The World Bank Economic Review, 37(4), 2023, 640–658 https://doi.org10.1093/wber/lhad014 Article Long-Term Effects of an Education Stipend Program Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 on Domestic Violence: Evidence from Bangladesh Raisa Sara and Sadia Priyanka Abstract Intimate partner violence (IPV) remains a widespread global phenomenon. Among various factors, a low level of education is considered a significant risk factor for experiencing IPV. This paper evaluates whether a sec- ondary school stipend program introduced in 1994 for rural girls affected the long-term prevalence of IPV in Bangladesh. The study exploits two sources of variation in the intensity of program exposure and geographic eligibility and finds that cohorts of rural women eligible for the program experienced significant declines in IPV. Evidence on mechanisms suggests that the program delayed marriage formation and changed partner quality, namely their education and employment, consistent with positive assortative matching resulting from women’s improved educational attainment. There are no significant changes in labor market outcomes, decision mak- ing within the household, or women’s attitude toward the acceptability of domestic violence. Marital matches present a plausible channel through which the program reduces the risk of IPV. JEL classification: I25, I28, J12, J16, O15, O16 Keywords: secondary education, intimate partner violence, Bangladesh, adolescence, conditional cash transfers 1. Introduction Intimate partner violence (IPV) is a major global public health concern, with one in three (35 percent) women reporting to have experienced either physical or sexual violence (WHO 2013). In Bangladesh, the incidence is higher, with 54.2 percent of women reporting experience of violence in their lifetime (Bangladesh Bureau of Statistics 2016). Although the costs of IPV are well known, there is mixed consensus on effective policies to reduce IPV, as programs designed to enhance women’s empowerment or change their resource endowment can have unintended consequences. In recent years, several adolescent empowerment programs have been designed to target girls’ human- capital accumulation and life skills (Bandiera et al. 2020; Edmonds, Feigenberg, and Leight 2021), rec- ognizing that early interventions can have significant life cycle effects. Yet there is limited evidence on whether empowering adolescent girls affects their experience of IPV (Kilburn et al. 2018; Chatterjee and Raisa Sara is an assistant professor at Sam Houston State University, Huntsville, USA; her email address is rts021@shsu.edu. Sadia Priyanka (corresponding author) is an assistant professor at Connecticut College, New London, USA; her email address is spriyanka@conncoll.edu. The authors thank the editor and two anonymous referees for their valuable comments. They also thank conference participants at the Southern Economic Association (2021), 16th Annual Conference on Economic Growth and Development–ISI Delhi, KDI School–World Bank DIME Conference (2021), CSWEP-sponsored session at the Eastern Economic Association (2022), and the Nordic Conference in Development Economics (2022) for their helpful feedback and discussions. A supplementary online appendix is available with this article at The World Bank Economic Review website. C The Author(s) 2023. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com The World Bank Economic Review 641 Poddar 2022), particularly in the long term, once these programs end. Moreover, a prior it is unclear how empowerment programs will affect IPV, as it may depend on whether there are male backlash effects. This paper addresses this gap and studies the long-term impact on IPV prevalence of a nationwide policy designed to augment girls’ human capital by reducing the cost of secondary education—the Fe- male Secondary School Stipend Program (FSSSP). Introduced in 1994 for rural girls, the FSSSP was a conditional cash transfer (CCT) that provided tuition and cash stipends that were conditioned on main- Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 taining attendance, grade requirements, and remaining unmarried. The program is often attributed to the significant advancement of girls’ education in Bangladesh. Adolescent empowerment through education can be an effective tool in reducing IPV through several channels. First, education can improve women’s labor market opportunities and their autonomy and bar- gaining status within the household (Anderson and Eswaran 2009; Majlesi 2016). However, instrumental theories of violence suggest that economic empowerment can lead to backlash if partners resort to violence because of deviation from male breadwinning norms or to extract newly acquired resources (Weitzman 2014; Erten and Keskin 2018; Svec and Andic 2018). Second, education can alter norms and attitudes around the acceptability of IPV and allow one to be better informed of laws, although it might not change the risk of IPV (Erten and Keskin 2022). Third, education can affect marriage prospects through assorta- tive matching, enabling one to marry higher-quality spouses (Lefgren and McIntyre 2006; Mansour and McKinnish 2014), who may be less likely to resort to violence. A distinct feature of the FSSSP was that eligibility was conditional on remaining unmarried, a provision that was designed to discourage early marriage, a pervasive issue in Bangladesh that is linked to a higher risk of IPV (Fakir et al. 2016; Yount et al. 2016). To identify the intent-to-treat program effects, this paper employs an identification strategy similar to Hahn et al. (2018), exploiting two sources of variation in the duration of program exposure and geographic eligibility. The study compares cohorts of rural girls eligible for the program to older cohorts who marginally missed receiving the stipend. To account for cohort trends, we use urban girls of the same ages, ineligible to be beneficiaries, as a comparison group. The two treatment cohorts are constructed based on the differential intensity of program exposure. As the initial program rollout was grade-specific, one cohort was eligible for the stipend for five years of secondary school, and a second cohort was partially eligible for two years. The strategy assumes that in the absence of the program, the trends in IPV would evolve similarly between treatment and comparison cohorts. We perform various placebo tests to validate the identification strategy. The main result shows that program exposure reduced the likelihood of experiencing IPV by 0.22 standard deviations (SD) for cohorts eligible for the stipend for five years. For partially eligible cohorts, IPV declines by 0.10 SD, but the effects are sensitive across specifications, suggesting that the duration of exposure matters. The study corroborates that the program increased girls’ education by 0.7 to 1.5 years (Hahn et al. 2018; Hahn, Nuzhat, and Yang 2018). We further find that treated women are significantly more likely to complete secondary and tertiary education. Evaluation of potential mechanisms does not indicate significant changes in labor market outcomes, decision-making autonomy, or attitudes towards the acceptability of wife beating. Program exposure delayed the age of first marriage and improved the quality of marital matching as captured in partner characteristics. Women marry men with higher education levels and better quality of employment. Edu- cated partners may hold more egalitarian gender views, and their improved earnings prospects can reduce household stress associated with resource scarcity, potentially mitigating IPV. The study further shows that men who married women eligible for the program are significantly less likely to perpetrate IPV. These results relate to the literature on cash transfers and IPV. Existing studies focus on the short-term impact of resources targeted at adult women. Although they mitigate IPV (see Buller et al. (2018) and Baranov et al. (2021) for a review), often through alleviating household resource stress and improvement in women’s intra-household bargaining, the evidence is mixed, whereby partners also resort to violence 642 Sara and Priyanka to extract resources or backlash effects arise from status inconsistency, shifting power dynamics, and deviation from norms (Bobonis, González-Brenes, and Castro 2013; Litwin, Perova, and Reynolds 2019). Relatedly, there is limited evidence of whether the effects are sustained once the transfer ends. The study contributes to this literature by showing that an education-specific CCT directed at adolescents can have a long-term impact on the prevalence of IPV almost 8–10 years post-transfers if the program affects the pathways to IPV. Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 The analysis suggests that human capital accumulation and delaying marriage formation during the transition to adulthood can set women on a positive trajectory in the long run. There is a marginally negative effect on women’s work, so violence could decrease if partners are less likely to resort to coercive tactics to extract resources. Still, Heath (2014) shows that the positive correlation between labor-force participation and IPV in Bangladesh manifests in women with low education and who marry young. Even with cash transfers, the impact on IPV can depend on the beneficiary women’s education and her educa- tion relative to her partner, so marital matches can be important to understand IPV dynamics (Hidrobo and Fernald 2013). Recent evidence suggests that complementary programs alongside cash transfers can have long-lasting effects on IPV (Peterman and Roy 2021). This study’s results indicate that complemen- tary programming can consider social contextual factors like child marriage in the design of adolescent empowerment programs. This article also contributes to the literature on the effects of education on IPV. Despite a negative as- sociation (Ackerson et al. 2008; Vyas and Watts 2009), the causal evidence based on compulsory school- ing is mixed. Erten and Keskin (2018) find an adverse impact in Turkey on psychological violence and financial control behavior, consistent with instrumental theories as the reform improved women’s labor- market outcomes, with no changes in partner characteristics. In Peru, compulsory schooling reduced IPV even though women’s occupational status improved, accompanied by a delay in marriage formation and changes in partner selection (Weitzman 2018). This paper contextualizes these mixed results as, taken together, the evidence suggests the effect of education may depend on accompanying changes in marital matching.1 The remainder of the paper is organized as follows: section 2 provides details on the FSSSP; section 3 elaborates on the empirical strategy; section 4 describes the data; section 5 presents the main findings and robustness tests; section 6 discusses mechanisms, and section 7 concludes. 2. Background: Female Secondary School Stipend Program (FSSSP) Secondary education in Bangladesh comprises grades 6–10 (ages 11–15), and higher secondary education consists of grades 11 and 12 (ages 16–17). While primary school (grades 1–5, ages 6–10) has been free and compulsory since 1990, secondary school students incur various expenses such as tuition and exam fees. In 1990, gross enrolment in secondary school was low, at 21 percent (14 percent for girls and 27 percent for boys), with a significant urban-rural differential. For instance, in 1991, 5 percent of rural girls compared to 12 percent of boys completed the tenth grade—equivalent to lower secondary schooling (Khandker 1996). The government introduced the Female Secondary School Stipend Program (FSSSP) nationwide2 in 1994 to address the gender disparity in secondary schooling in rural areas. Since the 1990s, girls’ sec- 1 Akyol and Kırdar (2022) study the Turkish reform with additional years of data and find declines in physical violence attributed to changes in schooling and partner characteristics. 2 Several development partners supported the program, including the World Bank, the Asian Development Bank, and the Norwegian Agency for Development Cooperation. The FSSSP was piloted in 1982 in a single upazila (subdistrict). From 1984 to 1992, 8 upazilas out of 492 were included. The success of these pilots led to the nationwide rollout in 1994. The pilots were not extended to all schools in an upazila or for all years before 1994. The limited number of girls in the pilot is considered not treated in the present analysis, which presents a lower bound program impact. The World Bank Economic Review 643 Figure 1. Women’s Average Years of Education by Rural-Urban Birth Cohorts Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Source: Data used Bangladesh Demographic and Health Survey (DHS) 2007, 2011, 2014, and 2017–2018. Note: The figure depicts a two-way scatter plot of women’s average years of education for each birth cohort with line of best fit. The gray dot represents urban women, and the black square represents rural women’s education. ondary school enrolment went up from 44 percent in 1998 to 50 percent in 2007 and 78 percent in 2019.3 In particular, attendance rates increased at a higher rate for rural relative to urban girls (Hahn, Nuzhat, and Yang 2018). The growth in rural secondary school enrolment is often attributed to the FSSSP, which covered more than 2 million girls each year (Khandker, Pitt, and Fuwa 2003; Hahn et al. 2018; Hahn, Nuzhat, and Yang 2018). Figure 1 shows that prior to the program, the trend in women’s average years of schooling remained similar between rural and urban cohorts. There is a steep rise in schooling years for the program-eligible rural cohorts from the 1980s onwards relative to a fairly similar trend for the non-eligible urban cohorts. The FSSSP was designed as a conditional cash transfer program and provided targeted recipients, girls in grades 6–10 residing in rural areas, with cash stipends to cover educational expenses that varied by grade level. Annual stipends ranged from US$18 for grade 6, US$20 for grade 7, US$22 for grade 8, US$36 for grade 9, and US$45 for grade 10. The stipend covered school tuition fees paid directly to the school. The remaining amount (intended to support exam fees, textbooks, stationery, uniforms, etc.) was paid in two annual installments into a savings account in the nearest state bank. To retain program eligibility, recipients had to maintain a 75 percent school attendance rate, satisfactory academic achievement with a test score of at least 45 percent in annual exams, and remain unmarried. 3 “School Enrollement, Secondary (Gross)—Bangladesh,” World Bank, https://data.worldbank.org/indicator/SE.SEC. ENRR?locations=BD (accessed June 21, 2021). 644 Sara and Priyanka Table 1. Years of Program Exposure by Birth Cohort and Grade Eligibility Birth Year 1994 1995 1996 1997 1998 1999 2000 Cohort 1988 6 7 Cohort 1 1987 6 7 8 1986 6 7 8 9 1985 6 7 8 9 10 Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 1984 6 7 8 9 10 1983 6 7 8 9 10 1982 9 10 Cohort 2 1981 9 10 1980 9 10 1979 10 Cohort 3 ... 1971 Source: Based on the author’s calculations. Note: Girls born between 1983 and 1988 were eligible for the stipend for the full five years from 1994, when the earliest cohort was in grade 6. Girls born between 1980 and 1982 were eligible for the stipend for two years. Girls born in 1980 would be 14 and in grade 9 in 1994. The program expanded to grade 10 in 1995, making them eligible to receive the stipend partially for two years. If born in 1981, they would be 13 and in grade 8 in 1994, so they would only receive the stipend for two years in 1995 and 1996 once the program expanded eligibility to grades 9 and 10. If born in 1982, they would be 12 and in grade 7 in 1994, so they would be eligible to receive the stipend in grade 9 and 10 in 1996 and 1997. Girls born between 1971 and 1979 were not eligible to receive the stipend as the youngest cohort was already in grade 10 in 1994. 3. Empirical Strategy To study the long-term effect of the FSSSP on IPV, the study exploits two sources of variation in program rollout. The FSSSP was introduced in phases by grade level to girls in rural areas in grades 6–10 to reduce the cost of secondary education. In 1994, stipends were provided to girls in grades 6 and 9. In 1995, the stipend was expanded to girls in grades 6,7, 9, and 10 (except for grade 8). Since 1996, all girls eligible for the program received the stipend at all grade levels. This program rollout feature generated exogenous variation in the duration of program exposure for different cohorts, a key source of variation utilized to identify program impact. The program targeted girls in rural areas, so geographical eligibility is used as a second source of variation.4 The program rollout by grade level led to some cohorts receiving the stipend for all five years of secondary education and some cohorts receiving the stipend for two years. Treatment and comparison cohorts are defined based on this initial grade eligibility. Cohort 1 represents birth cohorts from 1983 to 1988 (aged 6–11 in 1994) who were eligible to receive the stipend for five years, as they would be in grade 6 of secondary school in 1994. Cohort 2 includes girls born between 1980 and 1982 (aged 12–14 in 1994) eligible to receive the stipend for two years (grades 9 and 10), enrolled in grades 7 to 9 in 1994. The never-treated comparison cohort 3 is defined as girls born between 1971 and 1979 (aged 15–23 in 1994), enrolled in grade 10 or above in 1994, and ineligible to receive the stipend. Therefore, the first source of comparison is between rural cohorts eligible to receive the stipend with older cohorts who narrowly missed eligibility, similar to the approach employed by Duflo (2001) and Muralidharan and Prakash (2017). To account for cohort and nationwide trends, the study uses urban girls in the corresponding cohorts 1–3, typically ineligible for the program, as a second comparison group. Table 1 depicts years of program eligibility by birth cohort and grade level. 