The Impact of Maternal Literacy and Participation Programs: Evidence from a Randomized Evaluation in India Rukmini Banerji, James Berry, and Marc Shotland∗ September 2016

Abstract Using a randomized field experiment in India, we evaluate the effectiveness of adult literacy and parental involvement interventions in improving children’s learning. Households were assigned to receive either (1) adult literacy (language and math) classes for mothers, (2) training for mothers on how to enhance their children’s learning at home, or (3) a combination of the two programs. All three interventions had significant but modest impacts on children’s math scores. The interventions also increased mothers’ test scores in both language and math, as as well a range of other outcomes reflecting greater involvement of mothers in their children’s education.

JEL Classifications: C93, D13, I21, O15 Keywords: adult education, education inputs, field experiments ∗ Banerji:

Pratham and ASER Centre, [email protected]; Berry: Cornell University, [email protected]; Shotland: Abdul Latif Jameel Poverty Action Lab, [email protected]. We thank Annie Duflo for collaboration and insights on the study design and implementation. We also thank David Evans, Eric Edmonds, Marcel Fafchamps, Karthik Muralidharan, Hugo Reis, and seminar participants at the Berlin Social Science Center, Cornell University, Duke University, the NBER Education and Children’s Meeting, the Northeast Universities Development Consortium Conference, UC Santa Cruz, the University of Buffalo, and the World Bank for useful comments and suggestions. We are incredibly grateful to Jessica Chan, Nandini Gupta, Ravi Gupta, Rachna Nag Chowdhuri, and Nikhil Wilmink for superb work coordinating the field activities in Rajasthan and Bihar. Kevin Kho, Laurel Wheeler, and Raymond Lee provided excellent research assistance. This research was funded by the International Initiative for Impact Evaluation. All errors are our own.

1

1

Introduction

Improving the quality of education in the developing world remains a crucial issue for researchers and policymakers alike. In many developing countries, despite gains in educational attainment, learning levels are abysmally low, both when compared with developed countries and with national learning standards (Pritchett, 2013). For example, a 2011 survey in India found that less than half of fifth-grade children could read at a second-grade level (ASER Centre, 2012). In searching for methods to improve levels of learning in developing countries, much of the policy discussion focuses on investments that can be made to improve schools, such as improving school infrastructure, providing additional learning materials, changing pedagogy, or improving teacher selection or incentives. Indeed, there is a growing evidence base of such cost-effective in-school interventions that can improve learning outcomes in developing countries (Kremer et al., 2013). Beyond investments that governments can make in schools, evidence suggests that parents, and in particular mothers, also play an important role in children’s learning—through activities at home as well as through decisions that interact with the formal schooling system. The literature on intergenerational transmission of human capital shows that more educated (or, in some cases, more literate) mothers make decisions that improve their children’s learning (Andrabi, Das, and Kwajha, 2012; Behrman, Foster, and Rosensweig, 1999; Guryan, Hurst, and Kearney, 2008). The direct policy implications from this research imply that investing in the education of today’s children, particularly girls, should have positive spillovers on their children in the future. This literature does not, however, reveal obvious policy solutions for today’s children who live with today’s uneducated parents. There are nearly one billion adults worldwide who are currently illiterate (UNESCO, 2015). To address low levels of adult literacy, a number of developing countries have launched ambitious adult literacy education campaigns.1 For example, in 2009 the government of India launched the Saakshar Bharat adult literacy campaign, with a goal of bringing literacy skills to 70 million adults and reducing the gender gap in literacy. In South Africa, the Kha Re Gude Mass Literacy Programme was launched in 2008, aiming to reach 4.7 million adults by 2012. These campaigns often focus on women in order to address gender inequality in literacy and to promote women’s empowerment (UNESCO, 2015). Although the primary motivation of initiatives like these has been to improve the livelihoods of adult beneficiaries, also listed among the instrumental benefits are improvements in the education of children (Abadzi, 2003; UNESCO, 2005b). Specifically, many of the intermediate outcomes 1 With respect to adult literacy programs, “literacy” typically encompasses reading, writing, and basic mathematics. See UNESCO (2005a) for a discussion of the definition of literacy.

2

through which parents’ educational attainment leads to children’s learning could also be activated by these types of programs. For example, such programs could provide parents with the skills to facilitate children’s learning at home and foster an interest in children’s education. This could lead to increased participation in learning at home, more effective participation, increased availability of educational assets at home, or greater interest in formal schooling. By targeting women, the programs could also influence child education by increasing decision making power of mothers in the household, under the assumption that child education is a higher priority for mothers than for fathers. However, while adult literacy education may provide parents with the skills and interest to participate more in their children’s education, for those who were never formally educated, these programs may not be sufficient to provide them with knowledge of how specifically they can participate. If a key constraint to facilitating children’s learning at home is parents’ lack of understanding of how they can participate, parental “participation training” that provides direct instruction on how they can be involved may prove more effective. For example, parental participation programs could train parents to monitor their children’s completion of homework, or to understand the symbols teachers use to grade homework. Adult literacy and participation programs may also be complementary if parents need skills and the experience of learning, along with instruction on participation, in order to support their children’s learning. Although adult literacy and participation programs could serve as important tools to improve children’s education outcomes in developing countries, there are real challenges that may limit their effectiveness. Adults in developing countries have many competing demands on their time, and without immediate monetary benefits, these programs may not be seen as valuable relative to work, home production, or leisure activities (Wagner, 2000; Abadzi, 2003). Even when adults are motivated to participate regularly, part-time programs may not be extensive enough for formerlyilliterate parents to develop the skills necessary to support their children’s learning. Whether these programs are effective in improving children’s education outcomes is thus an empirical question. To address this question, we conducted a randomized evaluation of three interventions in rural India designed to improve maternal literacy—defined to include reading, writing, and basic mathematics—and mothers’ involvement in children’s education. Villages were randomly assigned to one of four groups. In the first group, mothers in the village were offered the Maternal Literacy (ML) program, consisting of daily language and math classes. In the second, mothers were given the Child Home Activities and Materials Packet (CHAMP) program: materials, activities, and training were provided each week to promote enhanced involvement in their children’s education at home. In the third, mothers were offered both the literacy and home-learning programs (MLCHAMP). The fourth group served as a control with no intervention. The evaluation was carried out in 480 villages in the states of Bihar and Rajasthan. 3

We find that the programs had small positive impacts on children’s learning levels: the ML, CHAMP, and ML-CHAMP interventions resulted in statistically-significant increases in children’s math scores by 0.035, 0.031, and 0.056 standard deviations, respectively. The only significant impacts on language scores came from the combined ML-CHAMP intervention. We also provide evidence that the programs improved mothers’ learning outcomes and indicators of participation and investments associated with improved children’s learning. The three interventions increased mothers’ overall test scores by 0.096, 0.043, and 0.12 standard deviations, respectively. All three programs had impacts on mothers’ participation in their children’s education and on availability of education assets in the household, while the ML and ML-CHAMP programs had impacts on formal school attendance. By contrast, we find no significant impacts on mothers’ time spent directly helping the children with homework. The programs also had limited impacts on mothers’ involvement in household decision making, as the only statistically significant impact arose from the combined ML-CHAMP intervention. The evidence supports our hypothesis that the home learning environment can be improved through these programs and provides suggestive evidence that doing so resulted in increased learning of children. However, a key caveat to this interpretation is the possibility that the programs affected children directly. Indeed, we observed both attendance of children in ML classes and direct participation of children in CHAMP activities. While we cannot rule out the possibility of these direct impacts, we present suggestive evidence that in the case of ML, the impacts we observe on children’s test scores were largely driven by mothers’ participation in ML classes. We also show that changes in mothers’ learning levels, participation in their children’s education, presence of educational assets in the home, and children’s school attendance account for almost all of the impacts of ML on children’s learning. In the case of CHAMP, these intermediate outcomes account for a smaller fraction of the impacts, and therefore direct participation of children could have played a greater role. From a standpoint of policy, we note that the magnitudes of the effects we find on children’s learning are small, and basic cost-benefit calculations suggest that other interventions that target children directly are more cost effective. Thus, the interventions as implemented may not be the most effective policies to improve children’s learning. In the case of ML, we observe relatively low participation and argue that evaluation of innovations to increase participation and sustain learner interest would be a useful area for future research. In addition, while the magnitudes of the impacts may not justify the programs on the basis of children’s learning alone, the effects we find on mothers’ learning suggest that these programs could be promising tools for policy makers interested in influencing adult literacy as well as children’s learning. Our study contributes to several strands of literature. First, it is one of only two randomized evaluations of developing-country adult literacy programs that we know of and the first to study 4

spillovers on children. Early studies find significant effects of adult literacy programs on learning of adults, but these impacts are identified using ex-post comparisons with non-participants (Carron, 1990; Ortega and Rodriguez, 2008). Aker et al. (2011) conduct a randomized evaluation of a program that provided cell phones to participants in existing adult education classes in Niger and find significant impacts of the cell phone program on participants’ writing and math scores. However, there is no evaluation of the adult literacy program, per se. Most recently, Aker and Ksoll (2015) conduct an evaluation of an adult literacy program in Niger that included both a standard adult literacy program and adult literacy plus additional monitoring. The researchers find impacts of approximately 0.2 standard deviations on reading and math scores of adults, with significantly larger effects when monitoring was included. Research on the effects of adult literacy programs on children’s outcomes is even more sparse, and existing studies rely on retrospective selection of comparison groups (Aoki, 2005; Abadzi, 2003).2 Second, our study contributes to the limited literature evaluating parental participation programs in developing countries, where parents have little or no experience with formal education. Kagitcibasi, Sunar, and Bekman (2001) conduct a randomized evaluation of a small-scale program in Turkey that included both a home learning component and training in general parenting skills for mothers of preschool-aged children. The authors find positive impacts on primary school enrollment and academic performance in subsequent years. Several recent studies evaluate programs to encourage parental participation in management of schools, with mixed results on effectiveness (Banerjee et al., 2010; Duflo, Dupas, and Kremer, 2015; Gertler, Patrinos, and Rubio-Codina, 2008).3 In developed country contexts, rigorous evidence on the effects of participation programs is also limited, and the existing evidence is mixed (Avvisati et al., 2010). A recent large-scale randomized evaluation of a program in a poor school district in France that aimed to enhance parental involvement in the education of adolescent children found significant positive effects on parental involvement and on student behavior in school, though there were no detectable effects on test scores (Avvisati et al., 2014). The remainder of our paper is structured as follows. Section 2 discusses the context and programs. Section 3 covers the evaluation design and data collection. Section 4 analyzes program take-up. Section 5 describes the results. We provide discussion of the results in Section 6 and conclude in Section 7. 2 Although

there are numerous evaluations of adult literacy programs in the U.S., much of the research also suffers from methodological limitations (Beder, 1999) . 3 Our study also relates to the small but growing literature evaluating early-childhood stimulation interventions in developing countries which typically focus on development of children aged 0 to 3 (see Baker-Henningham and Boo, 2010 for a review). The interventions we study here differ in that they focus on households with school-aged children.

5

2

Program Description

2.1

Context

The programs in this study were designed and implemented by Pratham, a non-governmental organization (NGO) specializing in child literacy and numeracy.4 Pratham conducted the interventions in two blocks (district subdivisions) of the Purnia district in Bihar and two blocks of the Ajmer district in Rajasthan. Bihar and Rajasthan were selected by Pratham based on the low literacy levels of the two states. According to the 2011 Census of India, these states have the lowest female literacy rates in India (Government of India, 2011). The intervention districts within each state were selected because of existing Pratham programs and infrastructure in those areas. Within the intervention districts, the study blocks were selected because they did not have any pre-existing Pratham programs. In Appendix Table 1 we use data from the 2011 Census of India and 2011 ASER surveys to compare the intervention states, districts, and blocks with our evaluation data (Government of India, 2011; ASER Centre, 2012). According to the Census, households in Rajasthan are more likely to have electricity than those in Bihar (58 percent compared with 10 percent) and have higher ownership rates of costly durable goods such as televisions, mobile phones, and motorcycles. Literacy rates, children’s school attendance, and children’s learning levels are broadly similar between the two states. Within each state, the intervention districts follow similar patterns to the state-wide averages, with slightly lower rates of female literacy and children’s schooling outcomes in both the Rajasthan and Bihar intervention districts compared with the state-wide averages. Characteristics in the intervention blocks are largely similar to those of the broader districts.5 As in the census data, households in our sample displayed generally higher asset ownership in Rajasthan than in Bihar and substantially higher likelihood of having electricity (81 percent compared with 15 percent). Households in our sample have have higher levels of asset ownership than those in the blocks as a whole, which could be due to our selection criteria leading to larger households. We also observe lower levels of school attendance in our sample than the district-level ASER data indicate. This difference could be due to a difference in survey timing: the ASER surveys took place in the fall of 2011, while the surveys for this evaluation took place during the summer vacation period just after the beginning of the school year. Running the interventions in multiple states in different areas of the country aids external validity of the evaluation. Although the interventions were identical in both states, they were implemented by different local teams and supervised by separate state-level Pratham leadership. And 4 For

more information, see http://www.pratham.org. comparisons suggest that Pratham may have been targeting districts with worse education outcomes within states for its programs in general, but the intervention blocks are largely representative of these districts. 5 These

6

while learning levels in both states were similar, the differences in wealth and preexisting activities of the mothers presented distinct implementation challenges in each area.6

2.2 Intervention Descriptions Each of the three interventions was implemented between September 2011 and June 2012. Recruitment of mothers for each program was focused on women with children aged 5 to 8. It was hypothesized that the programs would have the greatest effects on children that were just beginning to develop the most basic reading and math skills. The mothers of these children (referred to as “target mothers”) were identified from a census of communities selected to receive the programs (see Section 3). The Maternal Literacy (ML) intervention consisted of daily literacy classes held in the villages and was modeled after Pratham’s community-based classes for children (see Banerjee et al., 2010). Classes were taught by volunteers who were recruited from within the villages and were given several days of training from Pratham staff before setting up classes. Classes were to be held for two hours per day, six days per week, at a time and place identified by mothers to be the most convenient. While ML classes were open to any who wished to attend, Pratham staff and volunteers were given a list of target mothers to recruit into the classes.Volunteers received several refresher trainings and ongoing supervision by Pratham staff throughout the intervention. Within ML classes, volunteers utilized a version of Pratham’s “Read India” methodology to teach basic language and math skills. This approach, previously shown effective in teaching children to read (Banerjee et al., 2010), was modified to suit the interests of adults. Because most mothers started with very low language and math skills, the programs focused on the most basic competencies. Pratham’s prior experience with adult literacy interventions indicated that mothers were more interested in learning math, and responding to this demand, the program placed more emphasis on basic math skills than on language skills. Math activities included counting, number recognition, addition and subtraction, and activities with currency. The language component provided instruction on letter recognition, reading basic words, and writing names. Each week, volunteers also presented short topical paragraphs to facilitate discussion among mothers. The paragraphs contained messages about decisions mothers could make for themselves or their families, particularly in relation to family health, women’s employment, government programs, and to a lesser extent, the value of educating girls.7 The direct objective of the ML classes was to increase language and math skills of mothers. The weekly discussion topics, as well as the opportunity to meet in groups of women, were designed to promote empowerment, particularly mothers’ ability and confidence to make decisions 6 In 7 Of

Appendix A we explore heterogeneity in the program effects on children’s and mothers’ learning by state. 34 topics, two related to girls’ education.

7

for themselves or for their households. Skills taught in the classes could also increase mothers’ capabilities to make such decisions. Our primary hypothesis was that the ML intervention would improve children’s learning outcomes. This could arise if the skills learned, exposure to the learning process, and opportunity for discussion about education increased mothers’ interest and participation in their children’s education. The intervention could thereby change the type and quality of participation at home, time spent with children, education assets at home, aspirations for children, or schooling behavior outside of the home. We note that while two of the 34 weekly discussion topics related to the benefits of educating girls, there were no specific program components providing instruction on participation in children’s learning. Although mothers were the sole focus of class activities, children were not prohibited from attending class and listening passively or following along with their mothers, and could have thereby directly benefited from attending classes themselves. As we show in Section 4 below, some study children attended both with their mothers and by themselves. We discuss the potential for direct effects on children in Section 6.1. The Child and Mother Activities Packet (CHAMP) intervention was designed to directly encourage mothers to be engaged with their children’s learning, focusing on activities with 5 to 8-year old children in the household. Unlike the volunteers who implemented the ML intervention, CHAMP implementers were paid Pratham staff who often had prior experience as Pratham volunteers or supervisors.8 Pratham decided to use paid implementers primarily because each implementer was responsible for conducting CHAMP in specific households in 5 separate villages, and Pratham believed that unpaid volunteers would not be willing to take on these responsibilities. These implementers were given about one week of training by senior Pratham staff and met together several times per month to review upcoming activities.9 Once per week during the intervention period, the CHAMP implementer was to visit each target mother and gave her a worksheet to help her child complete and instruction on how to be involved in her child’s learning.10 Each worksheet contained several written language or math activities that the mother was instructed to help the child complete in the following week. The activities focused on first-grade language and math skills. For example, some worksheets instructed the child to write the letters of the alphabet or a sequence of numbers. In about half of the sessions, the 8 Hiring staff with prior Pratham experience was not strictly necessary for the program, but Pratham’s practices for recruiting paid staff for new projects often consist of promoting successful volunteers or reallocating existing staff to from previous projects. 9 The staff who implemented CHAMP were also responsible for supervising volunteers in ML. This was primarily for logistical reasons, as the evaluation required multiple interventions to be conducted in the same areas. However, because the ML and CHAMP interventions were distinct, we do not view the multiple roles of these staff as important to our results. 10 When the mother had more than one child aged 5 to 8, she was given multiple worksheets and instructed to do the activities with each child in this age group.

