Adolescent Girls’ Safety In and Out of School: Evidence on Physical and Sexual Violence from Across Sub-Saharan Africa

Abstract This study characterises rates of physical and sexual violence against adolescent girls and compares rates of violence against girls who are enrolled versus unenrolled in school, to contribute to an understanding of the relative risks associated with school attendance. We look at rates of violence across countries that together represent 80 per cent of sub-Saharan Africa’s girls aged 15–19. The analysis shows high rates of violence overall: 28.8 per cent of girls report having experienced physical or sexual violence. However, in none of the 20 countries do adolescent girls enrolled in school report a statistically significantly higher likelihood of having been sexually assaulted than girls not enrolled in schools. Another source of data sees significantly higher rates in just one country. This pattern of results is robust to the inclusion of a range of control variables, and to analysis using different sub-groups. The evidence on physical violence is more mixed. Girls face significant rates of physical and sexual violence whether they are enrolled in school or not. These findings underline the importance of confronting violence against girls both in school and in the community, with tailored programs appropriate to each setting.


Introduction
The benefits of educating girls both to girls themselvesthrough higher earnings (Psacharopoulos & Patrinos, 2018), delayed fertility (Duflo, Dupas, & Kremer, 2021), and reduced adolescent marriages (Boahen & Yamauchi, 2018;Masuda & Yamauchi, 2020)and to the people around themwith better outcomes for their children (Akresh, Halim, & Kleemans, 2021) are clear, compelling, and persistent. While much attention around girls' education has focused on improving access, boosting learning (Evans & Yuan, 2022), and preparing girls for a school-to-work transition (Rose, 2021), recent studies in individual countries have documented sexual abuse of girls in school as an important issue (Chitsamatanga & have been dramatically expanding access to secondary education (Evans & Mendez Acosta, 2021), with more planned in the coming years, so understanding the dynamics of violence both in and out of school is crucial.
Across these countries, 28.8 per cent of girls in the 15-19 age group report having experienced physical or sexual violence previously. Once one adjusts for basic differences between girls who are enrolled in school and those who are not enrolled in school (i.e. rurality, age, parental education, and household assets), enrolled girls are slightly less likely (1.1 percentage points) to report having experienced violence. That difference is driven by differences in reported sexual violence. Country-by-country, girls enrolled in school are only more significantly likely than unenrolled girls to report ever having experienced physical or sexual violence in one out of the 20 countries (Nigeriaand that result is driven by physical violence). In both cases, the difference is driven entirely by differences in reporting physical violence, not sexual violence. This pattern of results is consistentwith almost no exceptionsif we restrict the analysis to violence that has occurred in the last 12 months, if we focus on the accumulated risk of violence measured against total years of schooling, if we focus on slightly older girls, or if we restrict the sample only to girls who have never been married.
We next examine the VACS, which report the proportion of girls who experienced sexual violence in the past 12 months in six countries in sub-Saharan Africa. 3 On average across the six countries, 17 per cent of girls report at least one incident of sexual violence in the past year. This number is unsurprisingly lower than the number in the DHS, since it is restricted to sexual violence and to incidents in the past year. Among those who report having experienced sexual violence, on average 10 per cent of incidents took place at school, ranging from 2 per cent in Zimbabwe to 18 per cent in Kenya. In the VACS, girls in school are statistically significantly more likely to experience sexual violence in just one country (Nigeria), but they are significantly more likely to experience physical violence in most countries in our sample: most of that is from non-partners (and so may be driven by corporal punishment in schools). We discuss limitations to our approachincluding the fact that these numbers are likely underestimates overall, given hesitation to report violence in surveys, and potentially omitted factors that affect both education and the risk of violencein our methods section.
These results demonstrate that violence against adolescent girls is a pressing problem both in school and in the community. In the discussion section of the paper, we lay out potential courses of action to reduce violence in both contexts. While this study focuses on low-and middle-income countries, especially in sub-Saharan Africa, this problem is not unique to those contexts. Previous analysis demonstrates, for example, that overall rates of sexual violence against minors are uncorrelated with the income level of the country (Crawfurd & Hares, 2020).