4 Metropolitan subdistricts (classified as urban areas) were excluded from the program. As information on urban non- metropolitan subdistricts was not available, it is assumed that all subdistricts in urban areas did not receive the program. Notably, there are very few urban nonmetropolitan subdistricts. Even if women from these areas participated in the pro- gram, it would bias the estimates downward and present a lower-bound program impact. The World Bank Economic Review 645 The intent-to-treat (ITT) impact of program availability is estimated as follows: 2 Yi = α0 + β0 Rurali + β jCohorti j × Rurali + β3 (Xi × Rurali ) + φt + θd + εi (1) j=1 Yi , the outcome variable of interest, captures whether an individual woman i ever experienced physical and sexual violence in their lifetime. The standardized aggregate index of IPV based on binary variables captures various interrelated measures of violence. Cohorti j { j = 1, 2} represents dummy variables for Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 cohorts 1 and 2. Rurali is a dummy variable for whether an individual resides in a rural area. The primary parameter of interest β j captures the differential impact of years of program exposure for cohorts of eli- gible rural girls relative to ineligible older cohorts who narrowly missed the program and corresponding urban cohorts who were ineligible to receive the stipend. β1 captures the ITT impact of receiving the stipend for five years and β2 captures the effect of receiving the stipend for two years. If program expo- sure reduces the incidence of IPV, it is expected that β1 and β2 will exhibit a negative sign. The vector X controls for additional factors: religion (whether Muslim or not) and a wealth index that ranges from 1 to 5 (5 being the richest), with both variables interacted with the Rurali dummy to account for differential effects in rural-urban areas. The analysis incorporates cohort fixed effects φt to account for cohort-specific trends and division fixed effects θd to control for time-invariant geographical characteristics.5 The stan- dard errors are clustered at the program treatment level by birth year × rural-urban residence. Because program effects are assessed on a large number of outcomes, the analysis adjusts the standard errors for multiple hypothesis testing using the Benjamini-Hochberg procedure. The main results remain robust to this adjustment. A key assumption in the identification strategy is that the difference in IPV outcomes between treated and comparison cohorts by rural-urban residence do not exhibit differential trends prior to program intro- duction. To check for differential trends across cohorts, two placebo tests are performed in the robustness section. First, we create a placebo cohort ineligible to receive the program and include them in equation (1). If the analysis captures the ITT program effects, then the impact on the placebo cohort should re- main statistically insignificant while retaining consistent estimates for cohorts 1 and 2. Second, the main analysis is re-estimated using only non-eligible birth cohorts from 1957 to 1970, born 18 years before the youngest cohort in the main sample. The analysis retains the structure of the original cohort construction for this placebo sample. For this falsification test, we do not expect to detect significant effects on either cohorts 1 or 2. These tests validate and provide support for the identification strategy. We further check whether the identification strategy is capturing the program effects. The analysis uses two waves of the nationally representative Bangladesh Household Income and Expenditure Survey (HIES) 1995 and 2000, which contains information on contemporaneous program take-up to show that the treatment variation is associated with the actual recipient of the stipend. The study’s construct of program treatment is linked with program take-up rates of 73 percent for cohort 1 and 53 percent for cohort 2, relative to the mean rates of program recipients in the sample (column 1, table 2). As the outcome measure of being a program recipient is somewhat noisy in 1995, the same estimates are presented using only the HIES 2000. We continue to detect significant effects on cohort 1 (who are still eligible for program take- up) but not for cohort 2, who were beyond secondary school age in 2000. This result provides reassurance of the identification strategy. There are additional potential concerns in the identification of program impact. First, if there is delayed school enrolment beyond the officially recommended ages, cohort 2 may include individuals eligible to receive the stipend for five years. Similarly, cohort 3 could consist of girls who are otherwise considered to be not eligible. This is likely to underestimate the treatment effects for cohort 1, with the direction of bias less clear for cohort 2. Second, while there is a possibility of grade repetition to receive the stipend, repetition rates are low in secondary schools (Shamsuddin 2015). The cost of repeating a grade is high 5 There are six divisions in the data: Barisal, Chittagong, Dhaka, Khulna, Rajshahi, and Sylhet. 646 Sara and Priyanka Table 2. Effect of Program Treatment on Actual Recipient of Stipend Recipient of stipend (HIES Recipient of stipend (HIES 1995 and 2000) 2000) (1) (2) Cohort 1 × Rural 0.051∗∗∗ 0.145∗∗∗ (0.010) (0.023) Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Cohort 2 × Rural 0.037∗∗∗ 0.007 (0.009) (0.007) Rural − 0.004 − 0.028 (0.011) (0.023) Observations 14,486 6,621 R2 0.072 0.239 Mean 0.069 0.123 SD 0.254 0.329 Cohort fixed effect Yes Yes Division fixed effect Yes Yes Source: Data used the Bangladesh Household and Expenditure Survey (HIES) 1995 and 2000. Note: Robust standard error in parenthesis clustered at the birth year × rural/urban level. The analysis uses two waves of the HIES 1995 and 2000 that contain information on whether respondents are recipients of stipends to construct the outcome variable. As the HIES question captures contemporaneous program take-up in the survey year, the study focuses on the two waves of the HIES post-program rollout in 1994 that would include girls from the two treatment cohorts. The measure of stipend recipient is somewhat noisy in HIES 1995; so column 2 presents results using HIES 2000 only. Significance levels ∗p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01. relative to the small stipend amount, minimizing the incentive to remain in a grade for an extra year (Hahn et al. 2018). Third, another concern is the potential inclusion of women in the comparison group who migrated from rural to urban areas even though they may have received the stipend during adoles- cence. While we do not have data on where women were residing at the time of secondary schooling, we combine additional information to account for internal migration. The analysis uses data on whether the respondent has always lived in their current place of residence and, in the case of a move, if they remain in the same type of area as their prior residence (if they made a rural-to-rural or urban-to-urban move), to construct a subsample of individuals the analysis is less likely to misclassify for treatment assignment. The results remain robust in this subsample. Besides, internal migration is generally low in Bangladesh, with only 4.29 percent migrating from rural to urban areas (Hahn et al. 2018). 