8

mother received direct instruction on encouraging the child to do schoolwork at home, reviewing the child’s school notebooks, and discussing progress with the child’s school teacher. Some visits also included instruction on activities that the mother could do with her child such as sharing stories, telling time, or supervising simple writing exercises. Each weekly visit lasted approximately 10 to 20 minutes. When the mother was not available, the CHAMP session was conducted with another adult member of the household when possible. The primary objective of CHAMP was to improve children’s learning by involving the mother in the process. Mothers were shown how to participate in their children’s learning and were encouraged to spend more time on educational activities with their children. Beyond increasing participation of the mother, the program’s encouragement of home learning could also result in an increase in education assets available at home. Finally, the experience of participating in children’s learning could foster a broader interest in child education, resulting in increased aspirations for children and increased school attendance. In addition to its effects on children, CHAMP could also improve the learning and empowerment of mothers. Although the intervention did not directly provide math or language instruction to mothers, increased exposure to children’s education and interactions with CHAMP implementers could result in increased learning of mothers. Empowerment of mothers could also increase as a result of the program, through increases in learning, or through the experience of participating in their children’s education. As with ML, there was a possibility of direct effects of CHAMP on children beyond the involvement of the mother. CHAMP sessions were typically held at the mothers’ homes, and children could be present during these sessions. In addition, the worksheets provided at each session could have been completed without the involvement of the mother. We present evidence on children’s direct participation in CHAMP in Section 4 below. The combined intervention (ML-CHAMP) included both the ML and CHAMP programs. The combined intervention was not integrated—ML and CHAMP were simply conducted in the same villages with the same target group of mothers. As in the separate interventions, ML classes were conducted by volunteers, and CHAMP visits were conducted by separate Pratham staff. The MLCHAMP intervention was designed to test whether both mothers’ exposure to learning from ML and instruction on how to facilitate learning at home from CHAMP would be necessary for mothers to engage with their children and improve children’s learning. It also allows us to test whether one program “crowded out” another, in which case the impacts would not increase when both programs were implemented together. Although the programs were scheduled to be held from September 2011 through June 2012, the agricultural cycle led to lapses in implementation during this period. Programs were run between September 2011 and February 2012, upon which time the harvest season began in both states. In 9

Bihar, both ML and CHAMP continued with lower intensity in March and April, upon which time they stopped running. In Rajasthan, programs were halted in March and April and began again in May and June.

3

Research Design

3.1

Treatment Assignment

In each state, 240 hamlets (village subdivisions) were selected for the randomization. Hamlets were selected based on a target number of households (the approximate size that could support one maternal literacy class) and geographic distance from other target locations to limit spillovers. In Rajasthan, where villages are geographically distinct, one hamlet per village was randomly selected among those determined to be large enough to support a maternal literacy class. In Bihar, where hamlets may be close to one another (whether in the same village or in different villages), hamlets of the target size were included if they were sufficiently far from other included hamlets.11 For ease of exposition, we refer to the randomization unit as a “village” throughout. In each state, the 240 study villages were randomly assigned in equal proportions to the control group or to one of the three treatment groups. Randomization was stratified geographically to allow Pratham to organize its training and monitoring structure based on a known number of program villages in a fixed area. The 240 villages in each state were first divided into geographically proximate “sets” of 20 villages. Each set was then split into two “phases” consisting of 10 villages each. These phases determined the order of the rollout of the programs. Within each set of 20 villages, the Pratham team first initiated the interventions in Phase 1 villages and began in Phase 2 villages approximately three weeks later.12 These set-by-phase groups of 10 villages formed the strata for the randomization.13 Appendix Figure 1 provides a graphic representation of the stratification procedure.

3.2

Data collection

Within each village in the sample, a census was first conducted to determine a list of target mothers. Twenty-two mothers of children aged 5 to 8 years old were then randomly selected to be targeted. 11 In

Appendix B we provide more detail on how hamlets were selected in each state. from the small difference in implementation timing, Phase 1 villages were identical to Phase 2 villages in program administration and content. 13 Because there were four interventions to be assigned within a stratum of 10 villages, each stratum contained two “remainder” villages. These remainders were balanced within the larger set of 20 villages, such that each intervention was implemented in exactly five villages per set. 12 Apart

10

If there were fewer than 22 such mothers in the village, all mothers were targeted. There were a total of 8,888 target mothers in the study, with an average of 18.5 mothers per village. Baseline data were collected from households of target mothers at the onset of the interventions in summer 2011, and endline data were collected after approximately one year, in summer 2012. The primary evaluation data consisted of standardized tests of mothers and children, as well as household surveys administered to mothers. The standardized tests, designed to evaluate a basic set of Hindi language and math skills, were developed by the ASER Centre, Pratham’s research arm. These tests were expanded versions of the ASER Centre’s standard assessment tool used each year in its national assessments of children’s learning (ASER Centre, 2012). At baseline, the tests were administered to target mothers, children aged 5 to 8 in their households, other children in grades 1 to 4 outside that age range, and children aged 4 and below who were going to be enrolled in school in the next year. The endline testing included all mothers and children tested at baseline, in addition to the remaining children who were aged 3 or 4 at baseline. The mothers’ test contained the same questions as the children’s test, but included several additional word problems and other questions testing practical literacy and numeracy skills. Minor changes were made to the testing instruments between baseline and endline. At baseline a shorter test, designed to quickly assess basic reading and math knowledge, was also administered to other household members. The household surveys were administered to target mothers. These surveys were designed to measure the mothers’ involvement in household decisions, mothers’ interest and participation in their children’s education, and children’s schooling status. Mothers’ interest and participation were measured through modules on participation in learning at home, time use of the mother and child, and presence of educational materials in the household. A number of questions referred a single child aged 5 to 8 in the household. If the household had more than one child in that age range, one was randomly selected to keep the length of the survey manageable.14 The surveys also measured schooling status of all children in the household.15 While most questions appeared in both survey rounds, some additions were made to the endline survey based on observations made during implementation.16 Column 1 of Table 1 contains descriptive statistics from the baseline surveys and tests in the 14 The questions that focused on a single child included the questions on mothers’ and children’s time use, as well as several questions in the sections on participation in children’s education and children’s schooling behavior. We do not find evidence that the focus of the surveys on a single child had impacts on that child’s test scores. See Appendix C for details. 15 In the baseline survey schooling status was measured for children aged 5 to 8. The age range was expanded to 3 to 14 in the endline survey. 16 We also administered short questionnaires to fathers at baseline and endline. These questionnaires contained questions on participation in children’s learning, empowerment of mothers, and support for adult literacy programs. At endline only 57 percent of fathers were reached, and we therefore do not use these surveys in the main analysis.

11

control group. Mothers in our sample had low levels of education and similarly low levels of literacy and numeracy. The average years of education in the control group was 0.76, with a median of zero. Baseline test scores for mothers were relatively low: the tests focused primarily on first-grade competencies, yet the average test score in the control group was 30 percent in the language section and 22 percent in the math section. Columns 2, 3, and 4 of Table 1 present differences in means of the baseline variables between each treatment group and the control group. Two of the 21 variables are jointly significant at the 10 percent level between the three treatment groups and the control group, and one more is significant at the 5 percent level. Out of 63 individual comparisons between treatment and control groups, three are significant at the 10 percent level, and six more are significant at the 5 percent level—slightly more than what would be expected by pure chance, particularly between the ML-CHAMP and control groups. However, many of these significant differences are small in magnitude. For example, there are 0.055 more children aged 9 to 14 in ML-CHAMP compared with the control group mean of 0.98, and the fraction of male children in ML-CHAMP is 0.021 less than the control group mean of 0.52.17 Critically, there are no significant differences between any of the treatment groups and the control group in any of the children’s or mothers’ test score variables, our primary outcomes of interest.18 Approximately 3.5 percent of households reached for surveys and testing at baseline were not reached at endline. Endline child tests are available for 94 percent of children tested at the baseline. We do not observe evidence of differential attrition across treatment groups at the household level, but there is some imbalance of attrition levels among child test-takers between the CHAMP and ML-CHAMP groups and the control group. In Appendix D we present a formal analysis of attrition, including the construction of bounds on the children’s learning estimates using Lee’s (2009) trimming method. Because attrition was low overall, these bounds are relatively tight and do not alter the conclusions from our analysis.

4

Program Take-up

Participation in the ML and CHAMP programs is analyzed in Table 2. For ML, we use data from both the endline survey and Pratham administrative records. Because the endline survey did not ask questions about CHAMP, we rely on Pratham records for measures of CHAMP participation. Panel A.1 presents measures of ML attendance, using self reports from the endline survey. 17 Appendix Table 2 explores robustness of ML-CHAMP impacts on key outcomes to several different combinations

of control variables. The results are virtually unchanged when these different sets of controls are included. 18 Appendix Table 3 presents tests of equality of means between the ML, CHAMP, and ML-CHAMP treatment groups. The number of significant differences at the 10 percent level or less is zero for ML vs. CHAMP, three for ML-CHAMP vs. ML, and one for ML-CHAMP vs. CHAMP.

12

About two-thirds of mothers in both the ML and ML-CHAMP recall an adult literacy class running in the village in the past year. We note that this is somewhat low because according to Pratham, all mothers were approached at least once to be recruited for the classes. Mothers who initially refused to attend or attended only at the beginning may not have recalled these classes, particularly in villages with low attendance among other mothers.19 Self-reported take-up of ML classes is also relatively low: 40 percent of mothers in ML and 44 percent of mothers in ML-CHAMP reported having attended ML classes in the past year.20,21 Children attended the classes as well. The endline survey asked mothers whether and how often the selected child had attended with the mother and whether the child ever attended alone. Nineteen percent of selected children in ML villages and 24 percent of selected children in ML-CHAMP villages were reported to have attended with the mother. Of those, about one third always attended when the mother attended. In addition, about 20 percent of children in ML and ML-CHAMP villages attended alone at least once. Panel A.2 reports ML attendance collected by Pratham volunteers. These data show take-up rates of 76 percent in ML and 84 percent in ML-CHAMP, substantially higher rates than those selfreported by the mothers. There are two potential reasons for this discrepancy. First, because the administrative data were collected by Pratham volunteers and were not verified, they may overstate actual attendance. Second, self-reported attendance may be subject to recall errors. To avoid prompting, enumerators in the endline survey were not told the treatment status of a given village and therefore did not probe for recall of ML attendance in the treatment villages.22 Thus, the survey and administrative data likely provide upper and lower bounds on take-up. To avoid these discrepancies, future research examining these types of programs would benefit from measures of attendance observed directly by the research team. In spite of these differences, however, the 19 Indeed,

among mothers who did not attend, mothers’ knowledge of the classes is strongly related to average takeup in the village: 10 percentage points higher take-up in the village is associated with 2.8 percentage points higher likelihood of a non-attending mother knowing about the adult literacy classes (p-value < 0.01, not shown). 20 The surveys collected self-reports on any participation but not the intensity of participation because it was determined to be logistically infeasible to obtain accurate recall of the number of classes the mother attended over an entire year. It could also lead to surveyors learning the treatment status of the village, which could lead to differential surveyor prompting by treatment status. 21 We note that 7 percent of mothers in the CHAMP and control groups reported attending classes. We find no evidence of spillovers across program hamlets: there is no relationship between attendance in the CHAMP and control group hamlets and presence of ML or ML-CHAMP hamlets in the larger village (see Appendix E). The likely reasons for non-compliance are a combination of misunderstanding of the survey question and the government’s Saakshar Bharat adult literacy mobilization that occurred in Bihar during the spring of 2012. While virtually no Saakshar Bharat classes were actually held after the initial mobilization, initial promotion of the program occurred in many villages across the study area, and our respondents who participated in this mobilization may have considered this as class attendance. 22 There is some evidence supporting recall errors. Among those recorded in the administrative data as having attended, recorded attendance is 69 percent higher (41.2 days vs 24.3) for those who also reported attending in the endline survey. Within this group, the number of months since the last recorded attendance is also significantly related to to a mother not reporting attending in the survey (results not shown).

13

correlates of take-up are similar across self-reported and administrative measures (see Appendix F). Although the administrative data do show a relatively high percentage of mothers attending at least one class, the number of classes attended was low. The average targeted mother in ML and ML-CHAMP attended 25 and 27 classes over the course of the year, respectively. Attendance was also sporadic throughout the year: among those who took up the classes, attendance between September 2011 and February 2012 averaged between 3 and 5 days per month. As noted above, classes were held less consistently from March through the end of the program in June due to the onset of the harvest. Taken together, these measures indicate relatively low participation in ML classes. According to Pratham staff, much of the low take-up was driven by mothers’ interest and availability, although availability of volunteers did impact whether classes were held. While classes were scheduled to be held six days per week (about 24 days per month), the administrative data indicate that 12.5 classes were held on average per month in ML and ML-CHAMP between September 2011 and February 2012. Some of the “missing classes” were likely holidays: the peak festival period in India falls between October and December. However, many could be the result of a lack of volunteer availability or effort (which could itself be an endogenous response to low interest of mothers). Nonetheless, when classes were held, the average selected mother only attended about 25 percent of the time, indicating that the low participation was primarily due to choices of mothers. As shown in Table 2, there is slightly higher likelihood of attendance in the ML-CHAMP group compared with ML: self-reported take-up was 4.4 percentage points higher, and take-up in the administrative data was 6.7 percentage points higher (both significant at the 5 percent level).23 However, we note that the magnitudes of the differences are relatively small, representing 11 and 9 percent higher attendance in ML-CHAMP for the self-reported and administrative data, respectively. In addition, the administrative data do not indicate a significant difference in the number of classes attended between the two groups. CHAMP was a door-to-door intervention where Pratham staff conducted weekly visits with each mother at home. If the mother was not at home, then the CHAMP session was to be conducted with another adult in the household. Panel B of Table 2 presents take-up of CHAMP according to Pratham administrative data. Ninety-nine percent of target households were visited at least once during the intervention period. Households were successfully visited about 16 times on average. As with ML classes, the majority of CHAMP visits were held between September 2011 and February 2012, with fewer visits taking place between March and the end of the program in 23 When

we split the sample by state, only Rajasthan exhibits significantly higher take-up in ML-CHAMP (results not shown). We speculate that this could be driven in part by the lower overall participation of mothers in ML classes in Rajasthan, as described below. Mothers in Rajasthan could have been closer to the margin of participating and were therefore more responsive to the additional contact with Pratham staff.

14

June. During the second half of the CHAMP program, attendance data in both Rajasthan and Bihar included information on the mother’s and child’s participation in CHAMP activities. Within this subsample, mothers were present during 81 percent of successful CHAMP visits. In about half of cases, implementers also recorded children’s presence during CHAMP sessions and self-reports of who in the household helped the child with the previous week’s worksheet. These data show that a target-aged child was present during 58 percent of CHAMP visits. The mother helped the child with the CHAMP worksheet 53 percent of the time. In 19 percent of cases another household member helped with the worksheet, while the child completed the worksheet alone 28 percent of the time. In Appendix F we present a detailed analysis of the correlates of mothers’ participation in ML and CHAMP. For the ML intervention, several relationships suggest factors that could explain the low take-up discussed above. First, take-up is positively associated with slightly higher baseline test scores of mothers and a small amount of formal education. This highlights a potential challenge in targeting of adult literacy programs: even though they may be meant to reach adults with no education, a small amount of education may signal both experience with classroom learning and an interest in learning more. Second, experience with literacy classes in the past and experience with meeting in groups (as measured through self-help group membership)24 are also important determinants of the take-up decision. Thus, some mothers may not be comfortable in the group learning environments typical in many adult literacy programs. Third, we observe higher selfreported take-up in Bihar, where Pratham staff reported that mothers were more available to attend than in Rajasthan. This result is echoed by our qualitative interviews, which highlighted a lack of free time as a key constraint to attendance.25 Almost all households participated in CHAMP, and hence there is little variation in the extensive margin of participation. Examining the intensity of participation, a higher fraction of children in school and higher baseline child test scores are strong determinants of the percentage of sessions attended. This could reflect a complementarity between the CHAMP material and the school curriculum. In contrast with the determinants of ML participation, intensity of participation in CHAMP has a small negative relationship with mothers’ education and no relationship with mothers’ test scores. 24 A

self-help group is a group of villagers that pools savings and provides loans to members of the group. spent in household and market work was 43 percent higher Rajasthan than in Bihar. We note, however, that conditional on state, we do not find a significant relationship between time spent on household and market work and participation. On the other hand, availability may also be driven by the opportunity cost of that time, or the value of leisure. 25 Time

15

5

Results

Before the endline data were examined, a pre-analysis plan (PAP) was uploaded to the American Economic Association’s Randomized-Controlled-Trial Registry website.26 The PAP includes the weights used in aggregating test scores, the estimating equation for test scores, and the list of intermediate outcomes to be examined. The analysis presented in this paper follows the PAP as closely as possible. In the few cases where we deviate from the PAP, we discuss the choices and implications below.