Analysis of the Demographic and Health Surveys (DHS)
We used the most recent Demographic and Health Surveys of the 20 most populous countries in sub-Saharan Africa with the relevant data: Angola ( Figure A1, available in the Supplementary Materials). The DHS are nationally representative household surveys that cover a wide range of development indicators, including a section on the prevalence of domestic violence from individual surveys conducted with women aged 15-49 years old. For this study, we looked at the incidence of having experienced physical or sexual violence perpetrated by family members, friends, and figures of authority for the sample of girls aged 15-19 years old, separately by current attendance in school. Girls who are Adolescent girls' safety in and out of school 741 currently married or living with a partner, and girls who are widowed, divorced, separated or have lived with a partner before, answer additional questions on physical and sexual violence specifically perpetrated by current or former partner/husband. We report both the rates of ever having experienced violence before and rates of having experienced violence in the past 12 months across the sample.
We use simple linear regressions to estimate if the difference in the experience of violence between girls who attended school in the past year and those who did not are statistically significant. We report both the simple difference between the two groups and the difference when adjusted for girls' age as well as other demographic characteristics such as type of residence (urban/rural), parents' education, and a wealth index from household assets. We control for these particular variables because they may influence the exposure to violence beyond the effect of school attendanceas in Palermo et al. (2019) and because they are unlikely to be affected by an adolescent's experience of violence. While an adult's income may be affected by an earlier experience of violence in adolescence, violence is less likely to affect household wealth significantly while the child is still an adolescent. We also regress the probability of having ever experienced physical violence, sexual violence, or either on total years of schooling. (For more details on the survey years and the variables used from the DHS, see Appendix Section A.) Country-level tabulations and regressions use sample weights and the pooled tabulations and regressions use denormalised weights (Ren, 2013).
While the focus of this paper is on adolescent girls, we do examine the pattern for 20 -24year-old girls, both to ensure that we do not observe a major shift in pattern and because a number of 20 -24-year-old girls may still be in basic education. We compare the rate of physical or sexual violence experienced in the last 12 months by girls who attended school the previous year to the experience of girls who did not attend school (with and without the controls discussed above). We also ran the same comparison for older girls who are still in secondary school compared to unenrolled girls who have not completed secondary school.

Analysis of the Violence against Children and Youth Surveys (VACS)
We also explored the Violence Against Children and Youth Surveys, available through the Together for Girls partnership for countries in sub-Saharan Africa with surveys in a similar time period: Kenya (2010), Malawi (2013), Nigeria (2014), Tanzania (2009), Zambia (2014), and Zimbabwe (2017) (Together for Girls, 2017)) (see Appendix Figure A1). VACS are nationally representative household surveys that interview male and female participants ages 13-24 years old to identify prevalence and context of childhood violence including physical, emotional and sexual violence. The surveys also identify where the most recent violence has taken place (e.g. at home, in school) and if this violence happened in the past 12 months. (For more details on the survey years and the variables used, see Appendix Section B.) We use the subsample of girls aged 15-19 years old to enhance comparability with our DHS analysis. Sample weights are used for all country-level and pooled tabulations and regressions. We also report the difference in the experience of violence by attendance to school, both the simple difference and the difference controlling for age and household assets such as access to electricity, telephone, television, radio, and bicycle (in lieu of the wealth index variable used as control in the DHS). We use the controls that are common across surveys to report our main results, but for robustness, we added additional controls where availablesuch as distance to nearest clinic and highest schooling attained by any adult in the household (in Kenya) and highest schooling attained by head of household (in Tanzania). Differences in coefficients between the two specifications (common controls and the additional controls) are reported in the notes section of the result tables. Similar to the controls used for DHS, we selected these variables because they may influence the exposure to violence independent of school attendance and at the same time these variables are not likely to be directly affected by prior adolescent exposure to violence. While these two datasets have different country coverage and different measures of violence, we include both because together they provide a richer picture of violence against adolescent girls in and out of school. We discuss differences in findings across the two sets of surveys in the discussion (Section 3.3).

Limitations of this analytical approach
Our analysis faces several limitations. First, although the World Health Organisation defines adolescence from 10 years of age to 19 years, the DHS only interview girls 15 years old and older, which restricts the adolescent age range of our sample to ages 15-19 years. We restrict our analysis using VACS within this age range to enhance comparability across the two datasets.
