Policies and interventions to remove gender-related barriers to girls' school participation and learning in low- and middle-income countries: A systematic review of the evidence

Lack of support for girls' education (Barrier 1)

Community norms and parental attitudes about innate abilities of and appropriate roles for girls may undermine schooling. Norms may discourage girls themselves from learning, affect academic performance and lead to premature school dropout (Eble & Hu, 2019; Warrington & Kiragu, 2012). At the same time, when financial resources are limited, parents with inequitable gender role attitudes may prefer to keep sons in school rather than daughters (Lloyd & Young, 2009). Normative beliefs may thus affect both parents' decisions about whether, and how long, to send their daughters to school, as well as girls' own academic aspirations, school performance and behaviours (Global Education Monitoring Report Team, 2018). Indeed, where inequitable gender roles are entrenched, the incentives for girls to attend, and perform well in, school may be lower than those for boys (Colclough et al., 2000). This may be the case particularly where overall levels of education are low, with the result that in settings that perform poorly on other dimensions of development, gender gaps in attainment are largest (Evans et al., 2021; Global Education Monitoring Report Team, 2015; Psaki et al., 2018). While numerous studies have established that girls in some LMICs perform more household chores than their brothers (e.g., Amin & Chandrasekhar, 2012; Singh & Mukherjee, 2017) the effect of gender differences in time use on educational outcomes, which has been examined with observational data, is less clear (Hedges et al., 2018; Rees, 2017).

Child marriage and adolescent pregnancy (Barrier 2)

Norms around age at marriage also intersect with education. One widely discussed potential consequence of child marriage is school dropout. A recent analysis that modelled the effect of successful interventions to reduce early marriage on education outcomes estimated “substantial” increases in secondary school completion (Rasmussen et al., 2019). It has been asserted that “the timing of early marriage almost always disrupts girls' education” (Mathur et al., 2003). Cross sectional analyses and qualitative studies have documented strong associations between early marriage/marital aspirations and school dropout in settings where marriage before age 18 is common (Prakash et al., 2017; Raj et al., 2019). Social norms frame sexual activity and schooling as incompatible for girls in many low- and middle-income settings (Clark & Mathur, 2012; Eloundou-Enyegue, 2004; Frye, 2017; Lloyd & Mensch, 2008). Yet, as a recent report assessing the economic effects of child marriage indicates, few studies have adequately measured the impact of child marriage on education outcomes, in part, because decisions about marriage timing and schooling are jointly determined (Wodon et al., 2017). Likewise, while adolescent childbearing clearly disrupts schooling, few studies in LMICs have determined whether the effect is causal. Child marriage and adolescent pregnancy are selective. The social and economic factors that predispose girls to marry early and bear children are also critical factors in premature school leaving including poverty, gender norms, perceived value of schooling and academic performance (Bajracharya et al., 2019; Lloyd & Mensch, 2008; Psaki, 2014). Studies have shown that girls who are behind in school may be more likely to engage in sexual activity, become pregnant and search for a marital partner (Clark & Mathur, 2012; Grant & Hallman, 2008). While observational data provide useful insights, randomized controlled trials (RCTs) that successfully reduce child marriage and adolescent pregnancy and also measure education outcomes may help identify their unique effects on school participation and learning among girls.

Lack of information on returns to education/alternative roles for women (Barrier 3)

One reason given for families' reluctance to invest in girls' schooling relative to boys' is that less economic benefit is expected from educating daughters. This appears to be the case particularly in settings where adolescent marriage is the norm, where opportunities for paid employment among women are lower than for men and where earnings for women even for the same jobs are lower (Colclough et al., 2000). Research has shown that perceived returns to education affect the demand for schooling; furthermore, where perceived returns are underestimated, interventions that provide accurate information may raise educational attainment (Jensen, 2010). In addition, expanding economic opportunities for women and providing new information on alternative roles for women may increase girls' educational aspirations and human capital investment. However, a recent systematic review identified only four studies meeting their inclusion criteria that provided information to families on the returns to education and because the four studies assessed different outcomes, the benefits of providing that information for school participation and learning were not clear (Snilstveit et al., 2015).

School-related gender-based violence (SRGBV) (Barrier 4)

Gender-based violence in and around schools is believed to have consequences for school attendance, learning and attainment for all children, but particularly for girls (Global Education Monitoring Report Team and UNGEI, 2015). There is an extensive literature documenting SRGBV, perpetrated by both students and teachers, including sexual, physical and psychological abuse of students at, and in transit to, school (DevTech, 2004; Global Education Monitoring Report Team, 2015; Global Education Monitoring Report Team and UNGEI, 2015; Leach et al., 2003, 2014). While considerable attention has focused on sexual violence by boys or male teachers against girls, SRGBV takes many forms (UNGEI & UNESCO, 2013). SRGBV is said to reflect local attitudes and practices regarding the acceptability of corporal punishment and other forms of violence (Barasa et al., 2013; Crooks et al., 2007; Leach & Humphreys, 2007) as well as underlying social conditions and unequal gender relations (Parkes et al., 2013). A review for the 2003/4 UNESCO EFA Global Monitoring Report categorized gender-based violence into behaviors and acts considered “explicit,” such as sexual harassment, rape and intimidation, and those considered “implicit,” including corporal punishment, bullying, verbal and psychological abuse, and teachers' use of “free labour.” The report asserted that because educational authorities had not taken sufficient action to combat SRGBV in many settings, especially sub-Saharan Africa, it had flourished “unchecked” and become “institutionalised” (Dunne et al., 2003). SRGBV is thought to be most widespread in settings where other forms of inequality or disability are common (Parkes et al., 2016). However, despite assertions about the detrimental effects of SRGBV on education outcomes for girls in these settings, a 2013 report from UNGEI and UNESCO concludes that “we know little about how this violence impacts retention and achievement. The link is still tenuous…” (UNGEI & UNESCO, 2013).