5.1

Estimating Equation

Throughout the analysis we utilize the following estimating equation: Y1iv = β0 + β1 MLv + β2CHAMPv + β3 MLCHAMPv + β4Y0iv + δ Gv + πHiv + εiv

(1)

In this equation, Y1iv is the outcome for individual i in village v. MLv , CHAMPv , and MLCHAMPv are dummy variables indicating the treatment status of the village. Y0iv is the baseline value of the outcome of interest (when measured). Gv is a dummy for stratum, as described in Section 3.1, while Hiv represents a vector of household-level characteristics. εiv is the individual error term, clustered by village, the level of randomization. Equation (1) is estimated using ordinary least squares. For all impact results using child-level data, we use all variables listed in in Table 1 as controls. When the mother is the unit of observation, we use all household variables in Table 1, Panel A, and household-level averages of the child-level variables in Panel B. This includes two variables that were not originally listed in the PAP: child gender and child schooling status. We include these because we find some imbalance in these variables across treatment groups, as shown in Table 1, and because child schooling status is a key outcome of interest. Excluding these variables from the set of controls does not substantively affect any of the main results presented below (not shown). We present our results in three parts. First, we present effects on children’s test scores. Second, we examine mothers’ learning and involvement in household decisions. Finally, we estimate program impacts on mothers’ participation in their children’s education, home education assets, and formal schooling. 26 https://www.socialscienceregistry.org/trials/65

16

5.2

Children’s Test Scores

Table 3 presents the effects of the treatment groups on children’s language, math, and total (math plus language) test scores. We estimate Equation (1), controlling for baseline math and language test scores and the full set of variables presented in Table 1. For each of the three test score categories, scores are normalized based on the control group mean and standard deviation in each round of testing. As specified in the PAP, all children tested at the endline are included in the estimation, including the younger children not tested at baseline. We code missing control variables as zero and include dummy variables to indicate missing values. Column 1 of Table 3 displays the impacts of the ML intervention on language, math, and combined test scores. The intervention had an impact of 0.035 standard deviations on math scores (significant at the 5 percent level). The estimated effects on language and combined scores are smaller and not statistically significant. As shown in Column 2, the CHAMP intervention had a similar impact on math scores of of 0.032 standard deviations (significant at the 5 percent level). The effect on language scores is smaller and not statistically significant, but the effect on total test scores equals 0.024 and is statistically significant at the 10 percent level. In Column 3 we display the impacts of the combined ML-CHAMP intervention. The estimated effects are 0.042, 0.056, and 0.052 standard deviations for language, math, and total test scores (respectively), and all three estimates are significant at the 1 percent level.27 For each category of test score, the ML-CHAMP impacts are larger than the individual ML or CHAMP impacts, and we only reject that the sum of the ML and CHAMP impacts equals the impacts of ML-CHAMP in the case of language, where ML-CHAMP had an impact 0.037 standard deviations larger than the sum of the ML and CHAMP interventions (p-value = 0.085). We note, however, that these tests are relatively low powered: for example, the sum of the ML and CHAMP impacts on math scores has a 95 confidence interval between 0.013 and 0.12. To understand how the interventions impacted specific skills, Appendix Table 5 presents effects on a subset of the competencies covered on the test. With few exceptions, these competencies correspond to those tested in the ASER national assessments of reading and language (ASER Centre, 2013).28 We focus on math, where the main results show significant impacts of all three interventions. Among the individual math skills, the ML intervention had similar impacts across the range of competencies, with the largest impacts on two-digit addition. The CHAMP and MLCHAMP interventions had the largest impacts on one-digit number recognition. 27 In Appendix Table 4, we evaluate the robustness of the children’s test score results to two alternative specifications. The first set of regressions limits the control variables to baseline test scores and stratum dummies, and the second drops observations where no baseline test scores were measured. The results are virtually unchanged under these alternative specifications. 28 The primary difference between the competencies examined and those on the ASER test is that we include two digit addition, while the ASER test includes division.

17

5.3

Mothers’ Test Scores and Empowerment

Table 4 presents the effect of the interventions on mothers’ normalized math, language, and total test scores. We again estimate Equation (1), controlling for baseline language and math scores as well as the control variables listed in Table 1. All three programs had statistically significant impacts on mothers’ language, math, and total test scores. The ML program improved mothers’ test scores by 0.066 standard deviations in language, 0.12 standard deviations in math, and 0.096 standard deviations overall. All three estimates are significant at the 1 percent level. As shown in Column 2, The CHAMP program improved mothers’ test scores by 0.023 standard deviations in language (significant at the 5 percent level), 0.059 standard deviations in math (significant at the 1 percent level), and 0.043 standard deviations overall (significant at the 1 percent level). As shown in Column 3, the ML-CHAMP intervention had impacts of 0.088, 0.15, and 0.12 standard deviations on language, math, and total test scores, respectively (each significant at the 1 percent level).29 As with the results for children, the impacts of the combined intervention on mothers’ test scores are larger than those of either individual intervention. Indeed, we cannot reject that the effects are additive for any of the three test score categories. In addition, these tests are comparatively well-powered: for example, the 95 percent confidence interval on the sum of the ML and CHAMP math impacts is between 0.12 and 0.23 standard deviations, well above the impact of either individual intervention and relatively close to the point estimate of the ML-CHAMP impact. In Appendix Table 6 we examine the effects on a subset of competencies on the test, similar to the exercise for children in Appendix Table 5. In language, ML and ML-CHAMP had the largest impact on letter recognition, the most basic skill tested. CHAMP, by contrast, had impacts of similar magnitude across a broader range of competencies, although these impacts are considerably smaller than the ML and ML-CHAMP impacts on letter recognition. In math, all three interventions had the largest impacts on one-digit number recognition. The difference in the pattern of impacts between ML and CHAMP in language could be due to the fact that children were typically exposed to higher level competencies at school than mothers were exposed to in the ML classes, and therefore the CHAMP intervention, which encouraged mother-child interaction, gave mothers more exposure to those higher-level competencies. We next examine the impacts of the interventions on women’s empowerment. As described in Section 2, either the ML or CHAMP intervention could promote empowerment by providing mothers with the ability and confidence to make decisions for themselves or for their households. Our main empowerment analysis focuses on the mother’s involvement in household decisions vis-à-vis 29 Appendix Table 4 examines robustness of these results to exclusion of household-level control variables. As in the case of the results on children, running the impact regression controlling only for baseline test scores leaves the results virtually unchanged.

18

her husband.30,31 This dimension is among the most commonly used in studies of empowerment (Malhotra et al., 2002), and we hypothesized that it could be a channel through which the programs could impact children’s outcomes. We include 9 indicators of the mother’s involvement in various types of household purchases, as well as involvement in child schooling decisions. These measures are aggregated into an index using the method specified in Kling, Liebman, and Katz (2007). To construct the index, we compute the normalized value of each variable based on the control group mean and standard deviation at endline. We then take the average of the non-missing normalized variables and re-normalize the resulting index based on the control group mean and standard deviation. The final index has a mean of zero and standard deviation of one in the control group.32 Table 5 displays the impact estimates of the interventions on the index and the component variables. Among the three treatment groups, only ML-CHAMP had a significant impact on the full set of measures, with a point estimate of 0.091 standard deviations (significant at the 1 percent level). The estimates for both ML and CHAMP are positive, smaller in magnitude and not statistically significant.33 One interpretation of the pattern of estimates is that ML and CHAMP indeed had small impacts on mothers’ involvement in household decisions, as the sum of the point estimates of the individual interventions is almost exactly equal to the point estimate of the combined intervention.34 However, this is speculative as we cannot reject that the impact of either individual program is zero. Thus, while women’s empowerment has been a policy outcome of interest for adult literacy programs in particular, we find limited evidence for impacts on our measure of involvement in household decisions. This case is similar to that of group-based microcredit: while these programs 30 Enumerators

were instructed to isolate the mother from other household members when asking questions on empowerment. This is a similar procedure to that followed in the Demographic and Health Surveys, a set of largescale international surveys commonly used in studies of gender (ICF International, 2016). 31 We note that empowerment and the intermediate outcomes in the analysis are based on self-reports from the household survey. While we cannot directly verify the responses to these questions, the patterns of impacts we find are inconsistent with surveyor demand effects that would influence responses on all questions related to education, or only on questions relating to program content. We discuss this issue in detail in Appendix G. 32 The variables representing mother involvement in household decisions are a subset of variables originally listed in the PAP. We have narrowed the original list of 26 variables in the PAP to reflect those most frequently used in the literature. We analyze the full set of variables from the PAP in Appendix I. The additional variables include mothers’ mobility, capabilities to perform market transactions, beliefs and attitudes about female education, and happiness. When all 26 variables are included in an aggregate empowerment index, we find impacts of ML of 0.071 standard deviations (significant at the 5 percent level) and impacts of 0.14 standard deviations (significant at the 1 percent level). 33 The results are qualitatively similar if the components of the index are restricted to the 5 variables not related to children’s schooling decisions (not shown). 34 Results from the endline father questionnaire suggest that these impacts arose without increases in intrahousehold conflict. Eighty-nine percent of fathers indicated that they supported adult literacy classes for their wives, with significantly higher proportions of support in each of the the intervention groups compared with the control group (not shown). We note, however, that the response rate of this questionnaire was only 57 percent.

19

have been credited with increasing women’s empowerment, recent rigorous evaluations have found either small positive results on similar measures of involvement in household decisions (Angelucci et al., 2015) or no evidence for impacts (Banerjee et al., 2015; Crépon et al., 2015; Tarozzi et al., 2015).

5.4

Mothers’ Participation, Home Inputs, and Child Schooling

As described in Section 2, we hypothesized that the programs would impact mothers’ interest and involvement in their children’s education. In this section we present results on mothers’ perceptions of their role in child education, their participation in child education, mothers’ and children’s time use, education assets in the home, and children’s school participation. We measured mothers’ perception of their role in childhood education by asking whether they believe they should be responsible for child education at all (relative to fathers), and what activities they could do to assist their children. A detailed analysis is presented in Appendix H. While most mothers believed they have a role to play in child education, mothers in the ML, CHAMP, and MLCHAMP villages were 4.1, 3.0, and 4.0 percentage points more likely to report being responsible for child education, respectively (each significant at 5 percent or less). We also asked an openended question on what the mothers could do to help children in their studies, and we counted the total number of responses.35 Relative to a control group mean of 1.9 activities, mothers in ML and ML-CHAMP listed 0.069 and 0.094 more activities (significant at the 10 and 5 levels, respectively). Turning to actual participation in children’s education, our measures consist of survey questions on mothers’ involvement in the education of the selected 5 to 8-year-old child, focusing on monitoring and supervision of the child’s learning. We include nine measures of participation, including indicators of school visits, whether the mother helps with homework, whether she has looked at her child’s notebook recently, and the frequency with which she talks to her children and others about children’s education.36 As with the empowerment measures, we construct an index of mothers’ participation using the nine measures, following the aggregation procedure described above. As shown in Table 6, we find positive and statistically significant impacts of each program on the mothers’ participation index. The magnitudes are 0.074 standard deviations for ML (signif35 Lists

of potential activities were given to the surveyor, but not shared with mothers. If a mother included the activity in her response, the surveyor checked it off from that list. 36 These measures correspond to those listed in the PAP, with two exceptions. First, the question “What have your child and family members taught you?” was included in the list of participation measures in the PAP, but it refers to household participation in the mother’s learning and is therefore not appropriate for the index. Second, the PAP included educational assets in the participation measures, and we have chosen to analyze impacts on these assets separately.

20

icant at the 5 percent level), 0.12 for CHAMP (significant at the 1 percent level), and 0.11 for ML-CHAMP (significant at the 1 percent level). While both ML and CHAMP had statistically significant impacts on the mother examining the child’s notebook, talking to the child about school, and talking to others about the child’s studies, CHAMP and ML-CHAMP also had significant impacts on mothers helping with homework.37 The impact of CHAMP on the participation index is 62 percent higher than the impact of ML (although the difference is not statistically significant). The larger impact of CHAMP is likely due to the CHAMP program’s direct encouragement of participation. In addition, the sum of the ML and CHAMP effects is almost twice as large as the ML-CHAMP impact, and in contrast with the results on test scores, we can reject the null hypothesis that the program effects are additive (p-value = 0.049). In this case, the programs may have influenced the participation of similar groups of mothers in similar ways, and one program simply crowded out the other when they were combined. We next turn to the effects of the programs on mother’s and children’s time use. To the extent that the interventions encouraged participation in children’s learning at home, we hypothesized that they would increase mothers’ time spent directly helping children learn and child time spent learning at home. However, we find limited evidence of these effects. Panel A of Table 7 presents impact estimates of the programs on mothers’ weekly time spent with children and working inside and outside of the household.38 Across all measures, we see little evidence that the programs impacted time use. Time spent helping children with homework increased by about 0.1 hours for each of the three interventions, but these estimates are not statistically significant. Panel B of Table 7 presents the estimated impacts of the programs on weekly time use of the randomly-selected 5 to 8-year-old child in the household. As with the results for mothers, we find few significant impacts across the range of measures. There is suggestive evidence that the CHAMP program resulted in modest increases in time spent on homework: the impact was 0.13 hours per week for CHAMP and 0.21 hours per week for ML-CHAMP, but neither is statistically significant at conventional levels.39 37 We

note that the recall period of the survey questions overlapped with the end of the CHAMP intervention in Rajasthan, and thus some of the CHAMP results could be driven by activities undertaken by the mother during the intervention. However, the impact of the CHAMP program on the participation index in Bihar only is 0.092 standard deviations and is significant at the 10 percent level (not shown). 38 The measures presented here are a subset of those listed in the PAP. Appendix Tables 7 and 8 present impacts on all of the time use variables listed in the PAP for mothers and children, respectively. 39 The time-use results also provide suggestive evidence that ML and ML-CHAMP increased work hours by both the mother and child. ML-CHAMP increased paid, agricultural and livestock work of the mother by 1.6 hours per week (significant at the 10 percent level), and effect of both ML is 1.2 hours per week, falling just short of statistical significance at conventional levels. ML-CHAMP also increased the child’s time spent working in an household business by 0.30 hours (significant at the 5 percent level). We speculate that the skills taught and positive messages about employment in the maternal literacy classes could have given mothers more confidence and skills to work independently and to manage the child’s time working. Child labor could also have increased as a result of the the increase

21

Table 8 presents estimated treatment effects on the presence of education assets in the home, as reported by the mother.40 These assets include pencils, school books, other books, newspapers and magazines, and slates. The first row of Table 8 displays the impacts on an index of the five types of assets, where the index is constructed in the manner described above. The ML, CHAMP, and ML-CHAMP interventions had significant impacts on the index, with magnitudes of 0.057, 0.081, and 0.089 standard deviations. Among the individual asset types, the CHAMP intervention had statistically significant effects on the presence of pencils and newspapers/magazines, while ML-CHAMP significantly increased the presence of school books, other books, and slates. For the ML intervention, the only statistically significant effect is on the presence of school books. We note that although neither intervention provided any of these materials for home use, the ML program did provide small workbooks to mothers. Because the presence of the education assets were self reported we cannot rule out the possibility that some mothers confused these workbooks for school books. When school books are excluded from the index, the impact of ML falls to 0.046 and is no longer statistically significant (p-value = 0.14). CHAMP provided mothers with individual worksheets, which are not easily confused with the asset items. CHAMP also did not directly encourage purchase of any of these items. Finally, Table 9 presents the impacts of the programs on school participation. Panel A presents impacts on school enrollment and attendance for the sample of children tested at endline.41 MLCHAMP increased current or planned enrollment in the upcoming school year by 1.3 percentage points (significant at the 5 percent level) and attendance in school or aanganwadi (communitybased kindergarten) by 2.0 percentage points (significant at the 10 percent level).42 We also find impacts of about 2 percentage points of ML and ML-CHAMP on current attendance in formal school (significant at the 5 and 10 percent levels, respectively). Because the impacts on formal school attendance are larger than the impacts on school or aanganwadi attendance, ML and MLCHAMP may have encouraged mothers to shift their children from aanganwadi to first grade at earlier ages. Indeed, as shown in Panel B, the impacts on formal school attendance are nearly double for ML and ML-CHAMP when the sample is restricted to children around the age of initial enrollment in school (ages 4 to 6 at baseline). By contrast, we find little evidence for impacts on private school attendance, school days missed, or monthly tuition expenditures. These results are in child skills. However, given the lack of statistical significance of many of these estimates, and the large number of coefficients estimated on time use, these results may also be due to chance. 40 When the mother was unsure that a particular item fell into one of the categories, the surveyor verified it visually. 41 Panel A of Table 10 differs from the PAP in two ways. First, in order to more closely follow the test score results we have restricted our sample to children tested at endline (the majority of whom were aged 3 to 8 at baseline), while the PAP measured school participation for all children aged 3 to 14. We find no evidence for school participation on this broader group (results not shown). Second, we include participation in formal school (excluding kindergarten), which was not included in the PAP. 42 School attendance and aanganwadi attendance were measured through the questions “Does the child go to school or aanganwadi?” and “What grade does the child attend?”

22

supported by our qualitative interviews, in which many parents claimed that the most they could do within the formal schooling system was send their children to school and cited poor quality of government schools and poor access to private schools as barriers they could not overcome.

6 6.1

Discussion Children’s Attendance in Maternal Literacy Classes

As discussed in Section 2, the theory of change of the ML intervention was that mothers’ attendance in the classes would lead to changes in mothers’ decisions and behaviors, ultimately leading to improved children’s learning. However, as we document in Section 4, children attended the ML classes in addition to mothers. In this section we present suggestive evidence that attendance of mothers may have served as a more important mediator of impacts compared with children’s attendance. We focus on impacts on math scores, where we find statistically significant impacts of the ML intervention.43 We first analyze the potential impacts of mothers’ and children’s attendance by exploiting the design of the experiment. As described in Section 3, the treatment assignment was stratified geographically for ease of implementation. Sets of 20 geographically proximate villages were selected, and these sets were randomly split into two phases of 10 villages each to determine the order of program rollout. The randomization was stratified by these 48 set-by-phase groups of 10 villages. We thus have 48 “separate” randomized experiments within the larger experiment of 480 villages. In addition to geographic differences and slight differences in the timing of program initiation, each Pratham supervisor was assigned to a set of 20 villages, potentially giving rise to implementation differences across strata. These implementation differences—such as in the timing of classes and in the of recruitment of mothers—could give rise to different patterns of take-up of mothers and children in each strata. To the extent that these take-up patterns influenced program impacts, we can examine the relationship between mothers’ or children’s take-up in each stratum and the learning impacts on children. For each stratum, we estimate 48 separate ML treatment effects on mothers’ attendance, children’s attendance, and children’s test scores using the specification in Equation (1).44 For a consistent measures of attendance for both mothers and children across treatment and control groups, we rely on self-reported attendance from the endline survey. We then correlate the treatment effects on children’s math scores with the effects on mother’s or children’s attendance. The results are plotted in Appendix Figure 2. There is a clear positive relationship between the stratum-wise ML 43 As shown in section 4, some children also participated directly in CHAMP. Because we have limited data on the extent and nature of this direct participation, however, we focus the analysis on children’s attendance in ML classes. 44 This is implemented by interacting the treatment groups with dummies for the 48 strata.

