapped, (7) violence in the community, (8) violence within the home, and (9) unsafe shelter.Child marriage (2) in the analysis indicates whether any individuals under the age of 18 within the household are currently married or about to be married.I use this as an indicator of potential risk since the normalisation of child marriage within the household may increase future risks.Additionally, I consider a binary variable denoting whether the shelter is unsafe for girls (9) as unsafe living conditions may pose a threat to their overall well-being.Table 1 presents the summary of the descriptive statistics showing that compared to boys, girls tend to be at a greater risk of experiencing certain factors such as sexual abuse/violence, child marriage, psychological distress and trauma, and violence at home.

| Control variables
I incorporate five control variables in the analyses, which are disability status, engagement in income generation activities, receiving remittances, age group, and household size.Additional details about these variables and the rationale for their inclusion are provided in Supporting Information S1: Appendix S2 of the online supplement.
To address RQ1 by examining the association between different risk factors and Rohingya girls' attendance at LCs, I employ three-level multilevel logistic regression models (also known as mixed-effects models).Level 1 includes individuals or children nested within households at level 2, which, in turn, are nested within refugee camps at level 3. I fit Equation (1), where A is the outcome variable indicating attendance at an LC by an individual i in household j and refugee camp site k .The expected outcome in this logit model is the probability of the response being equal to one, A ijk = 1, given the predictors in the model.is an intercept, β 1 is a coefficient vector for the nine risk factors R ijk , and β 2 is a coefficient vector for the control variables Z ijk , as described in the variable section.u j is the camp-level variance component with a distribution of u j ~N(0, σ 2 u ) and r ij is the household-level random intercept with a distribution of r ij ~N(0, σ 2 r ).To address RQ2 about the interaction of risk factors and their association with attendance in educational facilities, I fit Equation (2).Here, β 3 is a coefficient vector showing the interaction effects between different risk factors, that is, R A and R B .
To address RQ3 about the extent to which risk factors explain variation between camps, I, first, examine the random variance components of u j .Specifically, I check whether adding risk factors and control variables to different models changes the estimated variances.I also visually illustrate (in the online supplement) the relationship between the proportion of girls experiencing different risk factors in each camp and the camp mean of the predicted attendance at LCs.I compute the predicted attendance after running regression using Equation (1).

| RQ1: Risk factors and attendance at LCs
Findings suggest that girls who are at a higher risk of certain issues are less likely to attend LCs than those who do not face similar risks.Specifically, as illustrated in Table 2, Rohingya girls are about 3% points (p < .05)less likely to attend LCs when they experience sexual abuse and violence.This relationship persists even after controlling for other factors in model 3 and incorporating camp fixed effects in model 4. The inclusion of camp fixed effects allows me to account for any unobserved camp-specific characteristics that may potentially bias the results.The probability of attending educational facilities becomes even lower, 14% points (p < .001),when girls experience child marriage at home.However, once I add controls to the model the probability becomes 4.4% points (p < .05).
Moreover, I find that the risk of experiencing psychological distress or trauma lowers the probability of girls' access to education by 7.7% points (p < .01)before adding any controls and 5.4% points (p < .05)after controlling for relevant characteristics, and camp fixed effects.Other risk factors such as child labour, recruitment by armed groups, kidnapping, violence within the community and family, and unsafe shelter do not appear to be significantly associated with attendance at LCs.Nonetheless, I observe that disability is strongly associated with a decreased chance of attendance by around 37% points (p < .001),which stands at 39% points after using camp fixed effects.
The proportion of disabled children is very low in the study as shown in Table 1, yet they tend to be largely excluded from receiving education.
As mentioned, I also provide results analysing how similar risk factors might be associated with LC attendance by Rohingya boys.As shown in Table S1 of the online supplement, results are different for boys.I find that boys who have experienced child marriage at home and psychological distress or trauma are less likely to attend LCs.The result is surprisingly reversed for the risk of sexual abuse/violence, as it appears to be correlated with higher attendance among boys.One possible explanation for this is that engaging in learning activities may provide boys with a means to avoid or cope with the risk of violence.Consequently, being at a higher risk may motivate them to attend LCs more regularly.
Overall, the risk of experiencing sexual abuse and violence seems to play a distinctive barrier to girls' engagement in learning activities, a trend not observed among boys.

| RQ2: Interplay of risk factors and attending LCs
Based on the analysis in RQ1, I examine the interaction between three risk factors and their association with attending LCs.Here, I mainly interact the risk factors that appear significant in Equation (1) to further examine whether they have a compounded effect.These are: risks of sexual violence/abuse, child marriage, and psychological distress/trauma.
As presented in Figure 2, the results are consistent across three sets of interactions.Specifically, when Rohingya girls are both likely to be at risk of experiencing sexual violence/abuse and distress, they are least likely to attend LCs compared to other groups.This is the same when girls are both at risk of experiencing sexual violence/ abuse and child marriage, and distress and child marriage.Additionally, I include the results from interactions involving risk factors found non-significant as robustness checks in Figure S1 of the online supplement.The interaction effects are predominantly non-significant for other combinations, including those shown in Figure S1.
Nevertheless, in contrast to the findings for girls presented in Figure 2, I find different results in the case of boys.As exhibited in Figure S2 of the online supplement, the probability of attending LCs does not significantly differ by the interaction between the risk of experiencing sexual violence and distress, and between sexual violence and child marriage.The interaction between the risk of distress and child marriage shows that both those who are and who are not at risk of being distressed are significantly less likely to attend LCs irrespective of their risk of child marriage (with the risk being slightly higher for those facing child marriage).
In brief, the interplay of the risk of sexual abuse, child marriage, and psychological distress may likely compound the effect and reduce the likelihood of girls' attendance at LCs.This is not the case for Rohingya boys.

| RQ3: Risk factors accounting for camp variations in attending LCs
As shown in Figure 3, there is a significant variation in girls' attendance at LCs among refugee camp sites.The random intercept (u j ) in Table 2 also confirms this variation.The random effects part of Table 2 further shows Interaction between different risk factors for girls and their probability of attending a learning centre (LC).Note: This graph is estimated based on Equation (2).F I G U R E 3 Random intercept predictions of refugee camp ranking for girls' LC attendance.Note: This caterpillar plot is estimated based on model three in Table 2. Camp identifiers are shown on the right side of the 95 percent confidence intervals (CIs).
Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1468-4446.13117by London School Of Economics And, Wiley Online Library on [29/05/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Summary of the descriptive statistics.Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1468-4446.13117by London School Of Economics And, Wiley Online Library on [29/05/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License The standard assumption in the model is that A ijk has a Bernoulli distribution.αTAB L E 1Notes: (a) Descriptive statistics are weighted by household survey weights except the number of observations.(b) The percentage of the 'yes' or discrete response is reported for all variables apart from 'household size'.
Risk factors associated with school attendance of Rohingya girls (age 5-14).Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1468-4446.13117by London School Of Economics And, Wiley Online Library on [29/05/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Marginal effects are presented, instead of log odds, for all the covariate coefficients, that is, coefficients above the 'Random effects' line.The coefficients can be interpreted in terms of the predicted probability of girls' attendance at LCs.
(b) Variance explained compared to the baseline model (model 1).(c) Model four is based on a two-level model while using camp fixed effects.(d) *p < .05**p < .01***p < .001.
Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/1468-4446.13117by London School Of Economics And, Wiley Online Library on [29/05/2024].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License