4. Data and Descriptive Statistics The study uses the nationally representative 2007 wave of the Bangladesh Demographic and Health Sur- vey (DHS) to evaluate the effect of the FSSSP on IPV.6 The DHS provides demographic and socioeconomic data on age, birth year, gender, educational attainment, religion, marital status, employment, and house- hold wealth, along with information on attitude towards acceptability of wife beating and participation in household decision making. The survey comprises 361 primary sampling units (PSUs) clustered at the subdistrict level, covering 121 urban and 240 rural areas. A total of 10,996 ever-married women between the ages of 15–49 were interviewed, with 4,467 selected to respond to the survey module on domestic violence. The primary outcome of interest is whether ever-married women experienced physical or sexual vio- lence from their intimate partner in their lifetime. The survey questions on domestic violence were admin- 6 Only the 2007 wave of the Bangladesh DHS contains data on domestic violence. The World Bank Economic Review 647 Table 3. Summary Statistics Obs. Mean Std. dev. Min Max Cohort 1 2885 0.32 0.47 0 1 Cohort 2 2885 0.18 0.39 0 1 Cohort 1 × Rural 2885 0.21 0.40 0 1 Cohort 2 × Rural 2885 0.12 0.33 0 1 Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Rural 2885 0.63 0.48 0 1 Muslim 2884 0.91 0.29 0 1 Wealth index (1–5) 2885 3.08 1.46 1 5 Domestic Violence (DV) Ever experienced any violence 2885 0.53 0.49 0 1 Ever experienced physical violence 2885 0.49 0.50 0 1 Less severe physical violence 2885 0.48 0.50 0 1 More severe physical violence 2885 0.24 0.42 0 1 Ever experienced sexual violence 2885 0.18 0.38 0 1 Domestic violence index 2885 − 0.002 0.67 −0.46 2.98 (Average z-score) Women’s Characteristics Age 2885 26.81 5.03 18 36 Years of education 2874 4.89 4.46 0 17 Completed secondary education 2874 0.16 0.36 0 1 Higher education 2874 0.09 0.29 0 1 Age at first marriage 2885 15.63 2.83 10 32 Worked in the last 12 months 2885 0.34 0.47 0 1 Worked in the formal sector 2882 0.03 0.18 0 1 Partner’s Characteristics Age 2766 35.93 7.17 17 80 Spousal age gap 2765 9.21 5.08 −7 50 Years of education 2880 4.99 4.94 0 17 Worked in the formal sector 2872 0.28 0.45 0 1 Source: Data used 2007 Bangladesh Demographic and Health Survey (DHS). Note: The table presents the mean, standard deviation, minimum and maximum values, and the number of observations. istered to one eligible male or female respondent in the household, with protocols to ensure the privacy of respondents. For this reason, the data on reported violence is not available from the same couple. The DHS only provides physical and sexual violence data, not psychological violence and financial control measures. The domestic violence data includes binary variables on whether a woman has ever experi- enced physical or sexual violence. Physical violence is captured by acts of pushing, shaking, or throwing objects; slapping; arm twisting and hair-pulling; punching with a fist or otherwise; kicking, dragging, or beating; choking or burning; and threatening with a knife, gun, or any other weapons. Sexual violence data captures whether a woman was physically forced to engage in sexual intercourse. The study con- structs z-scores using the mean and SD of these binary variables that capture interrelated measures of violence and aggregate them into a standardized index. The summary statistics are presented in table 3. The sample is limited to ever-married women aged 18–36 in 2007, who were 5–23 years old in 1994 when the FSSSP was introduced, resulting in a sample size of 2,885 women. The analysis constructs two sets of treated cohorts exposed to the FSSSP at varying intensities due to the program rollout features: 32 percent of women are in cohort 1, of which 21 percent 648 Sara and Priyanka were eligible to receive the stipend for the full five years in rural areas; 18 percent of women are in cohort 2, with 12 percent eligible to receive the stipend partially for 2 years. The disaggregated measures of domestic violence show that 53 percent of women report having ex- perienced either physical or sexual violence: 49 percent of women experienced some form of physical violence–48 percent in the sample experienced less severe forms while 24 percent experienced more severe forms of physical violence.7 The most common act of physical violence reported is slapping, experienced Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 by 46 percent of women; 18 percent of women report the occurrence of sexual violence at least once in their lifetime. On average, women in the sample have 4.9 years of education (slightly less than primary education), with an SD of 4.5 years; 16 percent of women completed secondary education, and 9 percent completed college. With 63 percent of the sample in rural areas, a very low proportion of women work in the formal sector (3 percent), even though 34 percent worked in the last 12 months. The formal sector is defined as work in professional occupations and businesses relative to agriculture or informal work as semi-skilled workers. Early marriage remains a challenge in Bangladesh, with an average age at first marriage of 15.6 and an SD of 2.8. In terms of partner characteristics, women are married to men with an average of five years of education (SD of 4.9), with 28 percent working in the formal sector, which typically pays higher wages. Women tend to marry older men, with an average spousal age gap of 9.2 years. The DHS 1993 data shows that household characteristics between treated and comparison cohorts are similar at baseline (table S1.1 in the supplementary online appendix S1, available with this article at The World Bank Economic Review website). The study estimates the raw difference-in-difference com- parison between cohort 1 and cohort 2 relative to the never-treated comparison cohort 3. There are no significant differences in characteristics such as household head age, gender, household size, literacy, and assets measures, except that cohort 1 girls are from households that are slightly less likely to be classified as poor (10 percent significance). 5. Results The analysis begins by evaluating the effect of the FSSSP on education outcomes. Program exposure increases years of education by 1.5 and 0.7 years for cohorts 1 and 2, respectively (column 1 of table S1.2). The estimates are similar to existing studies that find increases in schooling between 0.55 and 1.3 years (Hahn et al. 2018; Hahn, Nuzhat, and Yang 2018; Tanaka, Takahashi, and Otsuka 2021). Both cohorts are more likely to complete secondary schooling (column 2). Furthermore, there are spillover effects whereby incentives targeted at secondary education also translate into a higher likelihood of girls’ completing college (column 3).8 Table 4 presents the main findings using a standardized index measure of IPV. Column 1 shows esti- mates without individual and household controls or cohort and division fixed effects. Women in cohort 1 experienced a 0.17 SD decline in IPV, with a 0.04 SD decline for cohort 2. We progressively add controls for religion, wealth index, and cohort and division fixed effects. The preferred specification in column 4 with the full set of controls yields larger effects with a 0.22 and 0.10 SD decline in IPV for cohort 1 and cohort 2, respectively. On an annual basis, each year of program exposure corresponds to a 0.04 to 0.05 SD reduction in IPV. Table S1.3 examines individual components of the IPV index, indicating that the results are driven by the reduced likelihood of experiencing multiple forms of violence, includ- 7 The analysis defines less severe forms of violence as slapping, pushing, shaking, or throwing objects. More severe forms of violence include punching with a fist, kicking, dragging, beating; choking, burning; and threatening with a knife, gun, or other weapons. 8 As a placebo test, the analysis verifies that the program did not affect secondary schooling for a similar cohort of rural men who were ineligible for the stipend. However, men are more likely to complete college, potentially due to a changing pool of better-qualified brides. The World Bank Economic Review 649 Table 4. Impact of FSSSP on Domestic Violence Domestic violence (standardized z-scores) (1) (2) (3) (4) Cohort 1 × Rural − 0.169∗∗∗ − 0.166∗∗∗ − 0.223∗∗∗ − 0.219∗∗∗ (0.041) (0.041) (0.071) (0.071) Cohort 2 × Rural − 0.041 − 0.049 − 0.101∗∗ -0.100∗∗ Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 (0.035) (0.035) (0.043) (0.043) Rural 0.098∗∗∗ -0.406∗∗∗ -0.378∗∗∗ -0.373∗∗∗ (0.025) (0.056) (0.126) (0.126) Muslim 0.298∗∗∗ 0.298∗∗∗ 0.308∗∗∗ (0.000) (0.066) (0.067) Muslim × Rural 0.025 0.037 0.008 (0.042) (0.080) (0.082) Wealth Index − 0.185∗∗∗ − 0.184∗∗∗ − 0.191∗∗∗ (0.000) (0.025) (0.025) Wealth Index × Rural 0.087∗∗∗ 0.085∗∗∗ 0.095∗∗∗ (0.015) (0.030) (0.030) Observations 2885 2884 2884 2884 R2 0.004 0.044 0.048 0.055 Mean 0 0 0 0 Std. dev. 1 1 1 1 Cohort fixed effect No No Yes Yes Division fixed effect No No No Yes Source: Data used 2007 Bangladesh Demographic and Health Survey (DHS). Note: Robust standard errors in parenthesis clustered at the birth year × rural/urban level. To construct an aggregate standardized z score index of IPV the study uses eight questions from the DHS that ask if the husband ever perpetrated various types of physical or sexual violence. Significance levels ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 ing being slapped, punched, kicked, and dragged or getting arms twisted and hair pulled. The decline in these physical-violence measures ranges between 6 to 14 percentage points and is robust to adjusting for multiple hypothesis testing for cohort 1 only. The decrease in the incidence of forced sex for cohort 1 is not