Second, while the DHS have the advantage of being available for a wide range of countries, they use a relatively narrow definition of sexual violence, they do not identify where the violence took place, and they do not identify whether the girl was enrolled at the time that the violence took place. Similarly, the DHS collect data on violence under the domestic violence module and the questions are largely (but not entirely) focused on partner-perpetrated violence. On the other hand, the VACS are available for only a limited number of countries, but they collect information on different types of sexual abuse and harassment (e.g. unwanted touch and engaging in sexual activities because of pressure), whether this incident happened at school, and they explicitly collect information from a wider set of potential perpetrators (family members, peers, and community members).
Third, both sets of surveys are limited by the challenges inherent in collecting data about violence. Different interpretations of what constitutes violence and different social norms on what violence is 'acceptable,' as well as respondents' fears about confidentiality, make it challenging to collect accurate information from respondents about violence (Devries, Naker, Monteathvan Dok, Milligan, & Shirley, 2016;Tanton, Pearlman, & Devries, 2021). With both surveys, respondents may be more likely to underreport sexual or physical violence (e.g. from embarrassment or fear) than to overreport, so that even these striking estimates are likely underestimates of the problem. 4 To our knowledge, there are no conclusive estimates of whether enrolled or unenrolled girls in sub-Saharan Africa are more likely to report violence, or whether girls are more likely to report violence at school versus violence in other contexts. Two points merit consideration on this topic. First, previous analysis of the VACS in several countries suggest that children enrolled in school were more likely to informally disclose or to know where to seek help in the face of violence in most countries (Pereira et al., 2020). An extrapolation from that would suggest that our reports from unenrolled girls may be more likely to be underreports. Second, both the DHS and the VACS are household surveys, so girls may be less likely to report violence by household members relative to violence that takes place at school. Both of these points would signal that, if anything, reports of violence outside of school may be more underreported than those of violence in school.
A fourth limitation of our analysis is that all the data we analyse were collected pre-COVID. While we have some evidence that at-home violence against children rose during COVIDrelated school closures (see footnote 2), large-scale surveys from multiple countries have yet to reveal whether levels of violence in and out of school are distinctive in the wake of the global pandemic.
Finally, neither data set can be used to infer causal estimates of the impact of school enrollment on violence, for multiple reasons. First, the experience of violence could and likely does affect school enrollment and participation, just as school enrollment could affect violence. Some of the evidence on this suggests that the relationships are not always straightforward (Psaki, Mensch, & Soler-Hampejsek, 2017). Second, other factors could drive both school enrollment and the experience of violence. For example, youth from higher-income households Adolescent girls' safety in and out of school 743 may have both higher enrollment rates and more resources to protect them from violence (Gentz, Zeng, & Ruiz-Casares, 2021). In that case, a negative association between school enrollment and violence would not reflect the impact of enrollment on violence but rather the impact of wealth on both enrollment and violence. In the absence of the combination of an experiment increasing school participation and historical data on violence in and out of schools, confidently inferring a causal impact is difficult. Our controls for type of residence (urban/rural), parents' education, and a wealth index from household assets all help, but they cannot entirely resolve these issues simply because some differences that could drive both enrollment and violence may not be measured in cross-country household data sets.
In supplementary analysis, we examine the association (with controls) only for young women who have never been married or lived with a partner, since those who marry young may be different in many ways than those who do not. The pattern of results is similar. The reason we do not remove married young women in our primary analysis is that school enrollment can delay adolescent marriage (Duflo et al., 2021;Psaki et al., 2021), so by removing young women who have married from the sample, one is potentially removing one of the protective benefits of education. This is particularly true if girls in adolescent marriages are more likely to experience violence.