Gender insensitive school environment (Barrier 5)

While there are a number of studies documenting girls' treatment in the classroom relative to boys in high-income countries (e.g., Sadker & Sadker, 1994) systematic assessment of classroom dynamics in LMICs is less common. To the extent that classroom observations have been conducted, studies indicate that teachers interact more often with boys, which is said to lead to “passivity” among girls (Global Education Monitoring Report Team, 2015). In addition, it has been argued that girls may perform better with more collaborative learner-centred pedagogies in contrast to traditional teacher-dominated “chalk and talk” teaching practices that prevail in many low- and middle-income settings (Lloyd & Young, 2009; Mensch & Lloyd, 1998; Murphy-Graham, 2009).

In addition to pedagogical practices that may be detrimental to girls, if teachers have traditional views about gender roles, prefer teaching boys or have lower expectations of girls, particularly with regard to science, technology, engineering and mathematics (STEM) subjects, this may have an adverse effect on girls' education outcomes (Lloyd & Mensch, 1999). While it has been argued that gender sensitivity of teachers can help transform gender norms (Global Education Monitoring Report Team and UNGEI, 2015; Sperling & Winthrop, 2015), few studies have examined teacher attitudes and their effect on students (Lloyd et al., 2000; Mensch & Lloyd, 1998).

More so than other dimensions of the school and classroom environment, the role of female teachers in improving educational outcomes for girls has been examined. While female teachers are not necessarily more supportive of girls' education than their male counterparts (Mensch & Lloyd, 1998), it may be that parents feel more comfortable sending their daughters to schools with female teachers and/or that female teachers act as role models for their students (Global Education Monitoring Report Team and UNGEI, 2015). Rigorous studies have examined the effect of teacher gender on the gap in test scores between boys and girls. A prior review indicated that exposing girls in LMICs to female teachers may be beneficial for education outcomes (Lloyd & Young, 2009). On the other hand, an analysis of the gender gap in mathematics using administrative data from Chile, which included measures of the school and classroom environment, did not find the gender of the teacher explained the gap in performance in grades 4 and 8 which doubled during that period (Bharadwaj et al., 2012).

Lack of safe spaces and social connections (Barrier 6)

One potential barrier to adolescent girls' academic success is a lack of safe spaces to spend time and connect with their peers and trusted adults. Although they take different forms in practice, the goal of safe space groups, often led by female mentors, is to address social isolation, and often—through the delivery of life skills education—to build critical thinking and negotiation skills, address harmful gender norms, reinforce girls' commitment to staying in school, and/or strengthen support networks (Austrian et al., 2016; Temin et al., 2018). Proponents of these programs argue that the delivery of gender transformative content within an after school safe space platform could improve education outcomes for girls, especially in settings where girls' academic performance is weaker than boys. A recent narrative review of 30 “community-based girls' groups” another label for safe spaces, identified nine such programs in Africa and South Asia, all multicomponent, that investigated the effect on education outcomes. To be considered for the review, the program had to include adolescent girls, a trained female mentor, a community venue where regular meetings could take place, employ an experimental or quasi-experimental study design and be published between 2000 and 2017. Of the nine programs, six found significant effects on school participation or learning, which included numeracy, literacy, school enrolment, school retention and grade attainment. The nine programs measured a total of 18 education outcomes, of which 11 were significant (Temin & Heck, 2020). However, an important remaining question is whether safe spaces programs, on their own, are essential in driving improvements in education, or whether the content or structure of those meetings (e.g., the female mentor, the curriculum) or other components (e.g., cash or in-kind transfers) lead to improvements in education outcomes (Population Council, 2018).

Lack of teaching materials and supplies (Barrier 7)

Theoretically, lack of access to textbooks and educational materials is an obstacle to learning and may discourage enrolment and increase absenteeism. Girls may have differential access to teaching materials and supplies—for example, boys are more likely to own any type of phone and 80% more likely to own a smartphone compared to girls (Girl Effect and Vodafone Foundation, 2018)—this affects remote school participation and learning. There is also a potential gender issue in the form of textbook content. A comprehensive analysis of gender bias in textbooks found that it is a “hidden obstacle” to achieving gender equity in education in the global south (Blumberg, 2007). A review of studies documenting gender bias in textbooks for the UNESCO 2008 “Education for All Global Monitoring Report” described the following findings as being “near-universal” (Blumberg, 2007, 2008):

Women were found to be underrepresented relative to men;

A 2015 update noted minimal progress since the earlier period (Blumberg, 2015). While many education practitioners concerned with gender inequity assume bias in textbooks undermines education outcomes for girls—and research in high income countries suggests it may do so (e.g., Good et al., 2010)—initiatives to modify textbooks in low income countries are rare due to the time and expense of revision and replacement (Blumberg, 2007). Given this constraint, it has been suggested that teachers develop classroom exercises to engage students in identifying textbook bias (Blumberg, 2008). A qualitative analysis of an innovative curriculum in Honduras demonstrated that when gender is incorporated into the curriculum such that teachers and students are encouraged to challenge traditional views of gender roles and relations, more equitable constructions of gender can emerge (Murphy-Graham, 2009).