robust to adjusting the standard error for multiple testing. We test the sensitivity of the results to alternate measures of IPV. Using a binary variable for whether women ever experienced physical or sexual violence, the study finds that program exposure led to a 10 and 4.5 percentage point decrease in IPV for cohorts 1 and 2 (column 1 table S1.4). Relative to mean rates of IPV, this presents a decline of 19 percent for cohort 1 and 8 percent for cohort 2. There is a significant decline in both severe and less severe forms of physical violence (columns 2, 3), with a larger impact for cohort 1. Using an alternate method for constructing the IPV index based on factor analysis (column 4) finds comparable, although slightly smaller, declines in IPV of 0.16 and 0.08 SD for cohorts 1 and 2. We compare with other policy interventions to provide context to interpreting the program effect size. The compulsory schooling law in Peru that targeted students at similar grade levels led to an 8 percent reduction in IPV (Weitzman 2018). The FSSSP provided financial incentives and stipulations regarding delaying marriage, which could explain the larger, 19 percent, decline. Evidence from antipoverty cash transfer programs in Mexico and Ecuador finds decreases in physical IPV between 19 percent and 45 percent, although in some cases, some subgroups of women even experience a rise in violence (Angelucci 2008; Bobonis, González-Brenes, and Castro 2013; Hidrobo and Fernald 2013; Hidrobo, Peterman, and Heise 2016). In the context of Bangladesh, Roy et al. (2019) show that a program that provided poor households with cash or food transfers and behavior change communication led to a 26 percent decline 650 Sara and Priyanka Table 5. Robustness Tests for Impact of FSSSP on Domestic Violence District FE Division × Cohort FE Narrower cohort Placebo cohort Non-eligible cohorts Non-migrants (1) (2) (3) (4) (5) (6) Cohort 1 × Rural − 0.219∗∗ − 0.247∗∗∗ − 0.245∗∗ − 0.233∗∗∗ 0.005 − 0.254∗∗ (0.090) (0.074) (0.089) (0.077) (0.069) (0.106) Cohort 2 × Rural − 0.095 − 0.095∗ − 0.093∗ − 0.114∗∗ 0.049 − 0.105 Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 (0.100) (0.049) (0.048) (0.053) (0.057) (0.087) Cohort 3 × Rural − 0.031 (0.054) Rural − 0.252 − 0.378∗∗ − 0.418∗∗∗ − 0.357∗∗∗ − 0.265 − 0.863∗∗∗ (0.153) (0.139) (0.136) (0.130) (0.299) (0.209) Muslim 0.340∗∗∗ 0.296∗∗∗ 0.337∗∗∗ 0.309∗∗∗ 0.195 0.166 (0.083) (0.076) (0.072) (0.067) (0.121) (0.103) Muslim × Rural − 0.066 0.013 − 0.032 0.007 − 0.165 0.160 (0.096) (0.092) (0.087) (0.082) (0.158) (0.112) Wealth Index − 0.179∗∗∗ − 0.194∗∗∗ − 0.211∗∗∗ − 0.191∗∗∗ − 0.199∗∗∗ − 0.255∗∗∗ (0.024) (0.026) (0.025) (0.025) (0.054) (0.048) Wealth Index × Rural 0.081∗∗ 0.098∗∗∗ 0.113∗∗∗ 0.095∗∗∗ 0.096 0.179∗∗∗ (0.031) (0.030) (0.030) (0.030) (0.059) (0.051) Observations 2884 2884 2305 2884 1324 2055 R2 0.081 0.082 0.061 0.055 0.062 0.062 Mean 0.000 0.000 0.014 0.000 0.000 − 0.000 Std. dev. 1.000 1.000 1.010 1.000 1.000 1.000 Cohort fixed effect Yes Yes Yes Yes Yes Yes Division fixed effect No Yes Yes Yes Yes Yes Source: Data used 2007 Bangladesh Demographic and Health Survey (DHS). Note: Robust standard errors in parenthesis clustered at the birth year × rural/urban level. In column 3, the analysis defines the cohorts more narrowly where cohort 1 represents women born between 1983 and 1986, cohort 2 represents women born between 1980 and 1982, and the control cohort includes women born between 1973 and 1979. Column 4 performs a placebo test by including a placebo-treated cohort 3, which includes individuals born between 1976 and 1979. Column 5 performs another falsification test where the analysis tests the impact on cohorts born 14 years before the cohort in the main sample. This sample contains women born between 1957 and 1970, where cohort 1 includes women born between 1965 and 1970 and cohort 2 includes women born between 1962 and 1964. The non-migrant sample is defined as individuals who have always lived in the same place of residence or those who made rural-to-rural or urban-to-urban move, so there is no misclassification of treatment assignment. Significance levels ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 in physical violence 6 to 10 months after the program ended, with a larger, 54 percent, decline four years after (Roy et al. 2022). While the magnitude of our results is smaller, they are sizeable given the longer time horizon that is considered. Additional robustness checks are presented in table 5. First, we test whether the results are robust to other location- and cohort-specific trends. In column 1, the analysis controls for district fixed effects, a lower tier of administration than divisions, to better account for local area-specific differences. The results are identical to the baseline but are no longer statistically significant for cohort 2. Column 2 incorporates division × cohort fixed effects to account for cohort-specific trends, and the results remain broadly similar. If the age gap between the treated and comparison cohorts is large, the comparison may not be meaningful. The results are re-estimated with a narrower age group by dropping the youngest (bottom two years) and oldest (top two years) girls in the sample, comprising women born between 1973 and 1986. The results remain similar with a slighter larger effect for cohort 1 of a 0.31 SD reduction in IPV (column 3). We next construct placebo cohorts to test if there are differential trends in IPV between early and later cohorts. The identification strategy requires the trend in IPV to remain parallel for the treatment and comparison group in the absence of the program. First, similar to Hahn et al. (2018), we split the comparison cohort (born between 1971 and 1979) into two groups. We define a placebo treatment co- The World Bank Economic Review 651 hort 3 comprising women born between 1976 and 1979, intended to capture differential trends in IPV across never-treated cohorts. Column 4 finds that the effect on the placebo cohort 3 is not statistically significant. The impact on cohorts 1 and 2 remains similar, alleviating concerns about capturing differ- ential time trends instead of the ITT program effects. Second, column 5 presents a falsification test us- ing non-eligible cohorts born 18 years before the youngest cohort in the main sample between 1957 and 1970. We construct a placebo cohort 1, comprising women born between 1965 and 1970, and Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 a placebo cohort 2, of women born between 1962 and 1964, to capture the structure of the original cohort construction. Reassuringly, there is no statistically significant effect for the non-eligible placebo cohorts. The next set of checks addresses potential issues of sample bias. Internal migration can lead to the misclassification of treatment assignments. Column 6 presents results for a subsample of individuals who have always lived in their present location or have made a rural-to-rural or urban-to-urban move (so the analysis is less likely to misclassify treatment assignment). The effect is slightly larger for cohort 1 (statistically insignificant for cohort 2), suggesting migration could bias the estimates downward, so the analysis likely captures a lower-bound impact of the program. The sample consists of ever-married women on account of DHS procedures for collecting IPV data, so there could be concerns of sample selection bias as it excludes unmarried women.9 If the never-married treated women have greater autonomy and experience less IPV, the study could be underestimating the program’s impact. Conversely, if their increased autonomy leads to a backlash effect with an increase in IPV, it might overestimate the program effect. We conduct a bounding exercise to address this selection issue using data on all married and never-married women. The analysis determines how large the program- induced increase in IPV would need to be for the sample of never-married women to attenuate the main program effect to zero. To implement this, the analysis assigns IPV z-scores to the never-married women from a uniform distribution, repeating the exercise multiple times with new intervals of the uniform distribution. Figure S1.1 plots the simulated coefficient on IPV for the never married and the combined sample of married and never-married women to illustrate the relevant range that renders the program effect zero. The results indicate that the rise in IPV would need to be as high as 0.3 to 0.8 SD for the never-married sample to overturn the main results. Fakir et al. (2016) show that increased autonomy in Bangladesh is linked to a 27 percent rise in IPV. Assuming that never-married women have more autonomy, then based on mean IPV rates in the sample, they would have to experience almost three times higher rates of IPV (61 percent) to bias the program impact to zero. Overall, the results demonstrate that program exposure led to a significant decrease in IPV, predomi- nantly for rural women eligible to receive the stipend for the full five years of secondary education. The mixed effects on the cohort that was partially eligible for two years suggest that the duration of exposure matters for program efficacy. Dervisevic, Perova, and Sahay (2021) similarly find that short-term exposure of 1.5 years to a CCT program during adolescence did not significantly impact IPV in the Philippines. 