Social investment in schooling in low-income countries has increased greatly in the 1990s and 2000s because of the robust associations among schooling and demographic, economic, and health outcomes. This analysis investigates whether targeted school-attendance stipend programs succeeded in reducing gender and socioeconomic inequalities in school attainment among a sample of the rural poor in Bangladesh. Multivariate analyses find that targeted stipend programs helped to reduce the gender attainment gap. Females had an increased probability of participating in stipend programs, and returns to stipend participation were significantly higher for females. However, stipend programs failed to reduce the relative achievement gap between children of different socioeconomic backgrounds: low socioeconomic status (SES) was associated with a decreased probability of stipend participation, and stipend-related schooling gains for lower-SES females were matched by comparable gains for higher–SES females. Meanwhile, there was no significant association between stipend participation and schooling attainment for males.
Improvements in children’s schooling are a central component of long-term development strategies in low-income countries given the robust associations between schooling and improved demographic, economic, and health outcomes (King and Mason 2001; United Nations 2001). The many positive returns to schooling have led to large increases in social investment in schooling, ranging from the elimination of school fees to introduction of mass schooling incentive programs (Adato and Hoddinott 2010; Fiszbein and Schady 2009; Jones 2012). Bangladesh is a good example of a country with historically low rates of schooling that implemented an ambitious series of school reforms between 1990 and the early 2000s (Baulch 2011). In 1990, primary education was made free and compulsory. In 1993, the Food for Education program (FFE) incentivized primary school attendance for targeted vulnerable households by providing wheat and flour, conditional on school attendance. In 2002, FFE was replaced by the Primary Education Stipend program (PES), which provided cash stipends, conditional on school attendance of at least 85 %. In 1994, the Female Secondary School Stipend Project (FSP) incentivized secondary school attendance for rural adolescent girls by providing monthly cash stipends.1 Females in rural Bangladesh historically attained in school at lower rates than males, and the FSP strived to increase female attainment while delaying early marriage and childbearing.
Program evaluations of FFE, PES, and FSP have shown that stipend programs led to large enrollment increases (Ahmed and Arends-Kuenning 2006; Asadullah and Chaudhury 2009; Chowdhury et al. 2002; Schurmann 2009). Nonetheless, questions remain about the extent to which targeted incentive programs lessen the importance of family background in predicting schooling attainment (de la Briere and Rawlings 2006). As school reforms reduce the cost of schooling for households, family socioeconomic status (SES) might become a less important predictor of children’s schooling attainment (Yount et al. 2013). However, it is also plausible that targeted incentive programs do not always reach the poorest populations (Ahmed and del Ninno 2002). Targeted stipend programs may even amplify the benefit of a child having better-educated or wealthier parents who can navigate the education system and provide children with important resources and support. Qualitative evidence suggests that sociocultural and structural factors influence a household’s abilities to participate in conditional cash transfer programs (Adato et al. 2011).
In this analysis, I investigate whether targeted stipend programs reduce gender and socioeconomic inequalities in schooling attainment among a sample of the rural poor in Bangladesh. Stipend programs will reduce inequalities between groups if there are high participation rates among disadvantaged groups and positive returns to participation. In the first part of the analysis, I explore whether gender and socioeconomic disadvantage predicts participation in stipend programs. In the second part of the analysis, I investigate the factors that influence schooling attainment, including stipend participation, gender, and socioeconomic disadvantage.
The Chronic Poverty and Long Term Impact Study in Bangladesh data set was originally collected in 1996, and a follow-up survey of original households was conducted in 2006 (IFPRI-BIDS-IFNS 1998). The original survey included 1,012 households from 47 rural agrarian villages in Manikganj, a district in central Bangladesh; Jessore, a district located in southwestern Bangladesh; and Mymensingh, a district in northeastern Bangladesh. At baseline, in 1996, average per capita household expenditures were about US$200 in the sampled households across the three districts, which was 10 % to 20 % lower than the average annual income for rural Bangladesh (Rahman et al. 1996), meaning that the sample was quite poor even by Bangladesh standards. Nonetheless, within the sample was variation in SES, which is reflected in the differential levels of parental education and wealth (Table 1). This analysis focuses on a subsample of 1,341 school-aged children from 632 of the survey households. Children in this subsample were 18 years or younger at baseline and were school-aged when stipend programs were introduced. Both the 1996 baseline survey and the 2006 follow-up included questions on past and current participation in stipend programs for all school-aged children.