23

treatment effects on children’s math scores and mothers’ attendance (ρ = 0.31, p-value = 0.02), but not children’s attendance (ρ = −0.087, p-value = 0.55). These relationships are, of course, correlations, but they do suggest that attendance of mothers, rather than attendance of children, is most closely related to children’s test scores. We next analyze the potential effect of attendance by including attendance of mothers and children as regressors in the main children’s learning specification from Section 5.2. Inclusion of endogenous intermediate outcome variables in an impact estimation can provide suggestive evidence for the influence of these variables as mediators of impacts (Baron and Kinney, 1986). When the intermediate outcome variable is included, a decrease in the coefficient of the treatment variable and a positive coefficient on the intermediate variable suggests that this variable is a mediator of the treatment effect. However, this rests on several strong assumptions (Green et al., 2010; Imai et al., 2011). In particular, the analysis assumes that the intermediate variable is uncorrelated with other factors that influence the final outcome variable. If, for example, the intermediate variable is positively correlated with these unobserved factors, then the variable may appear to be a mediator due to omitted variable bias. Alternatively, if the intermediate variable is measured with error, then the analysis will tend to underestimate the effect of mediation.45 Table 10 presents the results of the analysis, including combinations of mothers’ and children’s attendance variables in the regression of children’s test scores on the treatment groups. Again, we rely on self-reported attendance from the endline survey, as it provide consistent measures across treatment and control groups. Because we only have attendance data for one 5 to 8-year-old child per household, we restrict the sample to those children, and hence the impacts using the main specification from Section 5.2 are noisier and fall just short of statistical significance at conventional levels (p-value = 0.12). Columns 2 and 3 include dummies for attendance of either the mother or child in the regressions. In each specification, the coefficient on the ML treatment decreases by about one-third, and the coefficient on the attendance dummy is about 0.04 (both significant at the 5 percent level). As shown in Column 4, when we include both variables in the same regression, the ML treatment variable decreases by 43 percent. There is a slightly larger coefficient on attendance of the child compared with that the mother (0.028 vs 0.020), but neither is statistically significant. Finally, Column 5 includes both attendance of the mother and two richer measures of attendance of the child–attendance with the mother weighted by frequency46 and attendance without the mother. In this regression, we observe a 39 percent decrease in the coefficient on ML. The coefficient on the mother’s attendance (0.028) is now larger than that of the child’s attendance with the mother (0.019) but again neither is statistically significant. The coefficient on the variable indicating the child’s attendance without the mother is effectively zero. In slight contrast with the stratum-wise 45 See 46 The

Green et al. (2010) and Imai et al. (2011) for a more formal discussion of these issues. weighted variable equals 1 for “always”, 0.67 for “sometimes,” and 0.33 for “rarely.”

24

analysis above, the results in this table suggest that attendance of the mother and the child could have had similar mediating effects on the children’s learning results.47 The slightly different results of the the two analyses presented in this section are likely due to the use of different sources of attendance in each one. When we isolate variation in attendance due to geographic and implementation factors, there is no evidence that children’s attendance mediated the impacts. When we use variation at the individual level by including attendance measures in our learning regressions, the analysis suggests that both could have served as mediators. Nonetheless, the results suggest that while we cannot rule out the possibility that children’s participation in ML influenced the learning results, much of the impact appears to be mediated by participation of the mothers.

6.2

Mediation of Impacts Through Intermediate Outcomes

The ML, CHAMP, and combined ML-CHAMP programs influenced a number of intermediate outcomes that could have led to the observed impacts on children’s learning. In the case of ML, we observe significant impacts on mothers’ average learning levels, increased participation in their children’s education as measured through our participation index, but not an increase in time spent directly helping their children with homework. We also find impacts on education assets available in the home and on attendance in formal schooling. CHAMP had smaller impacts on mothers’ learning, but larger impacts on mothers’ participation and education assets. In Table 11 we present suggestive evidence of the importance of these intermediate outcomes by including these measures as regressors in the estimation of impacts on children’s test scores, as was done in the previous section with measures of attendance. As before, we focus on math outcomes, where we observed significant impacts from all three interventions. We also restrict our sample to the one selected 5 to 8-year-old child per household to parallel the analysis in the previous section, and because some of the survey questions in the participation index focused on those children. As in Table 10, each column of Table 11 includes a separate variable or variables representing the intermediate outcomes, with the first column displaying the results using the specification from Section 5.2 for reference. In all cases, we include baseline values of the intermediate outcomes as controls, and thus the displayed coefficients on the endline variables can be interpreted as their influence relative to baseline. Column 2 includes endline mothers’ math and language scores. Inclusion of these variables decreases coefficient on the ML treatment dummy by 43 percent and the impact of CHAMP by 12 percent. One standard deviation higher math (language) score is 47 It

is also important to note that attendance of the child only appears effective when the child attended with the mother rather than alone, suggesting that to the extent that children did learn through attendance in ML, it only occurred alongside their mothers.

25

associated with a higher child test score of 0.073 (0.053) standard deviations (both significant at the 1 percent level). As shown in Column 3, inclusion of the participation index reduces the coefficient on the ML treatment by 13 percent and the CHAMP coefficient by 11 percent. We observe similar reductions in the coefficients when the education asset index is included. When we include formal school attendance, the coefficient on ML drops 34 percent, and the coefficient on CHAMP drops 3 percent. As shown in Column 6, when all intermediate variables are included in the regression, the coefficient on ML decreases 82 percent, and the coefficient on CHAMP decreases 22 percent. In this regression, the mothers’ test score variables, the asset index, and the school attendance variables have about the same magnitude as in the individual regressions and remain statistically significant, while the coefficient on the participation index becomes closer to zero and is not statistically significant. Finally, in Column 7 we include attendance of the mother and child in the ML classes to the specification in Column 6. Inclusion of these variables has virtually no influence on the coefficients on the other intermediate outcomes. Although we emphasize that this analysis rests on strong assumptions and must be taken as suggestive, the patterns from the analysis highlights several potential pathways for impact. First, our index of mothers’ participation in children’s learning appears to be a modest mediator of the learning impacts, while mothers’ learning appears much stronger, both in the regressions where each measure is included individually and in the regressions where all mediators are included together. This could suggest that mothers’ learning could be influencing the quality of interactions beyond what is measured in the participation index, which primarily measures the different types of involvement of the mother. Alternatively, it could suggest that the participation index is measured with error, in that it does not fully reflect all of the potential activities the mother could do with her child. Second, the analysis suggests that formal school attendance is a strong mediator of impacts for the ML intervention. This implies that ML could have increased interest in formal schooling, and mothers responded by ensuring attendance in school—a margin they considered particularly feasible, according to our qualitative interviews. Third, when all intermediate outcomes are included together, they account for almost all of the ML impact, but only about a quarter of CHAMP impact. This could imply more potential for direct impacts of CHAMP on the child beyond the involvement of the mother, but it is also possible that our measures are simply not picking up the precise changes in the home environment that led to increased children’s learning. In addition, the specification of Column 7 shows that the influence of these intermediate outcomes is not confounded by children’s attendance in ML, as the coefficients do not change when the attendance of the child is included in the regression. This provides further evidence to the finding of the previous section that the children’s learning results in ML were largely driven by mothers’ attendance. Nonetheless, we emphasize these analyses presented should be viewed as suggestive, and research designs explicitly testing the influence of specific intermediate outcomes 26

as well as mothers’ and children’s attendance in these programs would be a fruitful area for future work.

6.3

Cost Effectiveness

This section presents a discussion of cost effectiveness of the three interventions in improving children’s and mothers’ test scores. As with most cost-effectiveness comparisons across studies, we note that differences in target population, competencies tested, testing instruments, local prices, and methods of calculating costs may limit comparability. Our cost-benefit calculation follows the methodology in Kremer, et. al (2013). Details of the cost-benefit calculation are provided in Appendix Table 9. While both interventions included a similar amount of Pratham staff time–either supervising the volunteers and classes in ML or running the sessions in CHAMP–the cost of the ML intervention also included the opportunity costs of volunteer time. This difference leads to much higher cost per beneficiary household in ML ($34 compared with $17 per household in CHAMP). For children, each $100 spent on ML, CHAMP, and ML-CHAMP resulted in improvements in math scores of 0.22, 0.40, and 0.24 standard deviations, respectively. The greater cost-effectiveness of CHAMP is driven by the lower costs of the intervention, as both ML and CHAMP had similar treatment effects on children. For mothers, each $100 spent on the ML, CHAMP, and ML-CHAMP interventions resulted in improvements in mothers’ total test scores of 0.28, 0.26, and 0.24 standard deviations, respectively. The similarity in cost effectiveness between the ML and CHAMP interventions is due to the fact that while ML was about twice as expensive, the effects of ML on mothers were twice as large as the effects of CHAMP. Because there are no other cost-benefit calculations of program effects on adult learning that we know of, we focus our comparison with the literature on learning impacts on children. Our impact estimates, ranging 0.03 to 0.06 standard deviations, fall below the range of statistically-significant estimates in the Kremer et al. (2013) study. This latter set of estimates ranges from 0.14 standard deviations to 0.6 standard deviations. Out of the 15 studies in Kremer et al. (2013) that found statistically significant impacts, 14 are also more cost effective than the interventions studied here. Even though the interventions studied here may be less cost effective than others in improving child test scores, our study examines and finds impacts on a broad set of outcomes, including mothers’ learning—a key policy objective of adult literacy programs. However, given the lack of data for other adult literacy programs, we are unable to perform comparisons along these additional dimensions. Our study should serve as a starting point for future work to expand the evidence base so that more explicit cost-effectiveness comparisons can be made in the future.

27

7

Conclusion

In developing countries, adult literacy programs have been credited with improving children’s education outcomes by encouraging parental investments in education and an improved home learning environment. Programs that encourage parental involvement have also been cited as a means of improving children’s education through similar channels. In spite of the potential of these programs to achieve these objectives, there has been little rigorous evidence on whether they are actually effective. Our paper evaluates the impacts of three variants of these programs, each targeting mothers in rural India: an adult literacy program, a participation program, and a combination of the two. We show that each of these three programs had modest impacts on children’s learning levels, particularly math scores. The programs also influenced several intermediate outcomes related to decisions of mothers and the home learning environment that could in turn have affected children’s learning: all three had impacts on mothers’ learning levels, interest in participation in child education, reported levels of participation, and the presence of educational materials at home. By contrast, we find no significant impacts on time spent directly helping children with homework. This suggests that child outcomes could have been influenced by the quality of time spent, or frequency or type of interactions, even if there was no increase in time spent helping with homework. We also find that the maternal literacy and combined programs increased children’s attendance in formal schooling. A more direct mechanism could be that children learned by participating directly in the programs, independent of how the mother was affected. Indeed, children did attend ML classes and were often present during CHAMP sessions. If children’s participation was found to be the sole mechanism through which children learned, it would call in to question a key assumption in the programs’ theory of change—that the decisions of mothers and the home environment were integral inputs into their children’s learning. Based on the wishes of the partner NGO, it was not feasible to prevent children from attending ML classes or to prevent direct involvement of children in CHAMP. Our experimental design thus does not allow us to rule out the influence of direct participation of children on the impacts on children’s test scores. However, in the case of ML classes, we provide suggestive evidence that attendance of children was likely not the primary mechanism for the children’s learning impacts. We also show that for the ML intervention, improvements in mothers’ learning levels, participation in their children’s education, presence of educational materials at home, and children’s school attendance can almost entirely account for the impacts we observe on children’s learning. Of these, improvement in mothers’ learning itself appears to be the largest mediator of impacts. For the CHAMP intervention, the analysis shows that the intermediate outcomes were smaller mediators of children’s learning,

28

suggesting a greater role of direct children’s participation or household involvement that was not measured in our surveys. We conclude that although we cannot fully isolate the specific channels through which the programs influenced children’s learning, the evidence is generally consistent with our hypothesis that the programs would influence children’s learning through changes in mothers’ decisions and the household learning environment—mechanisms previously identified in studies of the intergenerational transmission of human capital. However, we note that while we are able to show that similar mechanisms can be activated by adult literacy and participation programs, we cannot speak directly to the precise mechanisms at play in the broader literature. Because of the relatively low magnitude of the impacts on children’s learning, the interventions we study are not cost effective relative to interventions that have been studied elsewhere. The low effect sizes were likely a result of low take-up. Most mothers reported never attending the classes, and those who did, attended infrequently. Our experience with low attendance of mothers echoes prior work highlighting learner interest as a key impediment to the success of adult literacy interventions in developing countries (Wagner, 2000; Abadzi, 2003). The pattern of impacts across the three interventions suggests scope for larger impacts with more intensive interventions. In our key results on children’s and mothers’ test scores, the point estimates for ML-CHAMP are approximately equal to the sum of the ML and CHAMP impacts. Based on these results, we speculate that innovations to the ML model that better sustain learner interest could have larger impacts than what we have found here. Evaluation of such strategies would be an important area for future research. We also note that this paper does not fully analyze the opportunity costs of these programs, or whether they are optimal from the perspective of every household. This paper can serve as a starting point for future work examining the welfare implications of these programs.

29

References Helen Abadzi. Adult literacy: A review of implementation experience. The World Bank Operations Evaluation Department, 2003. Jenny C. Aker and Christopher Ksoll. Call me educated: Evidence from a mobile monitoring experiment in niger. Working Paper, Tufts University., 2015. Jenny C. Aker, Christopher Ksoll, and Travis J. Lybbert. Can mobile phones improve learning? Evidence from a field experiment in Niger. American Economic Journal: Applied Economics, 4 (4):94–120, 2012. Tahir Andrabi, Jishnu Das, and Asim Ijaz Khwaja. What did you do all day? Maternal education and child outcomes. Journal of Human Resources, 47(4):873–912, 2012. Manuela Angelucci, Dean Karlan, and Jonathan Zinman. Microcredit impacts: Evidence from a randomized microcredit program placement experiment by Compartamos Banco. (7):151–182, 2015. Aya Aoki. Assessing learning achievements and development impact: Ghana’s national functional literacy program. Australian Journal of Adult Learning, 45, November 2005. ASER Centre. Annual status of education report 2011, 2012. URL http://www.asercentre. org/p/79.html. Accessed August 17, 2016. Francesco Avvisati, Bruno Besbas, and Nina Guyon. Parental involvement in school: A literature review. Economeie De L’Education, 120(5):759–778, 2010. Francesco Avvisati, Marc Gurgand, Nina Guyon, and Eric Maurin. Getting parents involved: A field experiment in deprived schools. The Review of Economic Studies, 81(1):57–83, 2014. Helen Baker-Henningham and Florencia Lopez Boo. Early childhood stimulation interventions in developing countries: A comprehensive literature review. IZA Discussion Paper No. 5282, 2010. Abhijit Banerjee, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. The miracle of microfinance? Evidence from a randomized evaluation. (7):22–53, 2015. Abhijit V. Banerjee, Rukmini Banerji, Esther Duflo, Rachel Glennerster, and Stuti Khemani. Pitfalls of participatory programs: Evidence from a randomized evaluation in education in India. American Economic Journal: Economic Policy, 2(1):1–30, 2010. Reuben M. Baron and David A. Kinney. The moderator-mediator variable distinction in social psychological research: Conceptual, strategic, and statistical considerations. (51):1173–1182, 1986. Hal Beder. The outcomes and impacts of adult literacy education in the United States. National Center for the Study of Adult Learning and Literacy Report 6, (6), 1999.

30

Jere Behrman, Andrew Foster, Mark Rosenzweig, and Prem Vashishtha. Women’s schooling, home teaching, and economic growth. The Journal of Political Economy, 107(4):682–714, 1999. G. Carron. The functioning and effects of the Kenya literacy program. African Studies Review, 33 (3):97–120, 1990. Bruno Crépon, Florencia Devoto, Esther Duflo, and WIlliam Parienté. Estimating the impact of microcredit on those who take it up: Evidence from a randomized experiment in Morocco. (7): 123–150, 2015. Esther Duflo, Pascaline Dupas, and Michael Kremer. School governance, teacher incentives and pupil-teacher ratios: Experimental evidence from kenyan primary schools. Journal of Public Economics, 123:92–110, 2015. Paul Gertler, Harry Anthony Patrinos, and Rubio-Codina. Empowering parents to improve education: Evidence from rural mexico. World Bank Policy Research Working Paper No. 3935, 2008. Government of India. Census of india 2011: Population enumeration data, 2011. URL http:// www.censusindia.gov.in/2011census/population_enumeration.html. Accessed July 5, 2016. Donald P. Green, Shang E. Ha, and John G. Bullock. Enough already about "black box" experiments: Studying meidation is more difficult than most scholars suppose. (628):200–208, 2010. Jonathan Guryan, Erik Hurst, and Melissa Kearney. Parental education and parental time with children. Journal of Economic Perspectives, 22(3):23–46, 2008. ICF International. The dhs program - protecting the privacy of dhs survey respondents. URL http://dhsprogram.com/What-We-Do/ Protecting-the-Privacy-of-DHS-Survey-Respondents.cfm. Accessed June 20, 2016. Kosuke Imai, Luke Keele, Dustin Tingley, and Teppei Yamamoto. Unpacking the black box of causality: Learning about causal mechanisms from experimental and observational studies. (105):765–789, 2011. Cigdem Kagitcibaci, Diane Sunar, and Sevda Bekman. Long-term effects of early intervention: Turkish low-income mothers and children. (22):333–361, 2001. Jeffrey R. Kling, Jeffrey B. Liebman, and Lawrence F. Katz. Experimental analysis of neighborhood effects. Econometrica, 75 (1):83–119, 2007. Michael Kremer, Conner Brannen, and Rachel Glennerster. The challenge of education and learning in the developing world. Science, 340(6130):297–300, 2013. David S. Lee. Training, wages, and sample selection: Estimating sharp bounds on treatment effects. Review of Economic Studies, 76:1071–1102, 2009.