6. Mechanisms This section evaluates plausible mechanisms through which the FSSSP could have reduced IPV incidence: women’s labor market outcomes and autonomy, attitude towards domestic violence, and changes in mar- ital outcomes. 9 Ten percent of women have never married in the combined data of all ever-married and never-married women born between 1971 and 1988 (the sample cohort). 652 Sara and Priyanka Table 6. Impact of FSSP on Labor-Market Outcomes, Autonomy, and Attitude towards Wife Beating. Work in formal Worked in the last 12 Decision & autonomy Attitude towards wife sector months index beating index (1) (2) (3) (4) Cohort 1 × Rural 0.022 −0.064∗ 0.060 −0.060 (0.015) (0.033) (0.049) (0.054) Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Cohort 2 × Rural 0.025 −0.025 0.031 −0.025 (0.019) (0.028) (0.040) (0.040) Rural −0.031 −0.036 -0.096 -0.376∗∗ (0.047) (0.071) (0.131) (0.142) Muslim −0.020 0.078∗ −0.027 −0.018 (0.025) (0.043) (0.077) (0.080) Muslim × Rural 0.015 −0.131∗∗ 0.056 0.210∗∗ (0.027) (0.064) (0.094) (0.099) Wealth index 0.008 −0.075∗∗∗ 0.033 −0.113∗∗∗ (0.007) (0.011) (0.020) (0.018) Wealth Index × Rural −0.002 0.029∗∗ -0.009 0.061∗∗ (0.007) (0.013) (0.027) (0.024) Observations 2881 2884 2883 2883 R2 0.021 0.080 0.061 0.040 Mean 0.034 0.342 0.036 −0.027 Std. dev. 0.182 0.475 0.837 0.842 Cohort fixed effect Yes Yes Yes Yes Division fixed effect Yes Yes Yes Yes Source: Data used 2007 Bangladesh Demographic and Health Survey (DHS). Note: Robust standard errors in parenthesis clustered at the birth year × rural/urban level. The study uses factor analysis to construct the aggregate index for decision making and attitude towards wife beating. The Kaiser criterion is applied to identify and compute the composite scores for the aggregate indices and retain only one factor with an eigenvalue higher than one. The results are similar when an aggregate standardized z-score index is used for each measure. Significance levels ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 Labor Market Outcomes and Participation in Decision-Making The analysis first tests whether the FSSSP led to changes in women’s labor market outcomes and autonomy in household decision making. As the program improved women’s educational attainment, it could enable them to take advantage of better labor market opportunities that increase their resources, strengthen their outside options and their autonomy, and bargaining status within the household (Anderson and Eswaran 2009; Aizer 2010; Chin 2012; Hidrobo and Fernald 2013). Moreover, Majlesi (2016) shows that better labor-market options can improve nonworking women’s decision making; thus, increased education could improve autonomy even without accompanying changes in labor market participation. Table 6 shows no significant effects on the likelihood of women working in the formal sector, a proxy for employment quality. In fact, there is a negative effect on the probability of working in the last 12 months (10 percent significance). Earlier research finds a mixed impact of the program on labor market outcomes. For example, Hahn, Nuzhat, and Yang (2018) do not detect an effect on the likelihood of women working, while Shamsuddin (2015) finds increases in labor market participation, albeit at low wages, as women have difficulty finding good jobs. Similarly, limited formal employment opportunities in rural areas could be a plausible explanation for the present study’s findings. In addition, the sample only captures ever-married women, so the labor market effects could be downward biased if more-motivated younger women who pursue professional opportunities delay marriage formation and are thus excluded from the analysis. The World Bank Economic Review 653 The present study does not detect a statistically significant impact of the program on an aggregate index of household decision-making and autonomy constructed using factor analysis (table 6, column 3). The index captures women’s participation in decisions related to their health care, large household purchases, household purchases for daily needs, and visits to friends and relatives. The individual components of the index mainly remain statistically insignificant (table S1.5 columns 1–4), except that there is a positive effect on women’s participation in decisions related to household purchases for daily needs. Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Based on these findings, it does not appear to be the case that improvements in labor market outcomes or changes in autonomy are likely causes of the decline in IPV. An alternate explanation consistent with instrumental theories of violence is that if the program lowers the probability of women working, their partners may be less likely to resort to violence to extract resources, reducing the risk of experiencing IPV. Attitude Towards Violence We next examine whether the program led to changes in attitudes towards domestic violence as captured by acceptability towards wife beating. Women with higher levels of education may be less tolerant of threats and acts of violence due to altered beliefs regarding gender equality and better knowledge of laws and their rights. The empirical evidence on whether increased education changes attitudes towards domestic violence is mixed (Dinçer, Kaushal, and Grossman 2014; Friedman et al. 2016; Erten and Keskin 2018; Weitzman 2018). Women were asked whether a husband is justified in beating his wife under a series of circumstances: if she goes out without telling him, neglects the children, argues with him, or refuses sexual intercourse. We construct an aggregate index of attitudes toward violence with the individual components using factor analysis. Descriptive statistics suggest that the acceptability of violence is lowest for women who com- pleted secondary schooling or higher, based on the percentage of women who agree that at least one of the reasons above justifies wife beating. Column 4 of table 6 shows that the program did not lead to a statistically significant change in the index on the acceptability of wife beating. However, examining the individual components (table S1.5 columns 5–8) shows a 0.04 SD reduction in the acceptability of wife beating for neglecting children for cohort 1, with other components remaining statistically insignificant. The effect on cohort 2 is mixed, with increased acceptability of wife beating for going out without informing her husband and reduced acceptability of violence for refusing sex. For this partially treated cohort, program exposure may not have been sufficiently long to counter social norms regarding the types of domestic violence that are acceptable or their perception of appropriate gender roles. There are no robust changes in attitudes towards domestic violence due to program exposure, sug- gesting that it likely does not explain the main results. However, this finding mitigates concerns that the analysis is overestimating self-reported IPV for better-educated women if they are expected to report incidences at different rates than less-educated women. Marital Outcomes Finally, the study evaluates whether changes in marital outcomes affected the decline in IPV. Given the pervasive prevalence of early marriage in Bangladesh, the timing of secondary school often competes with marriage formation and early childbearing. Field and Ambrus (2008) demonstrate that the social and financial pressure to marry early leads rural women in Bangladesh to attain less schooling. The FSSSP was designed to address these dual concerns of low levels of girls’ education and early marriage, which present substantial risk factors for experiencing IPV (Morrison, Ellsberg, and Bott 2007). Matching theory suggests that marriage formation can involve positive assortative matching on edu- cation (Fernandez, Guner, and Knowles 2005; Anderberg and Zhu 2014; Chiappori, Dias, and Meghir 2018). According to the WHO, low levels of education can augment the risk of perpetration of IPV along- side a greater likelihood of experiencing violence. As the FSSSP raised the likelihood of women’s secondary 654 Sara and Priyanka Table 7. Impact of FSSP on Marriage Market Outcomes Age at first Partner years of Partner in formal marriage education