Insufficient academic support (Barrier 8)

Insufficient academic support for girls may be reflected both in difficulties with school performance (e.g., repetition, poor exam scores), as well as lack of confidence in one's ability to stay in school and navigate an academic environment. Some empirical evidence has shown that poor school performance predicts subsequent dropout for girls and boys, which may reflect failure to meet academic standards (Hartley & Swanson, 1986; Soler-Hampejsek et al., 2018), as well as girls becoming less invested in school and more likely to get married or become pregnant (Grant & Hallman, 2008; Marteleto et al., 2008). In settings where parents and community members already question the value of girls' education, poor performance may have more serious consequences for girls than boys. Previous reviews have examined the effectiveness of interventions designed to provide academic support to students but have largely lacked a gender lens. As part of a comprehensive review of the effect of school and teacher characteristics on education outcomes in developing countries, the effects of tutoring on education outcomes were positive in all four studies and significant in two (Glewwe et al., 2011). After school remedial education in core skills that target disadvantaged students and those lagging behind also appear promising, although results have not been systematically disaggregated by sex (Glewwe & Muralidharan, 2016; Snilstveit et al., 2015). To our knowledge, efforts to increase girls' academic self-confidence have not been reviewed previously, and are often incorporated into other program components, such as safe spaces or life skills education.

Inadequate sports programs for girls (Barrier 9)

Advocates of expanding girls' participation in sports in LMICs argue that sports challenge traditional gender scripts and transform the way girls view themselves and the ways their families and communities regard them. If true, supporters of these programs argue that expanding girls' access to sports facilities or equipment at school, as well as opportunities to practice leadership, engage in teamwork, develop physical skills, and/or increase girls' presence in outside areas where sports are played, may not only have positive effects on academic outcomes, but may also lead to increased empowerment (Brady, 2011, 2016; Sperling & Winthrop, 2015). Research on a natural experiment in the United States, taking advantage of a policy change (Title IX) that mandated equal opportunities for girls and boys to participate in sports, provides support for this relationship in high-income countries. One study found a significant positive effect of high school sports participation on college attendance among young women as well as a significant effect on labor force participation and employment in male-dominated occupations (Stevenson, 2010).

Inadequate health and childcare services

¹

Nutritional supplementation interventions at school are included as a separate barrier due to their effects on the household spending, in addition to child health.

(Barrier 10)

While there is reason to believe that the provision of health and childcare services at school would improve educational outcomes, particularly absenteeism, a review of seven studies published between 1993 and 2003 that investigated the association between school-based health clinics providing such services as vision and hearing screening, immunization, treatment for chronic and acute illnesses and mental health counseling and a variety of education outcomes including attendance, promotion, suspension, test scores and grade point average in the US concluded that there is “insufficient evidence to demonstrate a link” (Geierstanger et al., 2004). In addition to documenting considerable variability in findings across these studies, the authors noted numerous methodological limitations including the absence of study designs that control for confounding and selectivity. However, two more recent studies in the United States that used propensity score analysis to control for selection bias found significant effects of school-based health center use on attendance, school dropout and grade point average (Kerns et al., 2011; Walker et al., 2010). As far as the provision of childcare services is concerned, even in countries that have liberalized policies, returning to school as a mother is rare, at least in sub-Saharan Africa (Lloyd & Mensch, 2008). Thus, even if the provision of childcare services has an effect on education outcomes, it may not be large unless other systems are put in place to support adolescent mothers.

Inadequate life skills (Barrier 11)

Life skills education aims to foster social and emotional skills such as communication, empathy, resilience, perseverance, agency, and critical thinking, build knowledge about sexual and reproductive health and one's body, and/or foster more equitable gender attitudes, respect for all people's rights, and equal power in relationships. These skills and mindsets are believed to contribute to better schooling outcomes, success in the work world, and better health and wellbeing. Indeed, economists and psychologists have shown that the predictive power of traits such as self-control and perseverance may be as or more important for schooling, wages, and other socioeconomic outcomes than cognitive indicators such as IQ (Borghans et al., 2008; Heckman et al., 2013). Multiple reviews, primarily of studies from high income countries, have found significant associations between social and emotional skills and education outcomes, including meta-analyses of school-based interventions aiming to enhance students' social and emotional learning (Corcoran et al., 2018; Durlak et al., 2011; Taylor et al., 2017), as well as of after school programs (Durlak et al., 2010) and early childhood education (Blewitt et al., 2018). Sexuality and human immunodeficiency virus (HIV) education may contribute to improved schooling outcomes through multiple pathways. To the extent that sexuality and HIV education decreases pregnancy, teaches important social and emotional skills such as assertive communication and critical thinking, or fosters more equitable views on gender norms, it may contribute to improved education outcomes. An evaluation of the Teen Outreach Program in the United States, a curriculum-based program that addressed many of these issues, found reduced adolescent pregnancy and reduced school failure (Allen & Philliber, 2001).