31

Daniel Ortega and Francisco Rodriguez. Freed from illiteracy? A closer look at Venezuela’s Mision Robinson literacy campaign. Economic Development and Cultural Change, 57(1):1–30, 2008. Lant Pritchett. The Rebirth of Education: Schooling Ain’t Learning. Center for Global Development, 2013. Alessandro Tarozzi, Jaikishan Desai, and Kristin Johnson. The impacts of microcredit: Evidence from Ethiopia. (7):54–89, 2015. UNESCO. Aspects of Literacy Assessment: Topics and Issues from the UNESCO Expert Meeting. 2005a. UNESCO. Education for All Global Monitoring Report 2006: Literacy for Life. 2005b. UNESCO. Education for All Global Monitoring Report 2015: Education for All 2000-2015: Achievements and Challenges. 2015. Daniel A. Wagner. EFA 2000 thematic study on literacy and adult education. International Literacy Institute, 2000.

32

Table 1: Randomization Check Mean Relative to Control Control (1) Panel A. Household-level Variables First principal component of -0.0341 durables ownership [2.250] Main source of income: farming 0.494 [0.500] Number of children 0-4 0.929 [0.900] Number of children 5-8 1.438 [0.602] Number of children 9-14 0.976 [0.936] Number of children 15-17 0.265 [0.500] Number of adults 18 and over 2.914 [1.500] Fraction of household (15 and over) 0.389 that can read [0.426] Fraction of household (15 and over) 0.246 that can do math [0.376] Mother's education (years) 0.764 [2.282] Father's education (years) 3.876 [4.438] Mother's age 32.29 [7.102] Mother has past experience with 0.111 literacy classes [0.307] Mother's language score (fraction) 0.299 [0.247] Mother's math score (fraction) 0.215 [0.241] Mother's total score (fraction) 0.250 [0.234]

ML (2) 0.00735 (0.0859) 0.0137 (0.0222) -0.0112 (0.0298) -0.0300* (0.0174) 0.0560* (0.0289) 0.0130 (0.0152) 0.0263 (0.0569) -0.0137 (0.0173) 0.00415 (0.0154) 0.0475 (0.102) -0.150 (0.203) -0.253 (0.254) -0.00477 (0.0119) 0.00442 (0.0110) 0.00639 (0.0110) 0.00557 (0.0108)

33

P-value: CHAMP ML-CHAMP All coeffs. 0 (3) (4) (5) 0.0939 (0.0944) 0.00796 (0.0243) 0.0179 (0.0332) -0.0265 (0.0191) 0.0282 (0.0299) 0.0125 (0.0156) 0.127** (0.0610) 0.0189 (0.0189) 0.0284 (0.0181) 0.152 (0.118) 0.133 (0.226) -0.193 (0.273) -0.00489 (0.0126) 0.0134 (0.0131) 0.0106 (0.0129) 0.0117 (0.0128)

0.0306 (0.0899) 0.0526** (0.0218) 0.0675** (0.0297) -0.0139 (0.0171) 0.0549** (0.0271) 0.00232 (0.0154) 0.117** (0.0562) 0.00887 (0.0183) 0.0254 (0.0160) 0.0694 (0.103) 0.234 (0.213) -0.232 (0.254) -0.0179 (0.0120) 0.00548 (0.0115) 0.00598 (0.0115) 0.00577 (0.0113)

N (6)

0.747

8888

0.064

8885

0.037

8888

0.332

8888

0.138

8888

0.769

8888

0.071

8888

0.365

7576

0.220

7579

0.625

8864

0.290

8181

0.764

8888

0.436

8659

0.785

8857

0.860

8857

0.830

8857

Table 1: Randomization Check (continued) Mean Relative to Control P-value: CHAMP ML-CHAMP All coeffs. 0 (3) (4) (5)

Control ML N (1) (2) (6) B. Child-Level Variables (tested children) Child is male 0.520 -0.0132 -0.00781 -0.0213** 0.231 15502 [0.500] (0.0108) (0.0110) (0.0106) Child attends school / aanganwadi 0.801 0.00766 0.0250* 0.0149 0.277 15501 [0.399] (0.0136) (0.0135) (0.0137) Child's language score (fraction) 0.350 0.00753 0.00912 0.0103 0.786 15502 [0.296] (0.0110) (0.0113) (0.0109) Child's math score (fraction) 0.277 0.0114 0.0131 0.00857 0.669 15502 [0.303] (0.0111) (0.0117) (0.0109) Child's total score (fraction) 0.310 0.00971 0.0113 0.00934 0.726 15502 [0.291] (0.0108) (0.0113) (0.0106) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display the differences in means between each treatment group and the control group. Column 5 displays the pvalue of the F-test that the differences in means between the treatment groups and control group are jointly zero. Differences in means are computed by OLS regression, controlling for stratum dummies. First principal component of durables ownership is computed based on a set of variables indicating ownership of table/chair, watch/clock, mobile phone, landline, bullock cart, radio, generator, refrigerator, television, electric fan, car/truck, scooter/motorcycle, bicycle, jugaad (a basic motorized transport vehicle), plough, thresher, tractor, and harvester; the number of cows/buffaloes, oxen/bullocks, sheep/goats, hens/ducks, pigs, pigeons, and other livestock owned; an indicator of a cement/stone/metal/beam/plastic roof; and the number of rooms in the house. Father education variable has fewer observations due to absence of fathers from some households. Fraction of household that can read or do math is constructed as the average number of non-mother members 15 years old and over who passed the basic reading or math test, as described in Section 3.2 of the text, ignoring missing values. These variables have fewer observations due to missing test scores for some household members. "Aanganwadi" is community-based kindergarten. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

34

Table 2: Take-up of ML and CHAMP Mean Control (1) Panel A. Maternal Literacy A.1 Endline Survey Mother knew about ML classes

ML (2)

Diff ML-CHAMP CHAMP ML-CHAMP - ML / CHAMP (3) (4) (5)

0.220 0.628 [0.414] [0.483]

0.214 [0.410]

0.675 [0.469]

0.0438** (0.0218)

Mother attended ML class

0.071 0.398 [0.257] [0.490]

0.067 [0.250]

0.436 [0.496]

0.0440** (0.0221)

Child attended ML class (with mother or alone)

0.050 0.270 [0.219] [0.444]

0.040 [0.196]

0.324 [0.468]

0.0546*** (0.0203)

Child attended with mother

0.026 0.192 [0.159] [0.394]

0.025 [0.156]

0.242 [0.428]

0.0521*** (0.0174)

Child attended with mother: Always

0.007 0.058 [0.083] [0.234]

0.005 [0.072]

0.091 [0.288]

0.0308*** (0.00847)

Child attended with mother: Sometimes / Rarely

0.018 0.132 [0.134] [0.339]

0.019 [0.136]

0.151 [0.358]

0.0202 (0.0128)

0.036 0.183 [0.187] [0.386]

0.028 [0.165]

0.213 [0.410]

0.0316* (0.0162)

Child attended alone A.2 Attendance Records Mother attended ML class

--

0.764 [0.425]

--

0.844 [0.363]

0.0674** (0.0265)

Days attended

--

24.658 [31.327]

--

27.255 [30.685]

2.680 (1.919)

Percent of classes attended

--

0.248 [0.259]

--

0.255 [0.241]

0.00752 (0.0121)

--

--

0.990 [0.098]

0.995 [0.070]

0.00640** (0.00273)

Number of successful visits

--

--

15.789 [5.281]

15.480 [5.224]

-0.302 (0.301)

Percent of visits successful

--

--

Panel B. CHAMP Attendance Records Household participated

0.843 0.849 0.00726 [0.212] [0.197] (0.00721) Notes: Columns 1-4 present raw means and standard deviations (in square brackets) of the indicated participation measure in each group. Column 5 presents regression-adjusted differences and standard errors (in parentheses) between the mean in the ML-CHAMP treatment and the ML treatment (Panel A) or CHAMP treatment (Panel B). Regressions in Column 5 control for stratum, and standard errors are clustered at the village level. "Child" refers to one randomly selected child per household aged 5-8 at baseline. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

35

36

0.0322** (0.0155)

0.0118 (0.0146) 0.0558*** (0.0152)

0.0424*** (0.0144) 0.851

0.227 0.236

0.006 0.609

0.085 18283

18283

N (7)

0.0166 0.0238* 0.0515*** 0.625 0.032 0.588 18283 (0.0150) (0.0142) (0.0136) Notes: Standard errors in parentheses. Columns 1, 2, and 3 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for all variables in Table 1. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 4 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 5 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 6 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. Test scores are normalized based on the control group means and standard deviations for each category of score and in each round of testing. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.0351** (0.0162)

Math

Total

-0.00679 (0.0158)

Language

ML (1)

Table 3: Children's Test Scores Impact of Treatment in Endline P-value: P-Value: P-value: CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive (2) (3) (4) (5) (6)

37

0.0592*** (0.0147)

0.0233** (0.0119) 0.149*** (0.0163)

0.0884*** (0.0130) 0.000

0.001 0.000

0.000 0.258

0.945 8580

8580

N (7)

0.0957*** 0.0433*** 0.124*** 0.000 0.000 0.434 8580 (0.0124) (0.0113) (0.0130) Notes: Standard errors in parentheses. Columns 1, 2, and 3 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 4 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 5 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 6 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. Test scores are normalized based on the control group means and standard deviations for each category of score and in each round of testing. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.117*** (0.0160)

Math

Total

0.0663*** (0.0125)

Language

ML (1)

Table 4: Mothers' Test Scores Impact of Treatment in Endline P-value: P-Value: P-value: CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive (2) (3) (4) (5) (6)

38

0.544 [0.498] 0.458 [0.498] 0.245 [0.430] 0.482 [0.500] 2437.8 [2258.8] 0.493 [0.500] 0.499 [0.500] 0.399 [0.490] 0.422 [0.494]

Involved in purchasing utensils

Involved in purchasing cot

Involved in purchasing cycle

Involved in purchasing educational materials

Value of goods can buy alone

Involved in deciding girl school enrollment

Involved in deciding girl school type

Involved in deciding boy school enrollment

Involved in deciding boy school type

0.00562 (0.0157)

0.00414 (0.0153)

0.0199 (0.0166)

0.0228 (0.0170)

111.4 (71.67)

0.0240 (0.0171)

0.00559 (0.0138)

0.00464 (0.0172)

0.0325* (0.0171)

0.0438 (0.0340)

0.0264* (0.0150)

0.0186 (0.0149)

0.0215 (0.0166)

0.0203 (0.0163)

35.24 (72.11)

0.0281* (0.0168)

-0.0232* (0.0124)

-0.00830 (0.0167)

0.0330* (0.0177)

0.0375 (0.0316)

0.0322* (0.0165)

0.0265* (0.0155)

0.0508*** (0.0164)

0.0399** (0.0175)

178.2** (73.83)

0.0521*** (0.0174)

0.00249 (0.0130)

0.0134 (0.0175)

0.0426** (0.0182)

0.0905*** (0.0347)

0.181

0.374

0.924

0.882

0.293

0.809

0.031

0.423

0.977

0.852

0.245

0.422

0.143

0.511

0.174

0.237

0.045

0.422

0.831

0.283

0.992 8211

0.871 8238

0.702 7757

0.899 7821

0.766 8581

0.997 8549

0.283 8494

0.484 8500

0.374 8546

0.851 8581

Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 5 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 6 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 7 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. The "Decisions Index" is an normalized average of z-scores of the other variables in the table, using the control group means and standard deviations. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.000 [1.000]

Decisions Index

Table 5: Mothers' Involvement in Household Decisions Impact of Treatment in Endline Endline Mean P-value: P-Value: P-value: Control ML CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive N (1) (2) (3) (4) (5) (6) (7) (8)

39

0.321 [1.125] 0.157 [0.364] 0.0598 [0.237] 0.762 [0.426] 0.789 [0.408] 0.218 [0.413] 2.994 [3.049]

Takes child to school (times/week)

Visited school in past 3 months

Reason for most recent visit was child's learning

Knows how often child receives homework

Helps child with homework

Opened child's notebook in past 2 months

Talks to children about school (times/week)

0.203** (0.0925)

0.0325** (0.0133)

0.00426 (0.0124)

0.00938 (0.0108)

0.00633 (0.00833)

0.00988 (0.0120)

-0.0380 (0.0375)

0.0739** (0.0309)

0.203** (0.0936)

0.0617*** (0.0132)

0.0286** (0.0123)

0.00614 (0.0107)

0.0147* (0.00819)

0.0132 (0.0124)

0.0338 (0.0415)

0.120*** (0.0325)

0.261*** (0.0961)

0.0467*** (0.0130)

0.0218* (0.0120)

0.00484 (0.0112)

0.00837 (0.00790)

0.0160 (0.0122)

-0.00294 (0.0356)

0.105*** (0.0329)

0.998

0.030

0.042

0.747

0.368

0.798

0.057

0.125

0.753

0.095

0.114

0.905

0.641

0.892

0.159

0.283

0.256

0.014

0.527

0.474

0.307

0.694

0.982

0.049

8525

8572

8572

8480

8480

8480

8480

8573

Talks to friends/family about child's 1.504 0.255*** 0.168** 0.246*** 0.217 0.405 0.087 8520 education (times/week) [2.248] (0.0721) (0.0722) (0.0762) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 5 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 6 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 7 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. The "Mothers' Participation Index" is a normalized average of z-scores of the other variables in the table, using the control group means and standard deviations. The baseline participation index only includes indicators for which data were collected. "Reason for most recent visit was child's learning" indicates whether the most recent visit to the child's school was to discuss child studies with school staff, attend a parent-teacher meeting, or to inquire or complain about lack of teacher effort. "Child" refers to one randomly selected child per household aged 5-8 at baseline. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.000 [1.000]

Mothers' Participation Index

Table 6: Mothers' Participation in Children's Education Impact of Treatment in Endline Endline Mean P-value: P-Value: P-value: Control ML CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive N (1) (2) (3) (4) (5) (6) (7) (8)

40

19.02 [8.008] 4.646 [3.679]

Housework

Look after children

1.886 [4.153]

Household business

0.207 (0.136)

0.0838 (0.133)

-0.0376 (0.143)

-0.175* (0.105)

0.264 (0.299)

1.217 (0.810)

0.0339 (0.131)

0.142 (0.131)

0.126 (0.133)

-0.125 (0.107)

0.112 (0.297)

-0.0313 (0.773)

-0.00384 -0.0377 (0.0391) (0.0356)

0.0940 0.109 (0.0899) (0.0931)

0.298** (0.137)

0.174 (0.133)

0.207 (0.136)

0.113 (0.104)

0.242 (0.288)

1.610* (0.831)

-0.00489 (0.0401)

0.0642 (0.0894)

0.226

0.661

0.236

0.627

0.612

0.122

0.350

0.867

0.162

0.799

0.226

0.009

0.859

0.112

0.560

0.871

0.769

0.782

0.550

0.005

0.753

0.718

0.508

0.269

8447

8449

7892

8581

8581

8555

8454

8533

Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and variables listed in Table 1. missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 5 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 6 displays the p-value of F-test that the impacts of ML, CHAMP, and ML-CHAMP are equal. Column 7 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. Time spent in each activity measured in hours per week. Child time use measured for one randomly-selected child aged 5-8 in the household. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

3.599 [4.110]

Housework

4.059 [4.331]

47.77 [25.12]

Paid / Agricultural / Livestock work

Panel B. Children Homework

0.320 [1.371]

2.308 [2.704]

Read with children

Panel A. Mothers Help with homework

Table 7: Mothers' and Children's Weekly Time Use Endline Mean Impact of Treatment in Endline P-value: P-Value: P-value: Control ML CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive N (1) (2) (3) (4) (5) (6) (7) (8)

41

Table 8: Home Education Assets Impact of Treatment in Endline P-value: P-Value: P-value: ML=CHAMP All coeffs. eq. Additive Control ML CHAMP ML-CHAMP N (1) (2) (3) (4) (5) (6) (7) (8) Education Asset Index 0.000 0.0571* 0.0808** 0.0893*** 0.433 0.566 0.292 8581 [1.000] (0.0311) (0.0315) (0.0323) Has pen / pencil 0.945 -0.000338 0.0126* 0.00784 0.052 0.151 0.671 8581 [0.227] (0.00794) (0.00680) (0.00746) Has school books 0.906 0.0168** 0.0112 0.0201** 0.471 0.542 0.500 8581 [0.292] (0.00845) (0.00825) (0.00863) Has other books / comics 0.251 0.0131 0.0235 0.0284** 0.483 0.531 0.693 8581 [0.434] (0.0131) (0.0150) (0.0139) Has newspaper / magazine 0.0561 0.00851 0.0244*** 0.00342 0.049 0.036 0.011 8581 [0.230] (0.00736) (0.00809) (0.00735) Has slate 0.889 0.0148 -0.00654 0.0227** 0.037 0.013 0.320 8581 [0.314] (0.00996) (0.0101) (0.00955) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 5 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 6 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 7 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. The "Education Asset Index" is a normalized average of z-scores of the other variables in the table, using the control group means and standard deviations. The baseline index only includes indicators for which data were collected. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01. Endline Mean

42

0.0790 [0.270]

Child attends private school

-0.135 (0.203)

-0.0298 (0.208)

-0.0726 (0.0737)

0.0881* (0.0516)

0.0610 (0.0536) 0.0177 (0.0698)

0.00578 (0.0164)

0.00202 (0.0125)

-0.00221 (0.00834)

0.00612 (0.0106)

-0.00120 (0.0105)

0.00390 (0.00535)

0.0409*** (0.0151)

-0.000349 (0.0130)

0.00852 (0.00844)

0.0248** (0.0102)

-0.000321 (0.0110)

0.00681 (0.00549)

-0.175 (0.191)

-0.0379 (0.0676)

0.0794 (0.0548)

0.0344** (0.0167)

0.0209* (0.0123)

0.00265 (0.00865)

0.0198* (0.0111)

0.0203* (0.0106)

0.0131** (0.00527)

0.632

0.214

0.584

0.024

0.845

0.207

0.075

0.931

0.571

0.790

0.447

0.858

0.066

0.136

0.449

0.191

0.056

0.186

0.973

0.870

0.362

0.588

0.269

0.767

0.476

0.144

0.757

Table 9: Children's Schooling Impact of Treatment in Endline P-value: P-Value: P-value: ML CHAMP ML-CHAMP ML=CHAMP All coeffs. eq. Additive (2) (3) (4) (5) (6) (7)