Demographic and Health Surveys (DHS) analysis
Across the 20 sub-Saharan African countries, 28.8 per cent of girls aged 15-19 report having experienced physical or sexual violence ever (Table 1), ranging from 14.0 per cent in Ethiopia to 44.6 per cent in Uganda. 5 Once we adjust the averages for country, age, parental education, socioeconomic status, and urban/rural status, enrolled girls are 1.0 percentage points less likely to report having experienced violence. Likewise, after adjusting for covariates, only five in 20 countries report statistically significant differences between enrolled and unenrolled girls. In Chad, Rwanda, Tanzania, and Zimbabwe, enrolled girls are less likely to report having been victims of violence; in Nigeria, enrolled girls are more likely to report having been victims of violence. This is driven entirely by reports of physical violence, not sexual violence.
If we break down physical violence and sexual violence, we observe that 26.0 per cent of girls report having experienced physical violence (supplemental Appendix Table A1) and 7.6 per cent report having experienced sexual violence (supplemental Appendix Table A2). Adjusting for covariates, we observe no significant difference in the likelihood of experiencing physical violence between enrolled and unenrolled girls; enrolled girls are 2.8 percentage points less likely to report sexual violence than unenrolled girls (with 99 per cent statistical significance).
If we restrict the sample to only those girls who have never married or lived with a partner, we observe no significant difference in the likelihood of experiencing either physical violence or sexual violenceon averagebetween enrolled and unenrolled girls (supplemental Appendix  Table A3 and A4). There are a couple of country exceptions: in Angola and Mozambique, once one restricts the sample to never married girls, the likelihood of sexual violence is significantly higher among those attending school. In Angola, the likelihood of physical violence is significantly higher among those attending school. As discussed in the methods section, this specification makes less sense if education plays a protective role against early marriage. Both Angola and Mozambique have high rates of child marriage (52 per cent and 57 per cent, respectively), albeit not the highest in our sample: Burkina Faso, Chad, and Nigeria all have higher rates (Yaya, Odusina, & Bishwajit, 2019). These results suggest that conditional on remaining unmarried, girls in school are at no higher risk of violence on average, but with exceptions.
To counter the concern that the violence may have taken place long ago, we also examine sexual or physical violence reported to have taken place in the last 12 months. 16.2 per cent of girls overall report experiencing physical or sexual violence in the past year, with no statistically significant difference between enrolled girls and unenrolled girls (supplemental Appendix Table  A5). In only one country (Ghana) do enrolled girls report statistically significantly higher rates of violence in the last twelve months. This is driven entirely by reports of physical violence. On average, across countries, 14.5 per cent of girls report having experienced physical violence in the past year, with no significant difference between enrolled and unenrolled girls (supplemental Appendix Table A6), and 3.9 per cent report having experienced recent sexual violence, with enrolled girls 2.5 percentage points less likely to report sexual violence than unenrolled girls (with 99 per cent statistical significance) (supplemental Appendix Table A7). We also find that girls with more years of schooling are slightly more likely to have experienced physical violence and slightly less likely to have experienced sexual violence in the last 12 months (supplemental Appendix Table A8).
For girls who have experienced sexual violence, we report the distribution of perpetrators of the first incident of sexual violence (Table 2). Overall, the most common perpetrators are boyfriends (19.5 per cent), friends (14.4 per cent), strangers (13.7 per cent), and family members (8.5 per cent). Teachers are much less likely to be named, at 2.1 per cent. We also observe that almost half of girls who did not attend school in the past year did not identify the perpetrator of the first incident of sexual violence, compared to only 5.9 per cent of girls who are still in school. If girls who are not enrolled in school are more likely to have experienced violence at the hands of someone in the household, then the fact that the survey is administered in the home may drive this differential reporting. We also examined the experience of violence of older girls (20-24-year-old) to see if there are changes to the experience of violence immediately succeeding adolescence. We observe the same pattern that on average, 20-24-year-old girls enrolled in school are less likely to be exposed to physical or sexual violence than girls who are not enrolled in school. This observation applies whether we simply compare enrolled to unenrolled girls with controls (supplemental Appendix  Table A9) or if we restrict to enrolled girls who are still in secondary school compared to unenrolled girls who have not completed secondary school (supplemental Appendix Table A10).