Inadequate menstrual hygiene management (MHM) (Barrier 12)

Over the past 15 years, increasing attention has been directed towards inadequate MHM as a possible barrier to girls' education (Phillips-Howard et al., 2016; Sommer et al., 2016; Sommer, Hirsch, et al., 2015; Sommer, Sutherland, et al., 2015). Much of this attention has been driven by qualitative research in LMICs. In a systematic review and meta-synthesis of the qualitative literature on women and girls' menstrual experiences, Hennegan et al. (2019) produced an integrated model of menstrual experience. The model posits that socio-cultural factors—including social support, behavioral expectations and knowledge—and resource limitations that dictate girls' physical and economic environments, influence girls' social participation, education, psychological and physical health through six dimensions of menstrual experience: shame and distress; confidence; menstrual practices; perceptions of environment and menstrual practices; containment of blood, odours, and materials; and individual menstrual factors, including experiences of symptoms related to menstruation. Interventions and policies focused on enhancing MHM have largely focused on the provision of menstrual management materials, improvement in school water, sanitation and hygiene infrastructure, and delivery of puberty education (Bobel, 2019; House et al., 2012; Sommer, Sutherland, et al., 2015; UNICEF, 2019). An increasing number of observational studies, many using a cross-sectional design, have attempted to quantify the relationship between girls' menstrual experiences and their education outcomes. While some studies find that girls report that menstruation is negatively associated with school attendance or girls' ability to participate in school activities (Esen et al., 2016; Sivakami et al., 2019; Vashisht et al., 2018), others find little evidence for these relationships (Grant et al., 2013). Weak study designs, and varying study populations and challenges in accurately measuring absenteeism make drawing inferences about the level of absenteeism attributable to menstruation difficult (Grant et al., 2013).

Lack of water and sanitation (Barrier 13)

School-level factors dealing with water, sanitation, and hygiene (WASH) infrastructure have been cited as potential barriers to schooling for boys and girls given the importance of proper sanitation to avoid illnesses that cause absenteeism, to ensure children are properly hydrated during the school day, as well as potentially for gender-related reasons including safety and the need for privacy and running water during menses (Alam et al., 2017; Davis et al., 2018; Miiro et al., 2018; Sperling & Winthrop, 2015; Vashisht et al., 2018). Sustainable Development Goal 6 focuses on clean water and sanitation, with a target of achieving access to adequate and equitable sanitation and hygiene for all by 2030 (UN General Assembly, 2015). UNICEF recommendations for menstrual health and hygiene-responsive toilets in LMICs include: having sex-separate toilet facilities, water access at a facility, presence of a door, having locks on doors, and including a pit or bin for disposal within a facility (Morgan et al., 2017; UNICEF, 2019). In a survey of 227 schools across six countries—Ethiopia, Kenya, Mozambique, Rwanda, Uganda, and Zambia—Morgan et al. (2017) found that less than 20% of schools in each country had at least four of the recommended services. While separate-sex toilet facilities were most commonly reported, ranging from 44% in Ethiopia to 98% in Uganda, a median of 13% of schools had locks for latrine doors and a median of 9% reported having a water supply at the latrine facility (Morgan et al., 2017). Previous reviews focused on the effects of WASH interventions on education outcomes were unable to identify sufficient evidence of this relationship (Birdthistle et al., 2011; Jasper et al., 2012).

Inadequate school access (Barrier 14)

Children's access to school is hampered by lack of schools or schools that are too far away, overcrowded and poor quality, and schools that lack flexible hours. School access is thought to particularly impact girls due to intersections with gender biases such as parental concerns about girls' safety, the disproportionate burden girls carry in terms of caring for younger siblings and household chores, and beliefs that educating girls is not worthwhile. Interventions or policies designed to improve school access may aim to make it easier for students to get to school, or to increase the hours they spend in school (Glewwe & Muralidharan, 2016). Distance to school may act as a barrier to enrolment and attendance, especially for girls in settings where they are unable to walk or travel to school on their own. Even when schools exist in close proximity, there may not be enough space for all students who want to attend, and parents may not be inclined to send their children if school infrastructure is perceived to be low quality (Li & Liu, 2014; Lloyd et al., 2005; Sharma & Levinson, 2019). Further, the timing of classes may not work with students' schedules—for example, if girls need to take care of siblings or their own children (Sperling & Winthrop, 2015). Previous reviews have found evidence that addressing access to school may increase enrolment and attainment, and possibly learning, for students overall (Glewwe & Muralidharan, 2016; Snilstveit et al., 2015), but have focused less attention on the question of gender differences in the effects of these interventions, unless they specifically targeted girls.