The first part of the analysis uses descriptive statistics and logistic regression to examine the relationship among gender, socioeconomic disadvantage, and participation in stipend programs. Low–SES females should have a higher probability of program participation if programs were successfully targeted to the most disadvantaged groups; thus, I also include an interaction between gender and socioeconomic disadvantage. The second part of the analysis uses ordinary least squares (OLS) regression to investigate the factors that influence schooling attainment, including stipend participation, gender, and socioeconomic disadvantage. Positive returns to stipend participation would indicate that stipend programs are successful in reducing inequalities, conditional on successful targeting of stipend programs. Higher returns to stipend participation for females and for socioeconomically disadvantaged children would mean greater reductions in inequalities; therefore, I look at the interaction between stipend participation and gender, the interaction between stipend participation and socioeconomic disadvantage, and the three-way interaction among stipend participation, gender, and socioeconomic disadvantage.
A dichotomous variable indicates whether the child is female (F = 1) or male (F = 0). In total, 48 % (n = 639) of the sample was female, and 52 % (n = 702) was male.
The analysis focuses on two dimensions of socioeconomic disadvantage: (1) low parental education and (2) low parental wealth.2 The associations between parental education and wealth are explored separately because education and wealth are differentially distributed among the sample and may have differential effects on children’s participation in stipend programs and on their attainment. The simple correlation between low parental education and low parental wealth is 0.08, which suggests that these two variables are largely independent of each other. In total, about 29 % (n = 393) of the sample came from households with both low parental education and low parental wealth.3
Low parental education is measured with an indicator variable that differentiates between the least-educated parents (E = 1), which are households where neither parent completed primary school; and more-educated parents (E = 0), referring to households where at least one parent completed primary school. Completion of primary school represents a high level of formal education for the parents in this sample, with 54 % (n = 725) of the sample having no parents who completed primary school and 46 % (n = 616) having at least one parent who completed primary school (Table 1).
Low parental wealth is measured with an indicator variable that differentiates between the poorest parents (P = 1), referring to households with the bottom 50 % of asset holdings; and less-poor parents (P = 0), referring to households with the top 50 % of asset holdings. I use the total value of assets that both parents brought into marriage to differentiate between the “poorest” half of the sample and the “less poor” half of the sample. Assets at marriage are a useful measure of wealth because marriage is an important time for resource transfers in rural Bangladesh, and assets at marriage are exogenous to decision-making in marriage, including decisions about investment in children’s education (Quisumbing and Maluccio 2003). Assets also represent secure stores of wealth, whereas income is subject to considerable temporal fluctuation (Deaton 1997). In total, about 51 % (n = 679) of the sample had the poorest parents, and 49 % (n = 662) had less-poor parents.
All models also include controls for additional child characteristics: birth order, birth cohort, and district of residence.
Participation in Stipend Programs
A dichotomous variable indicates whether the child participated in any stipend program (P = 1)—including FEP, PES, and FSP—or participated in no stipend programs (P = 0). In total, about 18 % (n = 239) of the children in the sample took part in at least one of the stipend programs at some point up to 2006 (Table 1). The vast majority of stipend participants took part in one stipend program, and a very small number of respondents (n = 3) participated in two stipend programs. If stipend programs were correctly targeted, participants should come from the most socioeconomically disadvantaged households because all the stipend programs were targeted to vulnerable households that tended to be low SES with low parental education and low wealth. The formal definition of “vulnerable households” included landless households; female-headed households; and households relying upon sharecropping, fishing, pottery, weaving, blacksmithing, cobbling, or day labor for income generation (Baulch 2011).