Violence against Children and Youth Surveys (VACS) analysis
The VACS data from the six sub-Saharan countries show generally high rates of physical violence in the last 12 months (both for girls in and out of school), ranging from 11.4 per cent in Zimbabwe to 41.6 per cent in Kenya (supplemental Appendix Table A11). Across all six countries, girls attending school are more likely to report experiencing recent physical violence than girls not currently attending school, once one adjusts for controls. Four of these differences are statistically significant. The highest difference is in Kenya, where girls enrolled in school are 16.4per cent more likely to report experiencing recent physical violence than girls not in school (statistically significant at 95 per cent). This figure is similarly high in Nigeria and Tanzania where the differences are 12.6 per cent and 13.8 per cent (both statistically significant at 95%). In Zimbabwe where there is the lowest rate of physical violence across the countries in the sample, girls in school are 2.4 per cent more likely to experience physical violence than girls not in school.
When separating the recent experience of physical violence by partner, girls in school are less likely to experience physical violence from a partner (supplemental Appendix Table A12), whereas they are more likely to experience physical violence from non-partners (supplemental Appendix Table A13). Detailed analysis of perpetrators in the VACS show that this physical violence by non-partners among enrolled girls is driven principally by teachers andto a much lesser degreeby peers (supplemental Appendix Table A14). 6 As discussed earlier, the DHS have a relatively restrictive definition of sexual violence and do not identify the location of the violence. The VACS data show much higher rates of sexual violence overall, likely because it includes a wider range of behaviours: unwanted touch, attempted unwanted sex, and pressured sex, as well as physically forced sex (Table 3). Across the six countries, 16.7 per cent of 15-19-year-old girls report at least one incident of sexual violence in the past year. The number ranges from 5.9 per cent in Zimbabwe to 27.7 per cent in Malawi. Across countries, unwanted touch and attempted unwanted sex are the two most common types of sexual violence, although which is most commonly reported varies across countries. Fewer girls report physically forced sex (with the max being 3.7 per cent in Nigeria) or pressured sex (with the max being 2.2 in Malawi).
Comparing girls who are enrolled in school and those who are not, the differencesonce adjusted for covariatesare insignificant in four of the six countries (Table 4). Only in Nigeria do enrolled girls report a statistically significantly higher likelihood of experiencing sexual violence; and only in Zimbabwe do they report a significantly lower likelihood. Of those who experienced sexual violence, 10.2 per cent overall reported having experienced that violence at school (Table 3). For girls who are currently enrolled in school, that number risesunsurprisinglyto 14.4 per cent (Table 4), rising as high as 28.3 per cent in Kenya.

Comparing results across the DHS and the VACS
The two sources of data (the DHS and VACS) have some overlap in covered countries, and while both sources are nationally representative, they do have differences. First, the surveys were collected during different time periodsof the six countries common to both surveys, four of them have surveys that are collected at least four years apart (e.g. Zambia's VACS was collected in 2014 against the more recent DHS in 2019).
Second, the scope of the definition of violence is different. While DHS mostly asks about physically forced unwanted sex (with a follow-up for 'other unwanted sexual acts'), VACS collects data from a broader set that explicitly includes less extreme forms of violence such as forced touch and being pressured into sex. As a result of these other forms of sexual violence being explicitly named in the VACS, respondents may be more likely to report them.
Third, and along similar lines, the violence-related questions (both physical and sexual) in DHS are collected under the Domestic Violence module and are therefore more likely to focus on intimate partner violence (there is only one question for violence from 'person other than husband/partner'; all other questions are for violence from current and previous partners). On the other hand, VACS asks about violence by partners, peers, family members, and community members in relatively equal measure, going so far as to identify the specific family member (e.g. mother, uncle) or community member (e.g. community leader, gang member) for both first and most recent instance of violence. These distinctions are in line with the general nature of the two surveys: the DHS are designed to capture a wide array of information beyond violence and therefore do not have the granularity of the VACS data. But this also makes it possible to control for additional variables in DHS such as parental education, which is something that the VACS do not collect.