Poor policy/legal environment (Barrier 15)

Policy and legal environments have the potential to address gender-related barriers to girls' schooling in two main ways: by guarding against the negative effects of gender norms and practices, such as forcing pregnant girls to leave school; and by directly removing barriers, such as school fees. In practice, such policies often take one of three forms: enacting free and/or compulsory schooling, allowing for automatic promotion (AP) between primary and secondary schooling, or allowing re-entry of pregnant girls and mothers into formal education. In areas with higher levels of poverty, without compulsory schooling policies in place, families may opt not to send girls to school and instead have them work for wages or in the home (Herz et al., 1991). A lack of AP policies may lead to high dropout between primary and secondary schools, especially among the most vulnerable (Ahmed et al., 2019). Without policies to formally allow for school re-entry, girls who experience early pregnancy may be shut out of the education system no matter their past school performance. Supplemental programs that provide information to policymakers and local stakeholders to improve attitudes and information about the school re-entry policy may help reduce remaining barriers to girls' re-entry into school (Sperling & Winthrop, 2015; Walgwe et al., 2016).

Inability to afford tuition and fees (Barrier 16)

Families that have limited financial resources must choose how they allocate those resources, and for many the immediate benefits of girls' education may not outweigh the costs of schooling (Herz et al., 1991; Sperling & Winthrop, 2015). Tuition and school fees may be prohibitively high for poor families, and as mentioned above, parents may opt to keep boys in school rather than girls when there are greater financial constraints (Lloyd & Young, 2009). This is particularly the case during income shocks, which can exacerbate already unequal access to schooling between boys and girls (Björkman-Nyqvist, 2013). Keeping girls out of school for these reasons can feed into a cycle of lower expected earnings for girls, which puts the next generation in a further disadvantaged position schooling- and earnings-wise (Moore, 2005). Many countries adopted free primary education policies as a strategy to address this continuing issue and as part of their Millennium Development Goals commitments (Inoue & Oketch, 2008; Oketch & Rolleston, 2007), while NGOs and international institutions such as the World Bank have implemented or funded projects to increase school participation among girls by providing tuition and fee waivers (Rose, 2009; World Bank, 2016).

Inability to afford school materials (Barrier 17)

In a similar vein as tuition and fees, girls in the poorest households may not be able to attend school due to the costs of buying exercise books, clothes for school, and writing implements (Kadzamira & Rose, 2003). The cumulative cost of school materials can reinforce gender inequalities in school attendance and completion (Chapman & Mushlin, 2008). This may be an issue even when tuition and fees are waived, as households are typically expected to bear the entire cost of purchasing these supplementary materials for school, and choice overload may disincentivize adults from purchasing each additional tool for their children (Blocker et al., 2013; Jackson, 2011; Williams et al., 2015). Parents and guardians may allocate more school materials such as books to boys than to girls, which reflects and reinforces perceptions of lower importance in education for girls (Chimombo et al., 2000). In addition, the prices for uniforms for girls may be higher in communities where cultural concerns about safety and privacy translate into particular styles of dress (Herz et al., 1991). Indeed, the cost of school materials is recognized as a barrier to schooling by governments such as India and Zambia, which have implemented policies and programs that make uniforms noncompulsory, or distribute materials such as book bags and uniforms for free to individuals and households (Department of Public Instruction Bengaluru, 2020; International Monetary Fund, 2005). Indeed, interventions that provide material incentives for schooling are among those that have been noted in other reviews to have the potential to improve education outcomes for children (Snilstveit et al., 2015).

Lack of adequate food (Barrier 18)

School feeding programs in the form of in-school lunches and take-home rations can be considered yet another way to reduce the cost of education and, at the same time, incentivize parents to send girls to school. These types of interventions also have the added potential benefit of reducing malnutrition among school-aged children, which is commonplace in low- and middle-income settings—30.3% of children in Africa and 23.2% of children in Asia are stunted (Global Nutrition Report, 2018). Although some studies disaggregate findings by gender, few have assessed the causal relationship between malnutrition and schooling outcomes such as academic performance focusing specifically on girls, despite the fact that girls in resource-constrained settings are less likely to get access to food and thus might benefit disproportionately from school feeding programs (World Food Programme, 2020). However, we might infer how malnutrition may affect girls through observational studies on anaemia and iron deficiency, which are known to be detrimental to cognitive health and are a common occurrences among girls in LMICs (Bahrami et al., 2019; Balarajan et al., 2011). The immediate reductions in school performance due to reduced cognitive capacity may have negative downstream effects in terms of educational and labour market outcomes, putting adolescent girls at a further disadvantage in the long run (Akramipour et al., 2008; Halterman et al., 2001; More et al., 2013). This may in turn have detrimental intergenerational impacts as well—experiences of childhood anaemia by parents has been associated with increased odds of anaemia in their children, which may further exacerbate already existing gaps in education (Onyeneho et al., 2018).

Types of studies

We limit our review to studies that use sufficiently rigorous methods to provide valid causal estimates of the effects of interventions on the outcomes of interest (i.e., that attempt to control for endogeneity

Studies reporting on both RCTs and quasi-experiments were eligible for inclusion. Articles that utilize study designs outside of those listed above were excluded, notably:

RCTs that only report post-intervention data from a small-scale intervention

Types of participants

We included studies that report education outcomes for girls, and/or interact results by sex. Studies that only reported results for boys, or combined girls and boys without reporting an interaction by sex, were excluded.

Types of interventions

Below is the list of gender-related barriers identified from the literature, or subsequently added based on the findings in our review. Descriptions of the types of interventions that are hypothesized to address these barriers are listed in the background section. Studies flagged for inclusion in the search present the results of interventions/exposures that attempted to remove or minimize one or more of the following barriers—otherwise they were excluded: 1.