7393

6993

8480

8165

8165

18065

18283

18283

18080

N (8)

19.87 2.830* 2.552 0.294 0.869 0.193 0.032 8457 [43.88] (1.578) (1.748) (1.466) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum, baseline values (where available), all variables in Table 1. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Column 5 displays the p-value of F-test that the impacts of ML and CHAMP are equal. Column 6 displays the p-value of the F-test that the ML, CHAMP, and ML-CHAMP impacts are equal. Column 7 displays the p-value of F-test that the impact of ML-CHAMP equals the sum of the ML and CHAMP impacts. "Selected child" refers to one randomly selected child per household aged 5-8 at baseline. School and aanganwadi attendance were measured through the mother's answer to the questions “Does the child go to school or aanganwadi?” and “What grade does the child attend?” In Panel C, days of school missed are coded as missing for children not in school. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

2.195 [5.419]

Days missed due to housework or other work in past month

Monthly tuition fees

1.379 [1.939]

Days of school missed in past week

0.691 [0.462] C. Selected children (aged 5-8 at baseline) Hours per day spent in school 4.073 [1.618]

Child attends school

0.887 [0.317]

0.728 [0.445]

Child attends school

B. Children aged 4-6 at baseline Child attends school / aanganwadi

0.882 [0.323]

0.941 [0.235]

Child attends school / aanganwadi

A. Children tested at endline Child is enrolled or will be enrolled next school year

Control (1)

Endline Mean

Table 10. Children's Math Impacts Controlling for ML Attendance Dependent Variable: Child's Math Score Included Attendance Variable(s)

Main (1)

Mother Attended (2)

Child Attended (3)

Mother Attended and Child Attended (4)

ML

0.0337 (0.0214)

0.0212 (0.0219)

0.0227 (0.0217)

0.0191 (0.0219)

0.0206 (0.0219)

CHAMP

0.0501** (0.0208)

0.0492** (0.0207)

0.0493** (0.0207)

0.0492** (0.0207)

0.0489** (0.0207)

ML-CHAMP

0.0546*** (0.0197)

0.0403* (0.0209)

0.0417** (0.0202)

0.0379* (0.0209)

0.0395* (0.0209)

0.0195 (0.0191)

0.0275 (0.0195)

Mother Attended

0.0356** (0.0162)

Child Attended

0.0406** (0.0171)

Mother Attended and Child Attended (5)

0.0282 (0.0200)

Child Attended with Mother (weighted)

0.0192 (0.0324)

Child Attended Alone

-0.00443 (0.0223)

R-Squared 0.709 0.709 0.709 0.709 0.709 N 8367 8311 8262 8262 8241 Notes: This table presents regressions using the specification from Table 3, including self-reported measures of mother's and child's attendance measured in the endline survey as regressors. The sample consists of one randomly selected child aged 5 to 8 in the household. Child's attendance with mother is weighted by frequency of attendance: "Always" = 1, "Sometimes" = 0.67, and "Rarely" = 0.33. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

43

Table 11. Children's Math Impacts Controlling for Intermediate Outcomes Dependent Variable: Child's Math Score Included Intermediate Variable(s)

Main (1)

Mother's Participation Education Learning Index Asset Index (2) (3) (4)

Attends School (5)

Intermediate All with Intermediate Attendance (6) (7)

ML

0.0337 (0.0214)

0.0192 (0.0214)

0.0294 (0.0214)

0.0288 (0.0212)

0.0223 (0.0209)

0.00611 (0.0208)

0.000516 (0.0214)

CHAMP

0.0501** 0.0439** (0.0208) (0.0208)

0.0446** (0.0207)

0.0437** (0.0206)

0.0484** (0.0202)

0.0391* (0.0201)

0.0394* (0.0201)

ML-CHAMP

0.0546*** 0.0379* (0.0197) (0.0198)

0.0495** (0.0197)

0.0483** (0.0195)

0.0538*** (0.0188)

0.0353* (0.0188)

0.0288 (0.0198)

Mother's language score

0.0528*** (0.0178)

0.0473*** (0.0175)

0.0475*** (0.0177)

Mother's math score

0.0729*** (0.0135)

0.0666*** (0.0136)

0.0654*** (0.0137)

-0.00392 (0.00606)

-0.00323 (0.00611)

0.0343*** (0.00685)

0.0338*** (0.00688)

0.221*** (0.0183)

0.220*** (0.0184)

Participation Index

0.0240*** (0.00587)

Education Asset Index

0.0523*** (0.00672)

Child attends school

0.241*** (0.0173)

Mother attended ML

-0.000911 (0.0192)

Child attended ML Observations R-Squared

0.0235 (0.0197) 8367 0.709

8307 0.710

8310 0.710

8311 0.711

8367 0.721

8298 0.724

8249 0.724

Notes: This table presents regressions using the specification from Table 3, including intermediate outcomes listed and binary measures of self-reported attendance in ML classes as regressors. The sample consists of one randomly selected child aged 5 to 8 in the household. Participation and Education Asset Indices are defined in Tables 6 and 8, respectively. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

44

Appendices - Not For Publication A

Impact Heterogeneity

We use the following estimating equation to examine heterogeneity in treatment effects:

Y1iv = β0 + β1Variv + β2 MLv + β3CHAMPv + β4 MLCHAMPv + β5Variv ∗ MLv +β6Variv ∗CHAMPv + β7Variv ∗ MLCHAMPv + β8Y0iv + δ Gv + πHiv + εiv In this equation, Variv is the interacted variable for individual i in village v, and the remainder of the variables are defined as in Equation (1).

A.1

Children

Appendix Table 10 examines heterogeneity in treatment effects on children’s test scores. We examine heterogeneity by state, mother’s baseline test score, mother’s age, mother’s education, child’s age, child’s baseline test score, and child’s gender. Overall, there is little evidence of heterogeneity by any of these variables. Point estimates suggest approximately double the impact of ML and ML-CHAMP in Bihar, but neither interaction is statistically significant. The ML-CHAMP intervention had significantly larger impacts on older children, but there is no similar interaction with either of the individual interventions.

A.2

Mothers

Appendix Table 11 examines heterogeneity in treatment effects of the interventions on mothers’ test scores. We focus on heterogeneity by the state where the intervention took place, the mother’s baseline test score, mother’s age, and mother’s education level. Consistent with the point estimates from the analysis of children’s test scores, there is evidence that the ML and ML-CHAMP interventions were more effective in Bihar. The point estimates imply that these interventions resulted in more than double the impact on test scores in Bihar compared to Rajasthan, and the differences between the two states are significant at the 1 percent level for both interventions. The greater effectiveness of the ML and ML-CHAMP interventions in Bihar is consistent with the fact that mothers in the Bihar sample were 12 percentage points more likely to report having attended a class (see Appendix F). We also find that the CHAMP intervention was significantly more effective for mothers with higher initial test scores and for mothers with some education. These results are consistent with 45

the finding in Section 5.2 that CHAMP impacts were spread across higher level skills, particularly in language. We also find that the CHAMP intervention was slightly more effective for younger mothers.

B

Location Selection

Because of the slightly different organization of the villages in Rajasthan and Bihar, a different procedure was used to select study hamlets in each state. The procedure focused on finding distinct hamlets in which the programs could run while limiting spillovers. Hamlet eligibility was therefore determined based on size and distance from other target hamlets. Size and location of hamlets were determined from “Rapid Rural Assessments” conducted in study blocks. In Ajmer District in Rajasthan, villages are geographically separate, and each village is divided into several smaller hamlets. Hamlets met the size eligibility requirements if they contained between 40 and 100 households. To limit spillovers, one hamlet per village was selected. All villages in the blocks of Kekri and Bhinay were targeted for the intervention. To identify a total of 240 hamlets, the boundaries of Kekri and Bhinay were extended into a third block. In Purnia District in Bihar, the village boundaries are less distinct, and villages are denser than in Rajasthan. As in Rajasthan, each village is comprised of smaller hamlets. All villages in the blocks of Dhamdaha and B. Kothi were targeted for the intervention. Within these villages, hamlets were considered eligible if they contained between 25 and 150 households.1 To limit spillovers, hamlets in Bihar were selected only if they were 500 meters or more from of other selected hamlets.2,3

C

Additional Analysis of Impacts by Age and Selection for Surveys

This appendix provides additional analysis by age group and selection for focus in the household surveys. The impacts are estimated running the regression Equation (1) including dummies for age 1 The

size criteria differed between Rajasthan and Bihar because the criteria for Rajasthan would not have produced a sufficient number of eligible hamlets were it applied in Bihar. Due to the higher upper bound on number of households, Pratham agreed to hold more than one class in a target hamlet where necessary in Bihar. 2 GPS coordinates were used to confirm distances between hamlets. Distances were checked between hamlets within villages as well as across villages. 3 In two cases, target hamlets were eliminated because Pratham determined that adult literacy rates were too high to sustain classes.

46

group and interacting these dummies with the treatment groups: J

J

J

J

Y1iv = β0 + ∑ β1 j G ji + ∑ β2 j G ji ∗ ML + ∑ β3 j G ji ∗CHAMP + ∑ β4 j G ji ∗ MLCHAMP j=1

j=1

j=1

j=1

+ β8Y0iv + δ Gv + πHiv + εiv

where G ji is a dummy indicating child i’s membership in group j. In this estimation, the coefficients β2 j , β3 j , and β4 j are interpretable as the impacts of the treatment within group j.

C.1

Children aged 5 to 8 vs. Other Children

As noted in Section 2, we initially hypothesized that the impacts of the interventions would be most effective among children aged 5 to 8 at baseline. In addition mothers in the CHAMP program were specifically instructed to do the CHAMP activities with the 5 to 8-year-old children in the household. Appendix Table 12 estimates separate impacts for the age groups of 3 to 4, 5 to 8, and 9 to 14. In all three interventions, the impacts on the younger age group are small (typically 0.02 or less) and are not significant for any intervention or any test score category. The impacts on the 5 to 8 age group, on the other hand, are larger and mirror the results for the full sample in Table 4. Turning to the differences between the 5 to 8 age group and older children, the impacts are qualitatively similar for both groups in most cases.We note, however, that the testing sample only included older children in grades 1 to 4, and we thus cannot analyze impacts on all of the 9 to 14-year-old children in study households. The small sample size of the older group also results in relatively low power to test differences between the 5 to 8 age group and older children. We do observe a larger impact on math in the CHAMP intervention among the 5 to 8 age group compared with the older group, although the difference is not statistically significant. This could be due to the CHAMP program’s specific focus on activities with children between 5 and 8.

C.2

Children Selected for Surveys vs. Other Children

As described in Section 3, a number of questions in the household surveys focused on a single child aged 5 to 8 at baseline. When there was more than one child aged 5 to 8 in the household, one was randomly selected. In this subsection we analyze whether improvements in test scores may have arisen from this feature of the survey. We do this by comparing the randomly-selected

47

child to other children aged 5 to 8, within households that had more than one child aged 5 to 8.4 If the focus of the survey on one particular child shifted the mother’s focus to that child during the interventions, we would expect larger impacts on selected children relative to other children in the same age group within these households. Appendix Table 13 presents the disaggregated impacts. The table shows that for all three interventions, the randomly-selected children have qualitatively similar impacts to other children aged 5 to 8. In math, non-selected children have slightly larger impacts (0.054 vs 0.039 standard deviations for ML, 0.035 vs 0.031 for CHAMP, and 0.095 vs 0.053 for ML-CHAMP), but in no case can we reject that the impacts on the two groups are equal.

D

Attrition

We analyze attrition across three samples: the mothers’ survey sample (3.5 percent attrition), the mothers’ testing sample (3.4 percent attrition), and the children’s testing sample (6.0 percent attrition). Attrition of mothers was primarily driven by a failure to locate the entire household (typically because the household had moved). The additional attrition of children arose primarily because of logistical difficulties in testing several children per household during each visit and in finding all of the children present at home. Appendix Table 14 analyzes the levels of attrition across the three treatment groups. In the household survey and mothers’ testing samples, there is no evidence of differential attrition. However, there is some evidence of differential attrition in the children’s testing sample: children tested at baseline in the CHAMP and ML-CHAMP groups were about 2 percentage points more likely to exit the sample compared with the control group. Appendix Tables 15 and 16 further analyze the characteristics of mothers and children who attrit across treatment groups. In Appendix Table 15, we regress mothers’ attrition (defined as not completing either the endline survey or test) on baseline characteristics, treatment group dummies, and interactions between the treatment groups and baseline characteristics. These interactions indicate whether the characteristics of mothers who attrit are different across treatment groups. As shown in the table, there are few differences in the characteristics of those who attrit across treatment groups: the interactions are jointly significant at the 10 percent level in only one of the 21 regressions. Appendix Table 16 repeats the exercise for children. Again, the characteristics of children who attrit are largely balanced; however, children in each of the three treatment groups who attrited had significantly lower math and language test scores than those who attrited in the control group. 4 When

we estimate impacts separately by age group for households with more than one 5 to 8-year old child (following the specification in Appendix Table 12), the pattern of impacts is similar to that of the full sample (not shown).

48

To evaluate the possible effects of attrition on validity of the estimates on children’s test scores, Appendix Table 17 analyzes the balance of characteristics across treatment groups in the sample of children that was tested at endline. Unlike the randomization check in Table 1, this table uses child-level data to mirror the children’s learning analysis sample in Table 3. Overall, the endline balance is similar to baseline: of the 21 variables analyzed, two are jointly significant across the three treatment groups at the 10 percent level, one more is significant at the 5 percent level, and two more are significant at the 1 percent level. Critically, there is still no evidence of imbalance in any of the test score variables. Thus, even though there is some evidence of differential attrition in the child sample, balance in the analysis sample is generally maintained because of the low overall level of attrition. We present two analyses to examine the robustness of the children’s learning results to differential attrition. First, as presented previously in Appendix Table 4, inclusion of household-level controls in the children’s learning regressions leaves the estimated program effects virtually unchanged. Second, we construct bounds around the children’s learning estimates using Lee’s (2009) trimming method. The estimation uses only the sample of children tested at baseline and thus follows the specification in Appendix Table 4, Panel B, Columns 3, 6, and 9. We trim the top or bottom test scores in the control, ML, and CHAMP groups such that the generated attrition rates in each group are equal to those of the ML-CHAMP group, in which we observe the highest level of attrition. We then regress endline test scores on dummies for treatment group, baseline test scores, and dummies for stratum. As shown in Appendix Table 18, the small absolute differences in attrition between treatment groups lead to relatively tight bounds. For example, the estimated effect size of the CHAMP treatment on math scores ranges from 0.033 to 0.050 standard deviations. For the ML-CHAMP intervention, the effect ranges from 0.052 to 0.071 standard deviations. There is thus little evidence that differential attrition substantially influences our estimated impacts on children’s test scores.

E

Spillovers of ML treatment in Bihar

As shown in Table 2, 7 percent of mothers in the control and CHAMP groups reported attending adult literacy classes. Almost all of these mothers were in the Bihar sample. In this appendix we analyze whether proximity to ML classes could drive spillovers to control and CHAMP mothers. As noted in the text and in Appendix B, treatments in Bihar were assigned by hamlet, which is a sub-unit of a larger village. In some cases, multiple hamlets in the same village were included in the study if they were at least 500 meters apart. Appendix Table 19 regresses knowledge of classes and take-up of control and CHAMP mothers on measures of proximity to ML and MLCHAMP villages. As measures of proximity we use the number of ML and ML-CHAMP villages 49

within 1km, the number of hamlets within the larger village, and indicators for at least one ML or ML-CHAMP village within 1km or in the larger village. As shown in Panel A, nearby ML hamlets had little influence on knowledge of ML classes in the mother’s village. We do find that having at least one ML or ML-CHAMP hamlet within 1km increased knowledge of classes in the village by 6.4 percentage points (significant at 10 percent), but the impacts of the number of nearby hamlets and the presence of ML in the larger village have smaller and statistically insignificant effects. The small impact of a nearby ML and ML-CHAMP hamlet on knowledge of classes in the village in the one specification could be due to the survey’s use of the word “village” and not “hamlet”. This choice was made because hamlets are referred to using different Hindi words in Bihar and Rajasthan. In Panel B we examine spillovers on mothers’ self-reports of attendance. In all four specifications, we cannot reject the null hypothesis of no spillovers. The largest positive coefficient is in the specification using an indicator for a ML or ML-CHAMP hamlet within 1km, implying about 2.5 percentage points more attendance in those villages (p-value = 0.32). The confidence interval allows us to rule out effects above 7.4 percentage points, about half of the self-reported take-up in the control and CHAMP groups.

F

Correlates of Participation in ML and CHAMP

In Appendix Table 20 we analyze the determinants of mothers’ participation in ML classes and CHAMP sessions. We regress participation on a set of variables including household characteristics, mother education, children’s schooling behavior, time spent working, prior experience with literacy classes, self-help group membership, baseline participation in children’s education, and state. Column 1 analyzes the determinants of any participation in ML using the survey-based measure, while Columns 2 and 3 use the Pratham administrative data to analyze any participation and the intensity of participation, respectively. The determinants of ML participation are broadly similar across survey-based and administrative measures, with the highest predictive power in the regression using the survey-based measure in Column 1. Mothers’ education has a non-monotonic relationship with attendance in ML classes: mothers were more likely to attend when they had some exposure to formal schooling, but more years of education made them less likely to attend. In addition, mothers who scored higher on the baseline test were significantly more likely to attend. Because the test covered only the most basic competencies, the test and schooling results are consistent in that mothers with a small amount of education were more likely to attend.5 Participation in ML is also weakly related to the percentage 5 We

note, however, that less than 10 percent of mothers have 1 to 5 years of education, the range over which

50

of children in school at baseline. Beyond mothers’ and children’s characteristics, experience meeting in groups is a strong predictor of attendance in the ML classes: members of self-help groups were 8 to 10 percentage points more likely to attend a class (both significant at the 1 percent level). Past experience with adult literacy classes also has a significant positive relationship with self-reported attendance and Pratham’s measure of the percentage of classes attended. Our survey data also show that mothers in Bihar had 12 percent higher take-up than those in Rajasthan. This mirrors Pratham and research staff observations that mothers were on average more motivated and had more time to attend in Bihar. This relationship is not reflected in either of our administrative measures of take-up, although it is supported by the larger impacts we observe on mothers’ test scores in Bihar compared with Rajasthan (See Appendix A). This could reflect higher work hours in Rajasthan, where mothers reported working 43 percent more hours (73 hours per week compared with 51 hours in Bihar). However, conditional on state, there is no significant relationship between total hours worked and take-up. Column 4 of Appendix Table 20 analyzes the determinants of intensity of participation in CHAMP. Because 99 percent of households were reached at least once in CHAMP, there is little variation on the extensive margin of participation, and we therefore focus on the percentage of classes attended.6 Less educated mothers were slightly more likely to participate, with an additional year of education reducing the percentage of sessions attended by about 1 percentage point. Unlike ML, those with zero education were not significantly less likely to participate. Participation is also associated with higher children’s test scores and a larger fraction of children aged 5 to 8 in school. This could be explained by a complementarity of the CHAMP material with the work that the children were doing in school. As with participation in ML, we find evidence that participation is related to self-help group membership, but unlike ML it is not significantly related to past experience with adult literacy classes.