Comparing the relationship between schooling and recent physical violence in the two surveys across the six common countries, we find that girls in school experience less physical violence in DHS (all coefficients in the six countries are negative, and the relationship is statistically significant in Zambia), but more physical violence in all of the countries based on their VACS compared to girls not in school (statistically significant in Kenya, Nigeria, Tanzania, and Zimbabwe). Table 5 compares the results across the two sets of surveys. This difference may be explained by how the two surveys capture violence from different perpetrators. When we disaggregate the exposure of physical violence from partners against non-partners (supplemental Appendix Table A15), we find that for both surveys, schooling is negatively associated with physical violence from partners in all six countries (significant in all six countries in DHS and significant in three countries in VACS) but is positively associated with violence from Adolescent girls' safety in and out of school 747

Notes:
We use the most recent publicly available VACS surveys. We control for age, household assets (electricity, radio, television, telephone, and bicycle) and report the adjusted difference. For Kenya: additional controls such as distance to nearest clinic and highest schooling of adult in household did not change the sign and significance of the coefficient of the adjusted difference of 'At least one incident of sexual violence' but reduced the significance level of '% of violence that happened at school' so that the adjusted difference is not statistically significant. For Tanzania: adding a control for the highest schooling attained by the head of household did not change the sign and significance of the coefficients. Country-level tabulations and regressions use sample weights.
non-partners for all six countries (significant in four countries in DHS and in five countries in VACS). This makes intuitive sense: girls in school may be less likely to have partners (certainly less likely to be married) and so may experience less violence from partners, but they are also more likely to be exposed to corporal punishment in school. The relationship between schooling and recent sexual violence is consistently negative for all countries in the DHS (significant in four countries) and mixed in VACS (half of the countries have positive associations) ( Table 5). But limiting the analysis of the VACS to the most extreme form of sexual violenceforced sexto make it more comparable to DHS yields findings that are more consistent: in five of the six countries, girls in school experience fewer instances of extreme sexual violence than girls not in school (supplemental Table A16). The relationship is statistically significant only in Zimbabwe. This suggests that girls in school may be less at risk of more extreme forms of sexual violence (i.e. physically forced sex) but may be more at risk of milder forms of sexual violence such as unwanted touch or pressure to engage in sexual acts.

Discussion
Our analysis illustrates the magnitude of the problem of violence against girls: across the Demographic and Health surveys in our study, one in four adolescent girls report having experienced violence, and one in seven reports having experienced sexual violence in the previous year. There is also no systematic evidence showing that school enrollment increases risk of sexual violence. Across the 20 countries for which we analyse DHS in our study, there is not a single country in whichafter adjusting for covariatesadolescent girls in school are significantly more likely to have experienced sexual violence, and there are only two in which girls are significantly more likely to have experienced physical violence. Across the six countries for which we have VACS, we observe higher rates of sexual violence for just one country (Nigeria). However, we observe consistently higher rates of recent physical violence in enrolled girls across the six countries in VACS, with statistically significant differences in four of the countries.  and Tanzania. They find that compared to girls not in school, girls in school are more likely to experience physical violence in five of the six countries (with statistically significant differences in three countries -Kenya, Malawi, and Nigeria). They also find that girls in school are more likely to experience recent sexual violence in five of the six countries, but with statistically significant differences only in two countries -Malawi and Nigeria; in Haiti, girls in school are less likely to experience sexual violence than girls not in school (statistically significant). Our paper examines a slightly older group of girls (15-19 years old) and uses different controls (i.e. household assets instead of a household wealth index and not controlling for living with biological parents), and we find similar trends: for countries common to both analyses (Kenya, Malawi, Nigeria, and Tanzania), schooling is positively associated with recent physical violence in all countries (statistically significant in three).
At first glance, our findings are less consistent with Palermo et al. (2019) when it comes to sexual violence. In that study, schooling is positively associated with sexual violence in all four countries that are common in both papers and statistically significantly so in two (Malawi and Nigeria). In our main analysis, the relationship is positive only in three countries and is only statistically positive in Nigeria, and the relationship turns negative (girls in schools are less likely to experience sexual violence) in Kenya, although this is not statistically significant. 7 These differences may be driven by varying controls: when we estimate an alternative specification using the same age range and as many of the same (or similar) controls as Palermo et al. (2019) as we can, we find the same direction of relationship as Palermo et al. (2019) across all common countries (supplemental Appendix Table A17).