Lack of support for girls' education

Although a common intervention approach, cash transfer programs targeting individuals and households were not included in this review because there have been several recent systematic reviews that have analyzed these interventions (e.g., Baird et al., 2013; Snilstveit et al., 2015).

Multi-component interventions in which one of the components is addressing gender-related barriers were included. Interventions and policies may take place at the primary or secondary levels or may be nonformal for school-aged young people. We identified a handful of multi-component programs that fit into one or more of the above-listed barriers, but also included a cash transfer component at the individual or household level. These studies have been flagged in Table 1 due to the inability to separate out the effects of the cash transfers from the effects of other components of the interventions.

Types of outcome measures

Only studies that measured the primary outcomes listed below were included. Outcomes of interest had to be measured consistently in all comparison groups in order for the study to be included.

Duration of follow-up

We did not exclude studies based on duration of follow-up but note duration in the data extraction table.

Types of settings

Studies that reported on the primary outcomes listed above using data from LMICs at the time of the intervention/exposure, as defined by the World Bank (2018), were included. Studies that only reported on outcomes from high-income countries, as defined by the World Bank, were excluded.

Electronic searches

The following databases were searched electronically:

Grey literature was identified using the databases listed above from DEC, ELDIS, OpenGrey, and DFID R4D. Studies published in all languages were included. Additional unpublished/ongoing studies were identified through searches of websites of specific organizations identified in the search to be key resources. These organization websites include: the Center for Global Development, CARE, CEDPA, High-Quality Technical Assistance for Results (HEART), International Center for Research on Women (ICRW), J-PAL (Poverty Action Lab), the Population Council, UNESCO, UNGEI, and UNICEF. References from these websites were reported in a general category (World Wide Web, i.e., WWW). The search strategy documented in Supporting Information Appendix Section A was used to conduct searches through ERIC and was adapted to conform to the search functions of the other databases. Searches were conducted in March and April of 2019.

Searching other resources

Reference lists and bibliographies in relevant review articles and reports of systematic reviews found in the search were also checked to identify additional articles eligible for inclusion. Reviewers contacted the authors of included studies as well as other relevant researchers and organizations to locate additional articles eligible for inclusion.

We concluded extra searches through databases, websites, reference lists, and so forth, in March of 2020.

Selection of studies

Articles were identified through searching the databases listed in the search strategy shown in Supporting Information Appendix Section A. Depending on the structure of each database, we used the inclusion criteria to further filter results when possible, excluding studies that were out of geographic scope, topical scope (e.g., included only animals), or describing girls' education without examining an intervention, as well as removing duplicates. Each title and abstract was screened by a set of two reviewers based on the inclusion criteria documented above through Covidence. Due to the large number of articles identified in the search, we employed a team of Masters and PhD-level reviewers in related disciplines to assist with selection of studies (Fiona Gambanga, Nura Anwar, Rachel Passmore, Lili Warren, Isabel Odean, Anne Smith, Grace Sheehy, Aditi Patrikar, Jeanette Shekelle). Before any review activities, all reviewers were trained on a random set of 10 articles to gain familiarity with voting procedures as well as address any gaps in understanding about the inclusion criteria. Randomized selection of articles was performed in Stata 16.

Group 1 (F. G., N. A., R. P., L. W., I. O.) were tasked with assessing which titles and abstracts to pull full text for. Members of Group 1 were split into pairs and the titles and abstracts were divided into equal parts for each pair to review. Once the final set of abstracts was agreed upon, full text was linked to each article. Ahead of the full text review, a sample of 10 randomly selected articles was jointly reviewed by authors (S. P., B. M., E. K. C.) and any disagreements were discussed to establish that interpretation of the content of each article was consistent. Subsequently, Group 2 (A. S., G. S., A. P., and J. S.) were similarly split into pairs and allocated portions of the full text articles to determine our final list of included studies. Where possible, online appendices, addendums and errata were identified and reviewed to determine inclusion. At both stages of screening, a third team reviewer (E. K. C.) settled disagreements regarding study inclusion.

Data extraction and management

Data extraction, along with intervention grouping, was completed by three reviewers (A. S., G. S., and E. K.C.) through an online tool (Google Forms). The complete tool can be found in Appendix Sections D–F. The data extraction form was designed in consultation with the Cochrane Handbook (Higgins & Green, 2011) and adapted from Psaki et al. (2019) and Mensch et al. (2019). Three reviewers (A. S., G. S., E. K. C.) divided the included papers into three groups, each taking one group for full data extraction. After full data extraction was completed for each paper by one reviewer, they rotated groups, and a second reviewer (A. S., G. S., E. K. C.) checked each paper's extracted data for errors. When errors arose, the two reviewers handling each paper (i.e., the reviewer who extracted the data and the reviewer who checked it) met to agree on how to resolve the error.

Assessment of risk of bias in included studies

While RCTs are generally considered the gold standard for identifying causal impact, even they may be subject to threats to validity (e.g., differential loss to follow-up and selective reporting of outcomes). Therefore, we performed an assessment of risk of bias in both experimental and quasi-experimental studies adapted from RoB 2 (Higgins et al., 2016) for randomized studies and ROBINS-I (Sterne et al., 2016) for non-randomized studies. In addition to using the full ROBINS-I tool for non-randomized studies, we also added questions on methods-specific criteria from Psaki et al. (2019), which were adapted from Baird et al. (2013). Risk of bias was assessed by two reviewers (A. S. and G. S.). A third reviewer (E. K. C.) assisted in resolving any disputes. The full assessment of risk of bias tool is presented in Supporting Information Appendix Section B. The relevant section of the risk of bias assessment tool was applied to each included study at the time of data extraction.