G

Surveyor Demand Effects

Although our main test score outcomes were based on tests conducted by our enumerators, our other outcomes are based on mothers’ self-reports. Because we do not have independent verification of these responses, we cannot completely rule out surveyor demand effects. However, the pattern of effects we find is inconsistent with surveyor demand effects that would generate posaverage participation is predicted to be highest as per the specification in Column 1, while 85 percent of mothers have zero education. In addition, as shown in the heterogeneity analysis in Appendix A, impacts on mothers’ learning were highest for mothers with zero education. 6 Because our data do not include mother presence in all cases, our measure of the percentage of sessions attended includes all sessions conducted at the household.

51

itive responses to the broad set of questions relating to education, or to only the questions that specifically relate to components of the ML or CHAMP programs. First, the patterns of impacts do not suggest that the mothers in the intervention groups responded in uniformly favorable ways to the broad set of questions relating to education. Although the interventions did impact our summary indices of participation, empowerment, and education assets in the home, the impacts on the underlying components are not universally and strongly positive, as would be the case if respondents exhibited surveyor demand effects for all topics relating to education. For example, the ML intervention, while influencing the overall participation index, did not have significant impacts on 5 of the 8 individual components of the index. In addition, as shown in Appendix Table 21, we see virtually no detectable impacts of any of the interventions on mothers’ aspirations for child attainment. A second possibility is that mothers were responding in ways that would be favorable to the specific components related to the programs in which they participated. Although we do observe larger impacts on some of these measures, this would be expected in the absence of surveyor demand effects. It is thus more challenging to separate surveyor demand effects from true program impacts. However, in this case we can also provide evidence inconsistent with systematic surveyor demand effects among these questions. In the ML program, the three specific outcomes that most closely relate to the program come from the set of empowerment variables listed in Appendix Table 22: the questions “Do you consider yourself literate?” “Do you count your change?”, and “Have you signed your name on official documents?” relate to skills taught in the classes, and we observe positive impacts on all three these outcomes. While it is not possible to fully verify the responses to these questions, we can check the report of signing one’s name against the mother’s ability to write her name in the endline test. Among mothers who were not able to write their name at endline, about 14 percent indicated they had signed their name on official documents. However, there is no significant difference in this “misreporting” between the villages where writing the mother’s name was taught (ML and ML-CHAMP) and villages in which no such instruction was given (Control and CHAMP groups). We also observe impacts from ML on a number of indicators that were not targeted by the program. For example, the ML program did not have a specific participation component, yet we do observe impacts on some of the participation measures, as well as the aggregate participation index. For the CHAMP program, we again see evidence of impacts on components targeted by the programs, particularly the intensive margins of helping children with homework and looking at the child’s notebook. However, there is no detectable impact on visiting the child’s school, which was also covered in the program. Beyond the participation measures, we also observe impacts on a number of components that were not explicitly part of the program, particularly the presence of 52

educational assets in the home.

H

Mothers’ Perceptions and Aspirations

In Appendix Table 21 we present impacts on a set of indicators that reflect mothers’ perceptions of parental involvement in education, aspirations for their children, and perceptions of their children’s reading and math ability. As shown in Panel A, the interventions had significant impacts on several measures of mothers’ perceptions of parental involvement. All three interventions significantly increased the mother’s belief that she should be involved in her child’s education. Through open-ended questions, we also counted the number of educational activities for which mothers perceived parents to be responsible and the number of activities the mother identified that she could do to help her child improve his or her studies.7 Mothers in the ML-CHAMP intervention identified more educational responsibilities of parents (significant at the 5 percent level), and the ML and ML-CHAMP treatment groups had significant impacts on the number of specific activities mothers identified they could do to help their children (significant at the 10 and 5 percent levels, respectively). By contrast, as shown in Panel B, we find we find little evidence that the programs influenced expectations and aspirations for child educational attainment. Across the four measures examined, the only significant impact is an increase of 0.25 years in the mother’s aspiration for her child’s grade attainment resulting from the ML-CHAMP intervention (significant at the 10 percent level). Panel C of Appendix Table 21 examines mothers’ perceptions of child language and math ability. Perceptions were based on integer scales of 0 to 4, corresponding to five levels of the ASER language and math tests. We find that the CHAMP and ML-CHAMP interventions significantly increased mothers’ perceptions of her child’s language and math scores. When compared to the child’s actual ability, however, the CHAMP and ML-CHAMP interventions caused mothers to be overly optimistic: both interventions increased the differences (in absolute value) between mothers’ perceptions and measured child ability in language and math. Our initial hypothesis was that through additional involvement, these interventions would increase the accuracy of the perception, and thus these results are surprising.8 One possible explanation for this result is that treated mothers overestimated the impacts of their involvement on their children’s learning. 7 For

both questions, the lists of activities were known to the surveyor, but not shared with mothers. If a mother included the activity in her response, the surveyor checked it off from that list. 8 In contrast, Dizon-Ross (2014) finds that providing information on child ability to parents in Malawi increases the accuracy of their beliefs and increases investment in level-appropriate educational materials.

53

I

Additional Analysis of Empowerment

As indicated in Section 5.3, the main analysis of empowerment in Table 5 includes a subset of the indicators originally listed in the PAP. This change was made to focus on the indicators that are most often used in the literature. In this section we discuss the impacts on the broader set of measures listed in the PAP. The impacts on these measures are analyzed in Appendix Table 22. As in the main empowerment analysis, we group these variables into categories of similar measures. Beyond the household decision variables analyzed in the main text, our measures included 17 additional variables that could reflect the mothers’ ability to make choices for themselves or their families.9 First, the “mobility and networks” group captures three measures of mothers’ mobility outside of the home, as well as two measures of self-help group membership. Women’s freedom of movement and access to external support networks reflect women’s autonomy outside of the household and have been frequently used in quantitative studies of empowerment (Malhotra et al., 2002). Second, the “capability” group captures seven measures of self-reported literacy, having signed her name on official documents, counting change during purchases, knowledge of the National Rural Employment Guarantee Scheme wage, and involvement in self-help group savings. This type of measure has also been referred to as “resources” that enable empowerment (Kabeer, 2009). Nonetheless, these measures have been seldom used as direct measures of empowerment in microscale quantitative work (Malhotra et al., 2002). We also note that these measures reflect life skills that could most directly be influenced by the programs. Third, our measures also include a set of beliefs and attitudes about girls’ and women’s education. These measures reflect views on social norms regarding women and are sometimes used in broad measures of empowerment (Pitt et al., 2006). Finally, we include one measure of selfreported happiness. Psychological well-being has been cited as a dimension of empowerment in the literature, but it is not typically analyzed in quantitative studies of empowerment (Malhotra et al., 2002). Appendix Table 22 presents impacts on these 17 variables, aggregated into 4 sub-indices based on the groups described above. The top row of the table aggregates these measures with the 9 decisions variables from Section 5.3 to create an broad empowerment index as per the PAP. Using this broader measure, we find significant impacts of ML of 0.071 standard deviations (significant at the 5 percent level) and ML-CHAMP of 0.14 standard deviations (significant at the 1 percent 9 We

exclude two measures listed in the PAP. The PAP includes one variable indicating whether the mother worked under the National Rural Employment Guarantee Scheme and one variable indicating whether she took advantage of other entitlements. The survey measured whether anyone in the household took advantage of these entitlements, and thus does not accurately measure empowerment of the mother. There were no significant impacts of the interventions on either of these measures (not shown).

54

level).10 These impacts are larger than those on an index created using only the indicators of mothers’ involvement in household decisions analyzed in Section 5.3. The larger impacts on the broader empowerment index are driven by the measures of capability and beliefs and attitudes. Both ML and ML-CHAMP had significant impacts on the capability subindex. As some of the capabilities were directly related to ML program content (e.g., signing one’s name), they also demonstrate the women putting these skills into practice to participate in the marketplace. All three interventions also had significant impacts on the beliefs and attitudes subindex, implying that both ML and CHAMP increased mother’s beliefs about the value of female education. We do not find significant impacts of any of the interventions on mobility and networks or on happiness.

10 When we split the index by when the components appear on the survey, find larger impacts on an index constructed

from the first 13 empowerment questions, compared with an index constructed from the last 13 questions (not shown). Although the difference in impacts could be driven by differences in the types of questions at different points in the survey, it is consistent with the possibility that the length of the survey could have decreased measured impacts.

55

References in Appendices Rebecca Dizon-Ross. Parents’ perceptions and children’s education: Experimental evidence from Malawi. Mimeo, MIT, 2014. Naila Kabeer. Resources, agency, and achievements: Reflections on the measurement of women’s empowerment. Development and Change, 30:435–464, 1999. David S. Lee. Training, wages, and sample selection: Estimating sharp bounds on treatment effects. Review of Economic Studies, 76:1071–1102, 2009. Anju Malhotra, Sidney Ruth Schuler, and Carol Boender. Measuring women’s empowerment as a variable in international development. In Deepa Narayan, editor, World Bank Workshop on Poverty and Gender: New Perspectives, pages 71–89. The World Bank, 2002. Mark M. Pitt, Shahidur R. Khandker, and Jennifer Cartwright. Empowering women with micro finance: Evidence from Bangladesh. Economic Development and Cultural Change, 54(4):791– 831, 2006.

56

57 Phase 1 (10 villages) No ML ML No CHAMP 2 3 CHAMP 3 2

-OR-

Rajasthan 240 Villages

Phase 2 (10 Villages) No ML ML No CHAMP 3 2 CHAMP 2 3

Phase 2 (10 Villages) No ML ML No CHAMP 2 3 CHAMP 3 2

Set (20 villages) ML No ML CHAMP 5 5 No CHAMP 5 5 Phase 1 (10 villages) No ML ML No CHAMP 3 2 CHAMP 2 3

Bihar 240 Villages

Full Sample 480 Villages

Appendix Figure 1: Stratification of Treatment Assignment

Appendix Figure 2: ML Impacts on Children's Math Scores vs. ML Impacts on Attendance, by Stratum

Impact on Children's Math Scores -.4 -.2 0 .2 .4

Panel A. Children's Math Impact vs Mothers' Attendance Impact Corr: 0.311 P-value: 0.024

0

.2

.4 .6 Impact on Mothers' Attendance

.8

Impact on Children's Math Scores -.4 -.2 0 .2 .4

Panel B. Children's Math Impact vs Children's Attendance Impact Corr: -0.087 P-value: 0.553

0

.2

.4 .6 Impact on Children's Attendance

.8

Notes: Each data point represents the impact of the ML treatment on child math scores within a stratum, graphed against the impact of ML on either mothers' or children's attendance within that stratum. Attendance is measured through household surveys.

58

59

96.7

In school - aged 6-14

95.5

45.8

61.44

17.26

24.2

64.43

25.56

13.95

58.25

5.42

91.1

41.29

59.13

25

35.3

66.4

32.1

15.6

71.3

5.27

--

36.69

55.3

21.4

36.49

61.5

25.17

10.81

60.18

5.04

82.22

--

--

42.98

56.29

86.75

36.5

11.46

80.79

6.75

97

49

59.78

6.59

48.78

51.42

10.2

25.81

10.37

5.48

93.3

39.28

48.38

4.6

44.8

36.6

6.1

19.2

8.1

5.03

--

45.23

54.25

4.87

46.23

37.48

7.59

26.58

7.64

5.12

82.63

--

--

5.14

59.53

57.9

8.56

22.24

15.4

6.53

Can read first grade text 57.5 52.7 51.9 --52.1 41.1 --students in grades 3-5 Notes: This table displays mean values from the 2011 Census of India and 2011 ASER survey (Columns 1-4 and 6-8), compared with data from the evaluation (Columns 5 and 8). Children's schooling and reading data come from the 2011 ASER survey, and the remainder come from the 2011 Census of India. Except for household size, all values are displayed as percentages. In the evaluation data, the percentage of children in school was calculated from the endline data for control group children aged 5 to 13 at baseline.

57.93

Literate - females, aged 7 and over

14.34

Motorcycle/Moped 67.77

46.16

Bicycle

Literate - aged 7 and over

51.22

Mobile Phone

17.32

Household Asset Ownership Radio 33.37

55.31

Household Has Electricity

Television

4.94

Household Size

Appendix Table 1: Comparison of State, District, and Block Characteristics 2011 Census of India, 2011 ASER Survey, and Evaluation Data Rajasthan (Rural) Bihar (Rural) India Ajmer Intervention Evaluation Purnia Intervention Evaluation (Rural) State District Blocks data State District Blocks data (2) (3) (4) (5) (6) (7) (8) (9) (1)

Appendix Table 2: Robustness of ML-CHAMP Impacts Impact of Treatment in Endline No Baseline All Significant Controls Only Controls Controls (1) (2) (3) (4) Outcome Child's language score

0.0685* 0.0456*** 0.0424*** 0.0463*** (0.0378) (0.0164) (0.0144) (0.0160)

Child's math score

0.0695* 0.0617*** 0.0558*** 0.0636*** (0.0361) (0.0174) (0.0152) (0.0170)

Child's total score

0.0685* 0.0557*** 0.0515*** 0.0570*** (0.0372) (0.0159) (0.0136) (0.0154)

Mother's language score

0.104** 0.0923*** 0.0885*** 0.0940*** (0.0465) (0.0149) (0.0134) (0.0150)

Mother's math score

0.163*** 0.145*** 0.145*** (0.0491) (0.0178) (0.0167)

0.145*** (0.0180)

Mother's total score

0.141*** 0.120*** 0.117*** (0.0484) (0.0153) (0.0140)

0.118*** (0.0155)

Decisions index

0.0673* (0.0382)

0.0673* 0.0918*** (0.0382) (0.0351)

0.0913** (0.0360)

Participation index

0.105** (0.0429)

0.102** 0.0966*** (0.0416) (0.0349)

0.104** (0.0417)

Education asset index

0.108** 0.108** 0.0768** 0.0837** (0.0435) (0.0426) (0.0335) (0.0408) Notes: Standard errors in parentheses. This table presents impact estimates of the ML-CHAMP treatment using varying sets of control variables. All specifications include dummy variables for the ML and CHAMP treatment groups. Column 1 controls only for stratum dummies. Column 2 controls for stratum dummies and the baseline value of the outcome variable (where measured). Column 3 controls for stratum, the baseline value of the outcome variable, and all controls listed in Table 1. For mother-level outcomes, child-level variables are averaged within the household. Column 4 displays impact estimates controlling for stratum, the baseline value of the outcome variable, and controls listed in Table 1 as significantly different at the 10 percent level between the ML-CHAMP and control groups. This includes an indicator for farming as the main source of household income, the number of children 0-4 in the household, the number of children 9-14 in the household, the number of adults in the household, and child gender. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Test score variables are described in Tables 3 and 4. The Decisions index, Participation index, and Education asset index are defined in Tables 5, 6, and 8, respectively. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

60

Appendix Table 3: Randomization Check Differences Across Treatments Mean Differences ML ML-CHAMP ML-CHAMP Control - CHAMP - ML - CHAMP (1) (2) (3) (4) Panel A. Household-level Variables First principal component of -0.0341 durables ownership [2.250] Main source of income: farming 0.494 [0.500] Number of children 0-4 0.929 [0.900] Number of children 5-8 1.438 [0.602] Number of children 9-14 0.976 [0.936] Number of children 15-17 0.265 [0.500] Number of adults 18 and over 2.914 [1.500] Father education (years) 0.389 [0.426] Mother education (years) 0.246 [0.376] Mother age 0.764 [2.282] Mother has past experience with 3.876 literacy classes [4.438] Fraction of household (15 and over) 32.29 that can read [7.102] Fraction of household (15 and over) 0.111 that can do math [0.307] Mother's language score (fraction) 0.299 [0.247] Mother's math score (fraction) 0.215 [0.241] Mother's total score (fraction) 0.250 [0.234] Panel B. Child-Level Variables (tested children) Child is male 0.520 [0.500]

N (5)

-0.0866 (0.0902) 0.00575 (0.0227) -0.0291 (0.0322) -0.00350 (0.0187) 0.0278 (0.0316) 0.000498 (0.0154) -0.101 (0.0623) -0.0326* (0.0192) -0.0243 (0.0177) -0.105 (0.112) -0.283 (0.220) -0.0599 (0.248) 0.000127 (0.0114) -0.00899 (0.0128) -0.00416 (0.0128) -0.00617 (0.0126)

-0.0233 (0.0881) -0.0389* (0.0206) -0.0787*** (0.0290) -0.0161 (0.0178) 0.00111 (0.0296) 0.0107 (0.0155) -0.0912 (0.0577) -0.0226 (0.0191) -0.0213 (0.0162) -0.0219 (0.103) -0.384* (0.212) -0.0204 (0.229) 0.0131 (0.0108) -0.00106 (0.0117) 0.000407 (0.0119) -0.000203 (0.0116)