sector Spousal age gap (1) (2) (3) (4) Cohort 1 × Rural 0.676∗∗∗ 1.053∗∗∗ 0.078∗∗∗ 0.396 (0.141) (0.252) (0.028) (0.275) Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Cohort 2 × Rural 0.242 0.460∗∗ 0.054 −0.366 (0.144) (0.203) (0.048) (0.341) Rural −0.374 1.700∗∗ −0.080 −0.515 (0.333) (0.640) (0.095) (0.651) Muslim −1.519∗∗∗ −0.066 −0.080 −0.547 (0.305) (0.408) (0.063) (0.384) Muslim × Rural 0.740∗∗ −0.606 0.061 0.016 (0.345) (0.515) (0.070) (0.484) Wealth index 0.723∗∗∗ 2.122∗∗∗ 0.104∗∗∗ −0.100 (0.051) (0.093) (0.010) (0.097) Wealth Index × Rural −0.224∗∗∗ −0.433∗∗∗ −0.015 0.120 (0.072) (0.125) (0.013) (0.168) Observations 2884 2879 2871 2765 R2 0.151 0.309 0.125 0.021 Mean 15.63 4.984 0.278 9.208 Std. dev. 2.825 4.934 0.448 5.084 Cohort fixed effect Yes Yes Yes Yes Division fixed effect Yes Yes Yes Yes Source: Data used 2007 Bangladesh Demographic and Health Survey (DHS). Note: Robust standard errors in parenthesis clustered at the birth year × rural/urban level. Significance levels ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 and tertiary education completion, it may have enabled them to marry better educated men with better employment prospects, which can affect the risk of IPV in two important ways. First, a “higher quality” husband may be less likely to resort to violence. Second, marrying men with better human capital and employment prospects can potentially alleviate financial stress and conflict within the household leading to lower odds of marital dissolution (Weiss and Willis 1997) and domestic violence (Alonso-Borrego and Carrasco 2017; Bhalotra et al. 2019). Table 7 shows that the FSSSP delayed the age at first marriage for cohort 1 by 0.7 years (column 1). With average age at first marriage of 15.6, this presents an increase of 4.5 percent to 16.3 years. While this is still low, Roychowdhury and Dhamija (2021) show that even delaying marriage by one year can lead to significant declines in physical violence. There are changes in marital matches as captured by partner characteristics. Women marry men with higher years of education (an additional one year) who are 7.8 percentage points more likely to work in the formal sector (a proxy for a better quality of employment). The education gap between spouses, where women are more educated than their husbands, is associated with a higher risk of IPV(Ackerson et al. 2008; Hidrobo and Fernald 2013). Therefore, matching with better-educated men could have mitigated potential adverse consequences of a widening spousal gap in education. There are no changes in the spousal age gap, which is nine years on average in the sample. The changes in marital outcomes mainly pertain to cohort 1, and the effects are not statistically significant for cohort 2, except that they also marry men with higher years of education. We conduct one other test of marital match quality. Specifically, we assess whether being married to women exposed to the stipend program affected men’s propensity to engage in IPV. Ever-married men were asked the same set of questions on the various indicators of IPV, soliciting whether they had engaged The World Bank Economic Review 655 in violent acts against their wives. Table S1.6 shows that men are significantly less likely to perpetrate violence with effect sizes between 0.09 and 0.20 SD across different measures of IPV. Consistent with the results for women’s marital outcomes, the changes in men’s behavior are captured for those potentially matched with women in cohort 1. Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 Discussion of Mechanisms Given the interlinkages between education and marital outcomes, the study evaluates whether years of education, the key focus of the FSSSP, is statistically the primary mechanism underlying the results using the method developed by Acharya, Blackwell, and Sen (2016). Table S1.7 shows that accounting for years of education reduces the program’s impact on IPV to 0.15 SD, but it is statistically not the only mechanism. Incorporating the remaining mechanisms in a regression-based mediation analysis, namely age of first marriage and partner’s education and employment, and the alternate explanation of instrumental violence operating through women’s work similarly reduces the effect on IPV to 0.13 SD, with results that are significant only at the 10 percent level. Quantifying the relative importance of these various mechanisms finds that years of education account for 19 percent of the treatment effect, marital outcomes account for 13.8 percent (7.5 percent attributed to the delay in age of marriage, 3 percent to partner’s education, and 3.4 percent to partner’s employment), and the decrease in women’s work account for 4 percent of the impact (table S1.8). However, the effect of partners’ education and women’s work is not statistically significant. It could be the case that the impact of partners’ education operates through greater participation in formal sector work. The mediation analysis corroborates that education and marital matching present the dominant pathways through which the program affected the long-term risk of IPV. Appendix S2 outlines the details of the mediation analysis. 7. Conclusion This paper evaluates whether an education-specific conditional cash transfer (CCT) targeted at secondary school-aged adolescent girls had long-term later-life effects on the prevalence of IPV in Bangladesh. The Female Secondary School Stipend Program, introduced nationwide in 1994 to rural women, aimed to reduce the cost of secondary schooling through tuition and cash stipends with eligibility conditional on satisfying grade and attendance requirements and remaining unmarried. This study exploits two sources of variation derived from program rollout features in the intensity of program exposure and geographic eligibility to identify program effects. The main results show that rural women eligible for the program are significantly less likely to experience IPV. Program exposure led to an average increase of 1.5 additional years of education for cohorts eligible to receive the stipend for the full five years of secondary school. There are no significant changes in women’s empowerment through labor market participation or increased autonomy and decision making within the household. The study detects significant changes in marital outcomes, including delays in marriage formation (the age of first marriage increases by 0.7 years) and changes in the quality of partner charac- teristics such as their schooling and employment. Further, men who married cohorts of women eligible for the program are significantly less likely to perpetrate violence. The results suggest that positive assortative marital matching on education may be important for understanding IPV dynamics. An important policy implication of the finding is that cash stipend programs targeted at adolescent girls before marriage can have long-term, sustained spillovers in boosting women’s well-being through human capital formation and, in turn, the quality of marital matches. This may be particularly relevant in settings with low education levels and a high prevalence of early marriage. While CCT programs in- volve significant government expenditures, their cost-benefit analysis ought to consider the externalities generated by women’s improved life outcomes. 656 Sara and Priyanka Conflict of Interest None. Data Availability Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 The Bangladesh Demographic and Health Survey (DHS) 2007 is available at https://dhsprogram.com/ data/dataset/Bangladesh_Standard-DHS_2007.cfm?flag=0. References ———. 2018. “Does Increasing Women’s Education Reduce Their Risk of Intimate Partner Violence? Evidence from an Education Policy Reform.” Criminology 56(3): 574–607. Erten, B., and P. Keskin. 2022. “ Does Knowledge Empower? Education, Legal Awareness and Intimate Partner Vio- lence.” Feminist Economics 28 (4): 29–59. Acharya, A., M. Blackwell, and M. Sen. 2016. “Explaining Causal Findings without Bias: Detecting and Assessing Direct Effects.” American Political Science Review 110(3): 512–29. Ackerson, L.K., I. Kawachi, E.M. Barbeau, and S.V. Subramanian. 2008. “Effects of Individual and Proximate Educa- tional Context on Intimate Partner Violence: A Population-Based Study of Women in India.” American Journal of Public Health 98(3): 507–14. Aizer, A. 2010. “The Gender Wage Gap and Domestic Violence.” American Economic Review 100(4): 1847–59. Akyol, P., and M.G. Kırdar. 2022. “Compulsory Schooling Reform and Intimate Partner Violence in Turkey.” Euro- pean Economic Review, 104313. Alonso-Borrego, C., and R. Carrasco. 2017. “Employment and the Risk of Domestic Violence: Does the Breadwinner’s Gender Matter?” Applied Economics 49(50): 5074–91. Anderberg, D., and Y. Zhu. 2014. “What a Difference a Term Makes: The Effect of Educational Attainment on Marital Outcomes in the UK.” Journal of Population Economics 27(2): 387–419. Anderson, S., and M. Eswaran. 