The data show us that girls are vulnerable to violence both in school and in other contexts, suggesting there is an urgent need for reforms to allow girls to study and to live their lives safely. The majority of the countries in our sample have some type of national policy in place to reduce violence against children, but none are fully funded (supplemental Appendix  Table A18).

Can cross-country variation in our findings be explained by country characteristics?
Many hypotheses could potentially explain why the in-school/out-of-school violence differential varies across countries. For example, norms about women being in school or working or broader norms about violence could drive differences. We test for associations between the differential for experiencing violence in versus out of school relative to a long list of covariates, including proxies for women's economic empowerment (labour force participation), schooling norms (primary enrollment, second enrollment, gender gaps in primary and secondary enrollment), and prevailing gender norms (gender inequality index, share of women subjected to physical or sexual violence, share of women who believe wife beating is justified in certain situations). These are reported in supplemental Appendix Table A19.
We find a small, negative association between the share of women who believe wife beating is justified for at least one reason and in-school/out-of-school differential on the experience of physical violence or combined physical and sexual violence: In practical terms, 20 per cent more women believing wife beating is justified is associated with school being 3.2 percentage point more protective against physical or sexual violence and 2.9 percentage point more protective against physical violence alone. This may result from the relative beliefs of school staff versus community members: in communities where most community members justify domestic violence, educated school staff may be less likely to incur violence, relative to other members of the community, to whom unenrolled girls are exposed for more of the day. When we look at associations with sexual violence, we also find a small, negative association between female labour force participation and the experience of sexual violence: in other words, 20 percentage points higher female labour force participation would be associated with school being 2.5 percentage points more protective against violence. We find a similarly sized coefficient on gross Adolescent girls' safety in and out of school 751 primary enrollment: countries with higher rates of primary enrollment overall have schools with a slightly more protective effect. As schooling for girls is increasingly normalised, professional norms against violence may also evolve.

Interventions that reduce violence
The analysis in this paper highlights that violence against adolescent girls is a major challenge both in and out of school settings. While it is beyond the scope of this paper to exhaustively review the evidence on interventions to reduce that violence, we do highlight existing strands of research and its limitations here. Only a handful of interventions that aim to reduce violence in school settings have been rigorously evaluated (e.g. with control groups and credible quasiexperimental or experimental strategies) in contexts of low-and middle-income countries Turner & Hares, 2021). Almost all of that evidence is limited to teacherperpetrated physical violence. For example, the Irie Classroom Toolbox intervention in Jamaica is a one-year training program for preschool teachers that successfully reduced teacher physical and emotional violence in classrooms (effect size À67.12 per cent) (Baker-Henningham, Bowers, Francis, Vera-Hern andez, & Walker, 2021).
Another intervention called the Good School Toolkit implemented in primary schools in Uganda over 18 months helped reduced prevalence of physical violence from school staff in the past week from 49 per cent to 31 per cent (Devries et al., 2015). A shorter intervention, Interaction Competencies with Children for Teachers, which uses a training workshop over 5.5 days implemented in secondary schools in Uganda and Tanzania reduced teacher physical and emotional violence and improved teacher attitudes towards use of violence in school (Nkuba, Hermenau, Goessmann, & Hecker, 2018;Ssenyonga, Katharin, Mattonet, Nkuba, & Hecker, 2022). Another intervention for secondary schools, this time in Pakistan, addressed physical, verbal, and psychological abuse from peers over a two-year program (Karmaliani et al., 2020). Finally, a 10-week program set in refugee camps in Tanzania had no effect on rates of physical violence from teachers towards primary and secondary students (Fabbri et al., 2021). These interventions ranged from preschool to secondary level and from shorter (1 week) to longer-term (over 18 months) interventions.
There is a larger body of evidence on interventions that aim to reduce violence against children in general but there are considerable gaps in the research available. A recently published evidence map of these interventions in low and middle-income countries finds over 150 studies on reducing violence against children, 97 of which are impact evaluations (Pundir, Saran, White, Subrahmanian, & Adona, 2020). The majority of the interventions focus on intimate partner violence, and a smaller number focus on parent, caregiver or teacher-perpetrated violence, peer violence or bullying. There is limited evidence available on longer-term economic and social outcomes and perpetration of violence as opposed to the more commonly reported outcome of experience of violence. In addition, most of the interventions are likely to focus on education and life skills training and similar siloed strategies with limited evidence around implementation and enforcement of laws, improving norms and creating a safe environment in general. Finally, many of the studies are concentrated in a few regions (e.g. a concentration of studies in South Asia and southern Africa), and only a couple out of the almost one hundred impact evaluations include cost analysis.