Measures of treatment effect

We extracted information from quantitative models that reported on our primary outcomes of interest. This included measures such as unstandardized regression coefficients, beta coefficients, odds ratios, and t statistics, among others. Typically, most systematic reviews that contrast groups compute standardized mean differences if the outcome is continuous and odds ratios if the outcome is dichotomous (Lipsey & Wilson, 2001). However, given the complex estimation methods used in included studies, we instead chose to convert effect sizes to partial correlation coefficients. In some cases, for example, linear models with a continuous or dichotomous education exposure, conversion was straightforward. But other conversions were more complex. See Supporting Information Appendix Section D for a more detailed description of the computation of partial correlations.

Partial correlations represent the relationship between two variables, controlling for (i.e., “partialling out”) covariates, and range in value from −1 to 1 (Aloe, 2014; Aloe & Thompson, 2013). Due to the similarities in the estimating equations underlying partial correlations and standardized regression coefficients (β), we can interpret the size of the effect as the standard deviation change in the dependent variable. Drawing on Cohen's (1988) conventions for interpreting effect sizes, correlation effect sizes ≤0.10 are considered small, values of 0.25 are considered medium, and values ≥0.40 are considered large. In the case of this study, which reports partial correlation coefficients that are adjusted for a set of key covariates, we might expect the magnitude to be somewhat smaller than the guidelines for bivariate correlations. Corresponding 90% confidence intervals (CIs) were calculated for all outcomes as many authors, given the conventions in the economics literature, use α = .10.

Criteria for determination of independent findings

In cases where a singular study provided results on more than one of our outcomes of interest, we presented each result separately. Similarly, if a singular study presented several measures for the same outcome, all results are presented for completeness, but are grouped by comparability of the effect size to other effect sizes, as categorized above. In addition, where multiple results for the same outcome of interest measured in the same way were presented within the same paper but were the result of different models/subgroup analyses, one effect size was reported based on the authors' indication that the effect was one of the primary results of the study. If multiple articles were identified that reported on the results from the same intervention, drew from the same study population and reported on the same or very similar outcomes, we reported the effects from the earliest published article or the longest term follow-up (Willson, 2018). In cases where the publication date of the earliest published article and the longest-term follow-up were in conflict, we included the longest-term follow-up among our reported results.

Unit of analysis issues

We followed the Cochrane Handbook recommendations for studies that do not correct for one or more types of unit of analysis errors. The following is a summary of the strategies we employed to treat the results subject to various manifestations of unit of analysis errors.

Repeated observations on participants o

If data from multiple follow-up rounds were reported, we chose the follow-up results that were furthest in time from the time of implementation.

Studies with multiple intervention groups are noted in Table 1, along with notes if an arm was excluded and the rationale. Only two studies (Bagby et al., 2017; Kazianga et al., 2012; Kazianga et al., 2019) reported on repeated observations on participants. In both cases, we reported the longest-term follow-up available for any primary outcomes. Other strategies that we outlined in our protocol are not listed as we did not identify studies with those unit of analysis issues. The full list of strategies can be found in the protocol (Chuang et al., 2019).

Assessment of reporting biases

We did not find a sufficient number of studies to run quantitative analyses on reporting bias (e.g., funnel plots, trim-and-fill plots, tests for meta bias) for any given barrier and outcome based on our criteria for meta-analyses. Further details about the criteria to conduct meta-analyses can be found below.

Data synthesis

Subgroup analysis and investigation of heterogeneity

As we did not identify enough studies to run meta-analyses, we in turn could not conduct moderator analyses using meta-regressions. We attempted to narratively describe the potential sources of variability of results based on various characteristics of the studies, such as whether interventions were single or multicomponent, the types of components employed, and methodology (e.g., experimental vs. quasi-experimental).

Dealing with missing data

Where data were missing that could determine the inclusion eligibility of a study, such as interventions that may fall in the “Other” category, or effect size conversion, reviewers contacted the study authors to request the relevant information; three attempts to contact the authors were made within 1 month. If the authors did not respond or did not provide the relevant information within 1 month of the first date of contact, then the study would have been excluded from quantitative synthesis but included in narrative synthesis. Fortunately, all authors of included studies responded to requests for information.

Results of the search

The PRISMA flow diagram (Figure 3) indicates the number of citations initially identified (N = 25,935) for the full systematic review though our database search, detailed in the Methodology section. There were a number of studies (n = 80) that were later suggested by our advisory group, authors of included papers, or through literature review searches. Of the articles initially identified either through our search or through the independent identification of additional papers, 18,780 were screened for relevance based on titles and abstracts, and 857 went through full text review. Ultimately, results from 82 studies were included in the review from a total of 88 papers (a few studies presented results in more than one paper).