2009. “What Determines Female Autonomy? Evidence from Bangladesh.” Journal of Development Economics 90(2): 179–91. Angelucci, M. 2008. “Love on the Rocks: Domestic Violence and Alcohol Abuse in Rural Mexico.” BE Journal of Economic Analysis & Policy 8(1). Bandiera, O., N. Buehren, R. Burgess, M. Goldstein, and S. Gulesci, Imran Rasul, and Munshi Sulaiman. 2020. “Women’s Empowerment in Action: Evidence from a Randomized Control Trial in Africa.” American Economic Journal: Applied Economics 12(1): 210–59. Bangladesh Bureau of Statistics. 2016. “Report on Violence against Women (VAW) Survey 2015.” https://asiapacific .unfpa.org/sites/default/files/pub-pdf/Bangladesh_VAW_survey_report_2015_compressed.pdf. Baranov, V., L. Cameron, D.C. Suarez, and C. Thibout. 2021. “Theoretical Underpinnings and Meta-Analysis of the Effects of Cash Transfers on Intimate Partner Violence in Low-and Middle-Income Countries.” Journal of Development Studies 57(1): 1–25. Bhalotra, S., U. Kambhampati, S. Rawlings, and Z. Siddique. 2019. “Intimate Partner Violence: The Influence of Job Opportunities for Men and Women.” World Bank Economic Review Bobonis, G.J., M. González-Brenes, and R. Castro. 2013. “Public Transfers and Domestic Violence: The Roles of Private Information and Spousal Control.” American Economic Journal: Economic Policy 5(1): 179–205. Buller, A.M., A. Peterman, M. Ranganathan, A. Bleile, M. Hidrobo, and L. Heise. 2018. “A Mixed-Method Review of Cash Transfers and Intimate Partner Violence in Low-and Middle-Income Countries.” World Bank Research Observer 33(2): 218–58. Chatterjee, S., and P. Poddar. 2022. “Women’s Empowerment and Intimate Partner Violence: Evidence from a Multi- dimensional Policy in India.” Economic Development and Cultural Change. https://doi.org/10.1086/721281. Chiappori, P.-A., M.C. Dias, and C. Meghir. 2018. “The Marriage Market, Labor Supply, and Education Choice.” Journal of Political Economy 126(S1): S26–72. The World Bank Economic Review 657 Chin, Y.-M. 2012. “Male Backlash, Bargaining, or Exposure Reduction? Women’s Working Status and Physical Spousal Violence in India.” Journal of Population Economics 25(1): 175–200. Dervisevic, E., E. Perova, and A. Sahay. 2021. “Long-Term Impacts of Short Exposure to Conditional Cash Trans- fers in Adolescence: Evidence from the Philippines.” Policy Research Working Paper; No. 9617. C World Bank, Washington, DC. http://hdl.handle.net/10986/35439, License: CC BY 3.0 IGO. Dinçer, M.A., N. Kaushal, and M. Grossman. 2014. “Women’s Education: Harbinger of Another Spring? Evidence Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 from a Natural Experiment in Turkey.” World Development 64: 243–58. Duflo, E. 2001. “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment.” American Economic Review 91(4): 795–813. Edmonds, E., B. Feigenberg, and J. Leight. 2021. “Advancing the Agency of Adolescent Girls.” Review of Economics and Statistics, 1–46. Erten, B., and P. Keskin. 2018. “For Better or for Worse?: Education and the Prevalence of Domestic Violence in Turkey.” American Economic Journal: Applied Economics 10(1): 64–105. Fakir, A.M.S., A. Anjum, F. Bushra, and N. Nawar. 2016. “The Endogeneity of Domestic Violence: Understanding Women Empowerment through Autonomy.” World Development Perspectives 2: 34–42. Fernandez, R., N. Guner, and J. Knowles. 2005. “Love and Money: Inequality, Education, and Marital Sorting.” Quarterly Journal of Economics 120(1): 273–344. Field, E., and A. Ambrus. 2008. “Early Marriage, Age of Menarche, and Female Schooling Attainment in Bangladesh.” Journal of Political Economy 116(5): 881–930. Friedman, W., M. Kremer, E. Miguel, and R. Thornton. 2016. “Education as Liberation?” Economica 83(329): 1–30. Hahn, Y., A. Islam, K. Nuzhat, R. Smyth, and H.-S. Yang. 2018. “Education, Marriage, and Fertility: Long-Term Evidence from a Female Stipend Program in Bangladesh.” Economic Development and Cultural Change 66(2): 383–415. Hahn, Y., K. Nuzhat, and H.-S. Yang. 2018. “The Effect of Female Education on Marital Matches and Child Health in Bangladesh.” Journal of Population Economics 31(3): 915–36. Heath, R. 2014. “Women’s Access to Labor Market Opportunities, Control of Household Resources, and Domestic Violence: Evidence from Bangladesh.” World Development 57: 32–46. Hidrobo, M., and L. Fernald. 2013. “Cash Transfers and Domestic Violence.” Journal of Health Economics 32(1): 304–19. Hidrobo, M., A. Peterman, and L. Heise. 2016. “The Effect of Cash, Vouchers, and Food Transfers on Intimate Partner Violence: Evidence from a Randomized Experiment in Northern Ecuador.” American Economic Journal: Applied Economics 8(3): 284–303. Khandker, S. 1996. Education Achievements and School Efficiency in Rural Bangladesh. Washington, DC: The World Bank. Khandker, S., M. Pitt, and N. Fuwa. 2003. “Subsidy to Promote Girls’ Secondary Education: The Female Stipend Program in Bangladesh.” https://mpra.ub.uni-muenchen.de/23688/. Kilburn, K., A. Pettifor, J.K. Edwards, A. Selin, R. Twine, C. MacPhail, R. Wagner, J.P. Hughes, J. Wang, and K. Kahn. 2018. “The Effect of a Conditional Cash Transfer for HIV Prevention on the Experience of Partner Violence for Young Women: Evidence from a Randomized Experiment in South Africa HPTN 068.” Journal of the International AIDS Society 21 (s1): e25043. Lefgren, L., and F. McIntyre. 2006. “The Relationship between Women’s Education and Marriage Outcomes.” Journal of Labor Economics 24(4): 787–830. Litwin, A., E. Perova, and S.A. Reynolds. 2019. “A Conditional Cash Transfer and Women’s Empowerment: Does Bolsa Familia Influence Intimate Partner Violence?” Social Science & Medicine 238: 112462. Majlesi, K. 2016. “Labor Market Opportunities and Women’s Decision Making Power within Households.” Journal of Development Economics 119: 34–47. Mansour, H., and T. McKinnish. 2014. “Who Marries Differently Aged Spouses? Ability, Education, Occupation, Earnings, and Appearance.” Review of Economics and Statistics 96(3): 577–80. Morrison, A., M. Ellsberg, and S. Bott. 2007. “Addressing Gender-Based Violence: A Critical Review of Interventions.” World Bank Research Observer 22(1): 25–51. Muralidharan, K., and N. Prakash. 2017. “Cycling to School: Increasing Secondary School Enrollment for Girls in India.” American Economic Journal: Applied Economics 9(3): 321–50. 658 Sara and Priyanka Peterman, A., and S. Roy. 2021. Cash Transfers and Intimate Partner Violence (IPV) in Low-and Middle-Income Settings: A Joint Research Agenda to Inform Policy and Practice. Intl Food Policy Res Inst. Roy, S., M. Hidrobo, J. Hoddinott, and A. Ahmed. 2019. “Transfers, Behavior Change Communication, and Intimate Partner Violence: Postprogram Evidence from Rural Bangladesh.” Review of Economics and Statistics 101(5): 865–77. Roy, S., M. Hidrobo, J. Hoddinott, B. Koch, and A. Ahmed. 2022. “Can Transfers and Complementary Nutrition Pro- Downloaded from https://academic.oup.com/wber/article/37/4/640/7192050 by University of Oxford user on 12 December 2023 gramming Reduce Intimate Partner Violence Four Years Post-Program? Experimental Evidence from Bangladesh.” Journal of Human Resources, 0720–11014R2. Roychowdhury, P., and G. Dhamija. 2021. “The Causal Impact of Women’s Age at Marriage on Domestic Violence in India.” Feminist Economics, 27 (3): 188–220. Shamsuddin, M. 2015. “Labour Market Effects of a Female Stipend Programme in Bangladesh.” Oxford Development Studies 43(4): 425–47. Svec, J., and T. Andic. 2018. “Cooperative Decision-Making and Intimate Partner Violence in Peru.” Population and Development Review 44(1): 63. Tanaka, T., K. Takahashi, and K. Otsuka. 2021. “Gains from Female Education in Rural Bangladesh: A Multidimen- sional Approach.” International Journal of Educational Development 81: 102344. Vyas, S., and C. Watts. 2009. “How Does Economic Empowerment Affect Women’s Risk of Intimate Partner Violence in Low and Middle Income Countries? A Systematic Review of Published Evidence.” Journal of International Development 21(5): 577–602. Weiss, Y., and R.J. Willis. 1997. “Match Quality, New Information, and Marital Dissolution.” Journal of Labor Eco- nomics 15(1, Part 2): S293–S329. Weitzman, A. 2014. “Women’s and Men’s Relative Status and Intimate Partner Violence in India.” Population and Development Review 40(1): 55–75. World Health Organization, 2021. “Violence Against Women.” https://www.who.int/news-room/fact-sheets/detail/v iolence- against- women. Yount, K.M., A.A. Crandall, Y.F. Cheong, T.L. Osypuk, L.M. Bates, R.T. Naved, and S.R. Schuler. 2016. “Child Marriage and Intimate Partner Violence in Rural Bangladesh: A Longitudinal Multilevel Analysis.” Demography 53(6): 1821–52.