Finally, another review of evidence on interventions to reduce violence against women and girls covers 114 studies, with over two-thirds of those studies from low and middle-income countries. The majority of the studies are quasi-experimental or experimental evaluations (Kerr-Wilson et al., 2020). The authors find that school-based interventions to prevent dating or sexual violence and similar school-based interventions against peer violence are 'effective, when well designed and executed' together with other broader programs such as economic transfers and parenting programs. They note a wide range of interventions where there is conflicting evidence or where more evidence is needed such as self-defence interventions and interventions driven by social media or digital technology platforms.

Policy implications
We propose that countries can learn more, do more, and test more regarding violence against children and adolescents. First, countries can learn more about the scope of violence, both nationally and in sub-national regions. Many countries do not gather any systematic data on violence against children and youth, and fewer still gather data on violence in schools (Crawfurd & Hares, 2020;United Nations International Children's Emergency Fund [UNICEF], 2015). Without such data, it will be impossible to know whether the situation is improving. While surveys like the DHS, which are administered in many countries, can shine a light on high levels of violence and could be used to catalyse public opinion in favour of action, more detailed modules like those in the VACS are essential to understand more about the nature, location, and perpetrators of violence against adolescent girls, thus pointing the way to designing effective interventions. To support countries to be able to gather accurate data, more methodological work to examine the best way to allow respondents to safely disclose in surveys including in cost-effective surveys that can be implemented at large scaleis needed.
Second, countries can do more. Too many countriesboth rich and poorfail to have comprehensive policies to combat violence against adolescent girls or to enforce those policies.
There are at least some tested interventions (as discussed in section 4.3 above), both schoolbased and community-based, that have been shown to reduce violence (Kerr-Wilson et al., 2020;Pundir et al., 2020). The lack of effective recourse may in part explain why women are often unlikely to report violence through official channels: one analysis of data from 24 countries estimates that gender-based violence is underreported in health systems data or police reports by anywhere between 11 and 128 times, depending on the country (Palermo, Bleck, & Peterman, 2014). International organisations can play a supporting role (Raman et al., 2021), while being careful to avoid pushing for replication of interventions from other contexts without appropriate local adaptation (Wessells, 2021).
Third, countrieswith support from the international communitycan test more. Recent reviews of effective interventions to prevent violence against adolescents reveal major knowledge gaps (Crooks, Jaffe, Dunlop, Kerry, & Exner-Cortens, 2019;Devries & Naker, 2021). Even as countries take action, it is worth investing in evaluation to understand which policies prove most effective.
While much of the discussion by policymakers focuses on violence towards girls, boys are often vulnerable to as muchand in some contexts, moreviolence as girls (Ruto, 2009). Analysis of VACS data in four countries (Cambodia, Kenya, Tanzania, and Swaziland) suggests there is no gender difference in the overall incidence of childhood violence (Ravi & Ahluwalia, 2017). Tackling violence against boys is an urgent priority too, and better data are needed on the differential vulnerabilities and solutions to violence against boys and girls.
Just as girls benefit from school in many ways, boys also experience returns to education. Research further suggests that those who justify intimate partner violence are more likely to have experienced violence themselves (Ravi & Ahluwalia, 2017), and boys who have completed secondary education are less likely to be perpetrators of violence (Abramsky et al., 2011). These are correlations, but they suggest that keeping boys in school may be one tool to help reduce violence against women and girls.
None of this will come for free. The prevalence of violence in schools is far too great, and must be addressed as a matter of urgency by education donors and policymakers. However, eliminating violence in schools alone will not make girls safe; action at both the school and community levels is essential. It will require much stronger financial and political commitment, alongside a sustained, coordinated effort to tackle the factors that make girls unsafe.
Adolescent girls' safety in and out of school 753