OPEN IN VIEWER

Figure 3

PRISMA diagram

Included studies

Of the 82 included studies (see Table 1), 41 employed an experimental design, and 41 utilized quasi-experimental methods. For one study (Benshaul-Tolonen et al., 2019; Phillips-Howard et al., 2016) we draw on the findings of one journal article and one working paper. Among the rest of the studies, about half of the publications were working papers (n = 42), 40% were journal articles (n = 35), 7% were reports (n = 6) and 3% were dissertations (n = 3). Almost all (80/82) studies provided girl-specific effects, 29% (n = 24) reported gender-differential effects of interventions and exposures, with 28% (23/82) reporting both girl-specific and gender-differential effects for girls relative to boys. Just over half of the studies were conducted in Sub-Saharan Africa (n = 43), followed by South Asia (n = 24). A total of 15 countries in Sub-Saharan Africa were represented, of which the most frequently represented were Kenya (n = 9), Uganda (n = 7), Ethiopia (n = 4), Zambia (n = 4), Zimbabwe (n = 4), and Ghana (n = 3). The majority of studies in South Asia were based in India (14), followed by Bangladesh (6), Pakistan (3) and Nepal (1). The rest of the studies were conducted in Latin America and the Caribbean (n = 5), Central Asia (n = 1), South East Asia (n = 3), East Asia (n = 4), and the Middle East and North Africa regions (n = 2). All East Asia studies were set in China. The earliest included study was from 2004, with the most recent study was published in 2020.

One characteristic of included studies that we expected to find challenging analytically is the issue of multicomponent programs. Indeed, many studies evaluate programs that include multiple components (see descriptions in Table 1). For example, an initiative may include school construction as well as teacher training. Thus, many of the studies are included under more than one barrier (see Tables  5.1 through 5.18). We explore the issues with synthesizing the results of multicomponent interventions in Sections 5 and 6.

Excluded studies

During the initial database search, 6654 studies were excluded because they were out of scope based on our inclusion criteria, and 254 studies were removed as duplicates. A further 315 duplicate studies were removed when the results of the database search were imported into Covidence. At the Title and Abstract Screening phase, we excluded 17,923 studies due to lack of relevance to the review, based on our inclusion criteria. Seven hundred sixty nine studies were excluded at the Full Text Review stage, of which the majority did not address a barrier to girls' education (n = 297) or did not use a quantitative method that adequately controlled for endogeneity (n = 217). 80 studies, while gender-informed, were purely qualitative, 64 studies did not include an exposure or outcome that was outlined in the protocol, 33 studies did not disaggregate the effects of the programs or policies by sex, 15 studies were conducted in high-income countries and 10 were purely cash transfer studies. We found and excluded 32 further duplicates, one study published before 2000 and three ongoing studies. We were unable to locate the full text for 16 studies.

Experimental studies

Of the 41 experimental studies, all but five were judged to have low risk of bias (Table 3). Four studies (Aber et al., 2017, Austrian et al., 2020, Cho et al., 2019, Freeman et al., 2012; Garn et al., 2013) were determined to have some concerns, and one (Wilson et al., 2012) was judged to have high risk of bias based on the tool we utilized. The most common potential sources of bias among these included studies were from issues with randomization, deviations in assignment and selection of the reported result. The main cause for concern related to randomization and assignment was the lack of blinding and concealment of intervention assignment among the participants and implementors throughout the programs. This is typical of the field at large, as any non-laboratory-based experimental social science studies are not blinded at either the participant or investigator level, if they report on blinding at all (Deaton & Cartwright, 2018). As for bias due to selection of reported results, the primary reason for the relatively higher numbers of papers with “some concerns” with bias is due to the lack of prespecified plans for analysis of the trial(s). This is again typical of social science studies—the publication of pre-analysis plans or study protocols before trial implementation, even for RCTs, is uncommon, though there has been a recent push among researchers to change this norm (Asendorpf et al., 2016; Lupia & Elman, 2014; Miguel et al., 2014).

Quasi-experimental studies

Among the 41 quasi-experimental studies included in the review, the overall risk of bias results were more mixed than those of the experimental studies (Table 4). Twenty two (22) were determined to have low risk of bias, while 16 had some concerns and three had high risk of bias based on our criteria. The most common sources of potential bias were due to confounding, missing data and methods-specific criteria. It is noteworthy that the majority of studies had either some concerns or high risk of bias due to missing data (n = 35 and n = 2, respectively). This is not altogether surprising given that most quasi-experiments tended to draw on secondary data, so most of our included studies did not report on how observations missing either from data collection or within their variable(s) of interest may have affected their analyses. Focusing on bias due to confounding, 27% (n = 11) were found to have some concerns and 5% (n = 2) had high risk of bias. This was primarily due to the phrasing of the items in ROBINS-I tool. For example, the item “Is there potential for confounding of the effect of the intervention in this study?” was far too open-ended for our team to confidently answer “No” or “Probably No.” Because the ROBINS-I tool was originally geared towards nonrandomized smaller-scale interventions, the phrasing of the items was perhaps not optimized for secondary analysis of non-randomized exposures such as policy changes. Finally, 10 studies were determined to have “some concerns” with bias due to methods-specific criteria, with two studies having a “high” risk of bias in this area. These studies used a variety of methods including difference-in-differences, two stage least squares, and regression discontinuity. Concerns included lack of Hausman test and insufficient information on specifications for propensity score matching.