School Attainment Outcomes
School attainment is measured using highest grade passed; respondents are censored at last observation. Some children (n = 87) were still of primary school age at the end-line survey in 2006; thus, there could be downward bias in this outcome if children had not yet completed schooling. Concerns about bias due to censoring are somewhat mitigated by the fact that the average respondent was beyond secondary school age by 2006 (the mean and median ages of the sample in 2006 were 19 and 20, respectively). To account for downward bias due to censoring, I also create a grade-for-age variable that subtracts the average grade attainment of the birth cohort from the grade attainment of the individual. If an individual’s attainment is above, at, or below the cohort mean, their grade-for-age will be positive, zero, or negative, respectively.
Stipend Program Participation
About 18 % (n = 239) of children in the sample participated in a stipend program (Table 1). All three stipend programs intended to reduce gender inequalities in attainment; thus, one would expect a disproportionately large share of participants to be female. In total, approximately 29 % (n = 187) of the female sample participated in stipend programs, and approximately 7 % (n = 52) of the male sample participated in stipend programs (Table 1).4 Eligibility for stipend programs was contingent on vulnerable SES; thus, it might be expected that children of the poorest parents and the least-educated parents would have participated in stipend programs at higher rates. Descriptive statistics indicate that this is not the case: about 49 % (n = 118) of stipend participants had more-educated parents, and about 51 % (n = 121) of stipend participants had less-educated parents; about 52 % (n = 125) of stipend participants had less-poor parents and about 48 % (n = 114) of stipend participants had the poorest parents. Thus, in spite of the fact that targeted programs were intended for the most socioeconomically disadvantaged populations, participation in stipend programs was equally distributed across children from lower and higher socioeconomic backgrounds in this sample of rural Bangladeshis.
I conduct logistic regression analyses of whether gender and socioeconomic disadvantage predicts participation in stipend programs (results are presented as marginal effects at the means). Being female is associated with a 21 percentage point increase in the probability of participation in a stipend program (p < .001) (Table 2, column 1), which is consistent with the fact that stipends were partially targeted to females. On the other hand, I find that having the least-educated parents is associated with a 4 percentage point decrease in the probability of participation (p < .05) (Table 2, column 1), and having the poorest parents is associated with a 5 percentage point decrease in the probability of participation (p < .05) (Table 2, column 1). The significant negative coefficients on the parental schooling and wealth variables are the opposite of what would be expected for a program successfully targeted to the most socioeconomically vulnerable populations.
Socioeconomically disadvantaged females should have a higher probability of program participation if stipends were correctly targeted; thus, I also run models with an interaction between female and least-educated parents and an interaction between female and poorest parents (Table 2, column 2). None of the interaction terms are statistically significant, which means that the relationship between socioeconomic disadvantage and participation in stipend programs does not vary by gender. Nonetheless, the poorest parents variable and the interaction between female and poorest parents are jointly significant (p < .05), and the least-educated parents variable and the interaction between female and least-educated parents are jointly marginally significant (p < .10).
Returns to Participation
I investigate the factors that influence schooling attainment using OLS regression. The first set of models confirms that gender and low parental education are associated with disadvantage in schooling attainment (Table 3, columns 1–2). Being female is associated with completing one less grade of school and having a grade-for-age that is 0.29 lower than males (p < .001) (Table 3, columns 1–2). Having the least-educated parents is associated with completing 1.7 less grades of school and having a grade-for-age that is 0.6 lower than having more-educated parents (p < .001) (Table 3, columns 1–2). There is no significant relationship between having the poorest parents and grade attainment or grade-for-age. Nonetheless, being a stipend recipient is associated with completing 1.3 more grades of schooling and having a grade-for-age that is 0.45 higher than being a nonrecipient (p < .001) (Table 3, Columns 1–2). Thus, stipend programs appear to be succeeding at increasing participants’ attainment.
In the next set of models, I include an interaction between female and stipend participation (Table 3, columns 3–4). In both models, the female coefficient remains negative and significant, confirming that females, on average, attain at lower levels than males. The stipend coefficient is no longer significant, indicating that there is not a significant relationship between stipend participation and attainment for males. There is a significant positive interaction between female and stipend participation for grades attained (2.23, p < .001) and grade-for-age (0.63, p < .001) (Table 3, columns 3–4). Thus, there are significant positive returns to stipend participation for females even though there are no significant returns to stipend participation for males.
In subsequent models, I include two additional interaction terms: (1) the interaction between stipend participation and least-educated parents, and (2) the interaction between stipend participation and poorest parents (Table 3, columns 5–6). Neither of these new interaction terms is statistically significant, although the interaction between female and stipend participation remains statistically significant, and the three interaction terms and the stipend variable are jointly significant in both models (p < .001). Taken together, these results suggest that there are positive returns to stipend participation, but these returns are not significantly different for children from different socioeconomic backgrounds.
In supplementary analyses, I introduced one set of interactions at a time rather than presenting multiple interactions in the same model, and the results were substantively the same. I also looked at the three-way interaction of gender, socioeconomic background, and stipend participation; this term is not statistically significant.
This analysis shows that targeted stipend programs succeeded at reducing the gender attainment gap between males and females in a sample of the rural poor in Bangladesh. The fact that females had an increased probability of participating in targeted stipend programs is unsurprising given that one of the three programs was specifically targeted to females. However, a less expected finding is the significant positive returns to stipend participation for females but not males. This gender discrepancy in the returns to participation could be because the FSP (the only stipend program aimed at increasing secondary school attainment) was restricted to females, whereas the FFE and PES (the stipend programs aimed at increasing primary attainment) were open to males and females.
The findings from this study illustrate that targeted stipend programs face challenges in reaching the most socioeconomically vulnerable populations. Having a lower–SES family was associated with a decreased probability of stipend participation, possibly because higher–SES parents have different preferences about children’s schooling or are better at navigating the school system and gaining access to stipends. Alternatively, this could have to do with how vulnerable status is designated and whether vulnerable status actually corresponds with the most socioeconomically disadvantaged households. Further research should continue to investigate the obstacles poor populations face in accessing social protection programs in order to improve program targeting.
Finally, the analysis demonstrates that stipend programs can improve absolute schooling for populations while failing to reduce relative achievement gaps between children of different socioeconomic backgrounds. In this case, stipend-related schooling gains for lower-SES females were matched by comparable gains for higher-SES females. Thus, stipend programs improved the absolute level of schooling of females in the sample but failed to reduce the attainment gap between females from lower- and higher-SES families. Meanwhile, there was no significant relationship between stipend participation and schooling attainment for males, which suggests a limited impact of stipend programs for male children. Additional research is needed to better understand whether similar patterns hold elsewhere in Bangladesh or in other countries that have introduced targeted stipend programs.
Background support for this study was provided by the grant Team 1000+ Saving Brains: Economic Impact of Poverty-Related Risk Factors for Cognitive Development and Human Capital “0072-03” provided to the Grantee, the Trustees of the University of Pennsylvania by Grand Challenges Canada. I am grateful to Jere Behrman for support and guidance throughout the development of the article. Pat Sharkey, Larry Wu, Florencia Torche, Paula England, Sara Duvisac, Abigail Weitzman, Monica Caudillo, Wahid Quabili, and Agnes Quisumbing provided helpful comments and assistance. I benefited from comments from participants of the American Sociological Association annual meeting, Population Association of America annual meeting, Sociology of Development annual meeting, NYU inequality workshop, and NYU graduate student conference.
A trial version of the FSP was launched by an NGO in 1982 on a small scale in six pilot areas. The program was scaled up to the national level with public financing in 1994.
Research suggests that mother’s and father’s resources may differentially affect children’s outcomes (Quisumbing and Maluccio 2003). I rerun all analyses using gender-disaggregated variables for mother’s and father’s education and wealth; however, I do not find significant differences in the predictive power of mother’s versus father’s resources for the outcomes of interest.
All descriptive percentages are approximate because of rounding.
The Female Secondary School Stipend Program was targeted exclusively to females. However, six male children in the sample reported receiving the Female Secondary School Stipend. This discrepancy could be due to measurement error or erroneous program targeting. I reran all analyses excluding these six respondents, and results were unchanged.