0.0633 (0.0968) -0.0447* (0.0228) -0.0496 (0.0327) -0.0126 (0.0194) -0.0267 (0.0306) 0.0102 (0.0159) 0.00956 (0.0617) 0.0100 (0.0206) 0.00298 (0.0189) 0.0829 (0.120) -0.101 (0.236) 0.0395 (0.251) 0.0130 (0.0115) 0.00793 (0.0137) 0.00457 (0.0137) 0.00597 (0.0136)

8888

-0.00544 (0.0107)

0.00802 (0.0104)

0.0135 (0.0107)

15502

8885 8888 8888 8888 8888 8888 7576 7579 8864 8181 8888 8659 8857 8857 8857

Child attends school / aanganwadi

0.801 [0.399]

-0.0173 (0.0128)

-0.00720 (0.0139)

0.0101 (0.0137)

15501

Child's language score (fraction)

0.350 [0.296]

-0.00158 (0.0113)

-0.00276 (0.0110)

-0.00118 (0.0114)

15502

Child's math score (fraction)

0.277 [0.303]

-0.00163 (0.0115)

0.00288 (0.0109)

0.00451 (0.0116)

15502

Child's total score (fraction)

0.310 -0.00161 0.000372 0.00198 15502 [0.291] (0.0112) (0.0107) (0.0113) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display the differences in means between the treatment groups indicated. Differences in means are computed by OLS regression, controlling for stratum dummies. See Table 1 notes for variable definitions and notes on observation counts. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

61

62

0.0166 (0.0150)

Total

0.0205 (0.0174)

0.0592*** 0.0628*** (0.0147) (0.0152)

0.0233** 0.0297** (0.0119) (0.0139)

0.0238* 0.0340** (0.0142) (0.0170)

0.0322** 0.0430** (0.0155) (0.0177)

0.0118 (0.0146)

--

--

0.0282 (0.0171)

0.0417** (0.0185)

0.00976 (0.0178)

0.149*** 0.150*** (0.0163) (0.0167)

0.0884*** 0.0925*** (0.0130) (0.0141)

--

--

0.0515*** 0.0555*** 0.0575*** (0.0136) (0.0159) (0.0164)

0.0558*** 0.0598*** 0.0624*** (0.0152) (0.0172) (0.0182)

0.0424*** 0.0460*** 0.0472*** (0.0144) (0.0163) (0.0173)

ML-CHAMP Limited Missing Controls Baseline (8) (9)

0.0957*** 0.0968*** -0.0433*** 0.0485*** -0.124*** 0.127*** -(0.0124) (0.0132) (0.0113) (0.0126) (0.0130) (0.0137) Notes: Standard errors in parentheses. Columns 1, 4, and 7 display estimated coefficients from Tables 3 and 4 for Panels A and B, respectively. Columns 2, 5, and 8 use the same specifications as in Tables 3 and 4, controlling only for baseline test scores. Panel A, Columns 3, 6, and 9 use the same specification as in Table 3, controlling only for baseline test scores, and including only the sample of children tested at baseline. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

Total

--

Math

0.117*** 0.117*** (0.0160) (0.0164)

--

0.0139 (0.0174)

0.0364* (0.0191)

-0.0142 (0.0184)

Panel B: Mothers Language 0.0663*** 0.0680*** (0.0125) (0.0137)

0.0162 (0.0165)

0.0351** 0.0346** (0.0162) (0.0176)

Math

Panel A: Children Language -0.00679 -0.00725 (0.0158) (0.0170)

Appendix Table 4: Children's and Mothers' Test Scores, Alternative Specifications Impact of Treatment in Endline ML CHAMP Limited Missing Limited Missing Main Controls Baseline Main Controls Baseline Main (1) (2) (3) (4) (5) (6) (7)

Appendix Table 5: Treatment Effects on Children's ASER Learning Levels Endline Mean Impact of Treatment in Endline Control ML CHAMP ML-CHAMP N (1) (2) (3) (4) (5) Panel A. Language Competency Read letters 0.389 0.000331 0.00656 0.0262*** 18283 [0.430] (0.00813) (0.00783) (0.00794) Read words

0.176 [0.353]

-0.00320 (0.00606)

0.00184 (0.00556)

0.00803 (0.00562)

18283

Read paragraph

0.118 [0.322]

-0.00391 (0.00541)

-0.00265 (0.00503)

0.00875* (0.00481)

18283

Read story

0.0751 [0.264]

0.000962 (0.00568)

0.00297 (0.00535)

0.00566 (0.00538)

18283

0.562 [0.474]

0.0189** (0.00778)

0.0297*** (0.00786)

0.0353*** (0.00778)

18283

Two-digit number recognition

0.201 [0.368]

-0.000204 (0.00604)

-0.00238 (0.00582)

0.00267 (0.00581)

18283

Two-digit addition

0.198 [0.399]

0.0203** (0.00811)

0.00242 (0.00731)

0.00709 (0.00735)

18283

Panel B. Math Competency One-digit number recognition

Two-digit subtraction

0.0579 0.0106** -0.00503 0.00295 18283 [0.234] (0.00516) (0.00499) (0.00469) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

63

Appendix Table 6: Treatment Effects on Mothers' ASER Learning Levels Endline Mean Impact of Treatment in Endline Control ML CHAMP ML-CHAMP (1) (2) (3) (4) Panel A. Language Competency Read letters 0.173 0.0355*** 0.00929* 0.0471*** [0.339] (0.00555) (0.00497) (0.00592)

N (5) 8580

Read words

0.0924 [0.278]

0.00538 (0.00386)

0.0100*** (0.00387)

0.0112*** (0.00395)

8580

Read paragraph

0.0776 [0.268]

0.00209 (0.00391)

0.00572 (0.00412)

0.00180 (0.00383)

8580

Read story

0.0622 [0.242]

0.000696 (0.00455)

0.00355 (0.00444)

0.00116 (0.00435)

8580

0.470 [0.461]

0.0675*** (0.0101)

0.0276*** (0.00967)

0.107*** (0.0108)

8580

Two-digit number recognition

0.0960 [0.275]

0.0150*** (0.00399)

0.00831** (0.00397)

0.0155*** (0.00393)

8580

Two-digit addition

0.0781 [0.268]

0.00503 (0.00474)

0.00791 (0.00511)

0.0152*** (0.00504)

8580

Panel B. Math Competency One-digit number recognition

Two-digit subtraction

0.0355 0.00632 0.00676 -0.00197 8580 [0.185] (0.00409) (0.00450) (0.00394) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

64

Appendix Table 7: Mothers' Weekly Time Use, PAP Categories Endline Mean Impact of Treatment in Endline Control ML CHAMP ML-CHAMP (1) (2) (3) (4)

N (5)

Help with homework

2.308 [2.704]

0.0940 (0.0899)

0.109 (0.0931)

0.0642 (0.0894)

8533

Read with children

0.320 [1.371]

-0.00384 (0.0391)

-0.0377 (0.0356)

-0.00489 (0.0401)

8454

Play with children

1.334 [3.286]

0.106 (0.113)

0.0611 (0.118)

-0.0279 (0.107)

8551

Share stories with children

0.497 [1.422]

0.0321 (0.0458)

-0.0198 (0.0426)

0.0238 (0.0503)

8536

Paid / agricultural work

31.14 [21.08]

0.926 (0.599)

0.542 (0.581)

0.956 (0.598)

8561

Livestock work

9.662 [7.040]

0.116 (0.252)

-0.336 (0.231)

0.474* (0.251)

8576

Collect animal feed

6.963 [6.773]

0.180 (0.250)

-0.242 (0.245)

0.184 (0.270)

8577

Collect wood

3.278 [5.102]

0.123 (0.165)

-0.0382 (0.177)

0.145 (0.174)

8570

Housework

19.02 [8.008]

0.264 (0.299)

0.112 (0.297)

0.242 (0.288)

8581

Buy household supplies

1.201 [2.843]

-0.0767 (0.0913)

-0.0539 (0.0931)

0.0239 (0.0897)

8572

Look after children

4.646 -0.175* -0.125 0.113 8581 [3.679] (0.105) (0.107) (0.104) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Time spent in each activity measured in hours per week. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

65

Appendix Table 8: Children's Weekly Time Use, PAP Categories Endline Mean Impact of Treatment in Endline Control ML CHAMP ML-CHAMP (1) (2) (3) (4)

N (5)

Homework

4.059 [4.331]

-0.0708 (0.140)

0.102 (0.133)

0.181 (0.133)

7892

Read

0.450 [1.561]

-0.0626 (0.0461)

-0.0362 (0.0497)

0.0310 (0.0497)

8156

Draw / paint

0.685 [1.562]

0.00942 (0.0501)

0.0486 (0.0492)

0.0751 (0.0514)

8153

Play with adult

0.542 [1.982]

0.0167 (0.0650)

-0.0351 (0.0646)

-0.109* (0.0612)

8405

Tuition classes

2.266 [4.908]

0.206 (0.179)

0.110 (0.195)

0.00590 (0.171)

8466

Television

3.701 [5.067]

-0.127 (0.166)

0.0845 (0.171)

0.0133 (0.163)

8414

Housework

3.599 [4.110]

0.0594 (0.132)

0.140 (0.129)

0.160 (0.132)

8449

Household business

1.886 0.198 0.0483 0.304** 8447 [4.153] (0.135) (0.131) (0.137) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables listed in Table 1. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Children's time use measured for one randomly selected child per household aged 5-8 at baseline. Time spent in each activity measured in hours per week. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

66

Appendix Table 9. Cost Effectiveness of Interventions ML (1)

Intervention CHAMP ML-CHAMP (2) (3)

Panel A: Cost Summary ($) Pratham Staff 37,521 30,699 68,219 Volunteer Time 21,653 — 21,653 Training, monitoring, materials 14,861 4,730 19,590 Total 74,035 35,428 109,463 Panel B: Standard Deviation Improvement per $100 spent—Children Language — — 0.18 Math 0.22 0.40 0.24 — 0.30 0.22 Total Children affected 4,572 4,448 4,653 Panel C: Standard Deviation Improvement per $100 spent—Mothers Language 0.19 0.14 0.17 Math 0.34 0.35 0.29 Total 0.28 0.26 0.24 Mothers affected 2,173 2,105 2,140 Notes: Costs incurred in rupees converted to dollars using 2011 exchange rate of 46.7 rupees/dollar. Volunteer time estimated based on average daily wage in non-agricultural occupations. Mothers and children affected calculated using the same inclusion criteria as Tables 3 and 4, respectively. Cost-effectiveness estimates in Panels B and C are computed based on the effect sizes reported in Tables 3 and 4, multiplied by the number of mothers or children affected, and divided by the total costs of each program in 100s of dollars. Cost-effectiveness is only calculated for effects significant at the 10 percent level.

67

68

-0.0681 (0.0520) 0.0112 (0.0155) 0.0187 (0.0258) 0.0153 (0.0166)

Mother's age

Mother has some education

Child's age

Child's baseline test score

0.0204 (0.0155)

0.0129 (0.0269)

0.0195 (0.0144)

-0.0158 (0.0513)

0.0225 (0.0143)

0.0180 (0.0207)

0.0549*** (0.0152)

0.00168 (0.0258)

0.0462*** (0.0140)

0.0344 (0.0541)

0.0520*** (0.0137)

0.0339* (0.0196)

0.0107 (0.0137)

-0.000224 (0.00430)

0.0436 (0.0386)

0.00263* (0.00156)

0.0114 (0.0129)

0.00901 (0.0299)

0.00144 (0.0136)

0.00180 (0.00432)

0.0326 (0.0366)

0.00123 (0.00150)

0.0113 (0.0129)

0.0106 (0.0283)

-0.00371 (0.0127)

0.00797* (0.00419)

0.0361 (0.0383)

0.000527 (0.00158)

0.0194 (0.0131)

0.0323 (0.0267)

Variable * ML-CHAMP (6)

14576

18282

18235

18283

18221

18283

N (7)

0.0129 0.0369** 0.0467*** 0.00740 -0.0258 0.00964 18283 (0.0183) (0.0176) (0.0171) (0.0199) (0.0202) (0.0200) Notes: Standard errors in parentheses. Each column displays the results of a regression of the child's normalized total test score on treatment dummies, the interaction variable indicated, and interactions of the variable and treatment dummies. Regressions include all control variables used in Table 3. Mother's and child's test scores are the mother's and child's total normalized test scores. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.0170 (0.0151)

Mother's baseline test score

Gender = boy

0.0118 (0.0216)

State = Bihar

Interacted Variable

ML (1)

Appendix Table 10. Heterogeneity in Impact Outcome: Child's Total Test Score Impact of Treatment in Endline Variable * Variable * ML CHAMP CHAMP ML-CHAMP (2) (3) (4) (5)

69

0.0808* (0.0451)

Mother's age

0.144*** (0.0463)

0.0426*** (0.0109)

0.0604*** (0.0144)

0.156*** (0.0472)

0.125*** (0.0129)

0.0811*** (0.0162)

0.000469 (0.00131)

-0.00453 (0.0143)

0.0731*** (0.0243)

-0.00314** (0.00133)

0.0321** (0.0129)

-0.0342 (0.0212)

-0.000968 (0.00137)

-0.00764 (0.0136)

0.0865*** (0.0254)

Variable * ML-CHAMP (6)

8580

8552

8580

N (7)

0.0978*** 0.0315*** 0.131*** -0.0224 0.0724* -0.0439 8556 (0.0127) (0.0113) (0.0131) (0.0430) (0.0384) (0.0416) Notes: Standard errors in parentheses. Each column displays the results of a regression of the mother's normalized total test score on treatment dummies, the interaction variable indicated, and interactions of the variable and treatment dummies. Regressions include all control variables used in Table 4. Mother's test score is the mother's total normalized test score. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.0957*** (0.0123)

Mother's baseline test score

Mother has some education

0.0582*** (0.0147)

State = Bihar

Interacted Variable

ML (1)

Appendix Table 11. Heterogeneity in Impact Outcome: Mother's Total Test Score Impact of Treatment in Endline Variable * Variable * ML CHAMP CHAMP ML-CHAMP (2) (3) (4) (5)

Appendix Table 12: Children's Test Scores by Age Group Impact of Treatment in Endline 3-4 year old 5-8 year old 9-14 year Difference: Difference: children children old children 5-8 vs. 3-4 5-8 vs 9-14 All Children (1) (2) (3) (4) (5) (6) Panel A: ML Language

-0.00679 (0.0158)

-0.00291 (0.0158)

-0.00761 (0.0202)

-0.0202 (0.0377)

-0.00471 (0.0236)

0.0125 (0.0393)

Math

0.0351** (0.0162)

0.0152 (0.0159)

0.0385* (0.0210)

0.0454 (0.0405)

0.0233 (0.0234)

-0.00686 (0.0422)

Total

0.0166 (0.0150)

0.00723 (0.0147)

0.0182 (0.0194)

0.0162 (0.0344)

0.0110 (0.0215)

0.00203 (0.0361)

Panel B: CHAMP Language 0.0118 (0.0146)

0.0130 (0.0169)

0.00678 (0.0190)

0.0361 (0.0370)

-0.00622 (0.0239)

-0.0294 (0.0393)

0.0322** (0.0155)

0.00665 (0.0155)

0.0464** (0.0204)

0.00771 (0.0390)

0.0397* (0.0225)

0.0387 (0.0427)

0.0238* (0.0142) Panel C: ML-CHAMP Language 0.0424*** (0.0144)

0.00989 (0.0151)

0.0294 (0.0187)

0.0214 (0.0342)

0.0195 (0.0214)

0.00802 (0.0374)

0.0205 (0.0170)

0.0306* (0.0181)

0.131*** (0.0391)

0.0101 (0.0226)

-0.100** (0.0407)

Math

0.0250 (0.0154)

0.0677*** (0.0199)

0.0360 (0.0395)

0.0427* (0.0222)

0.0316 (0.0423)

Math Total

0.0558*** (0.0152)

Total

0.0515*** 0.0238 0.0527*** 0.0819** 0.0289 -0.0292 (0.0136) (0.0154) (0.0178) (0.0344) (0.0210) (0.0372) Notes: Standard errors in parentheses. Column 1 displays estimates from the main analysis in Table 3. The remaining columns in the table display the results of regressions of of child test score on treatment dummies interacted with dummies for each of the three age groups. See Appendix C for full estimating equation. Regressions include all control variables used in Table 3. Regressions include 18,283 observations, of which 4192 are in the younger age group, 11,783 are in the middle age group, and 2307 are in the older age group. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

70

Appendix Table 13: Selected Children vs. Other Children Aged 5 to 8 Impact of Treatment in Endline All 5-8 Other 5-8 Difference: year old Selected year old Selected vs. children children children Other 5-8 (1) (2) (3) (4) Panel A: ML Language 0.00913 0.0107 0.00770 0.00298 (0.0237) (0.0268) (0.0300) (0.0317) Math

0.0464* (0.0254)

0.0387 (0.0310)

0.0535* (0.0304)

-0.0148 (0.0346)

Total

0.0306 (0.0232)

0.0269 (0.0267)

0.0339 (0.0283)

-0.00699 (0.0299)

Panel B: CHAMP Language 0.00990 (0.0220)

0.0140 (0.0264)

0.00615 (0.0290)

0.00789 (0.0341)

0.0335 (0.0241)

0.0314 (0.0298)

0.0354 (0.0308)

-0.00398 (0.0370)

0.0236 (0.0219) Panel C: ML-CHAMP Language 0.0327 (0.0223)

0.0244 (0.0260)

0.0229 (0.0282)

0.00146 (0.0321)

0.0300 (0.0267)

0.0350 (0.0300)

-0.00500 (0.0355)

Math Total

Math

0.0754*** 0.0534** 0.0953*** (0.0234) (0.0269) (0.0311)

Total

-0.0419 (0.0348)

0.0580*** 0.0443* 0.0704** -0.0261 (0.0215) (0.0243) (0.0290) (0.0321) Notes: Standard errors in parentheses. The sample used in this table is restricted to children aged 5 to 8 in households multiple children in that age group. "Selected children" are those children aged 5 to 8 selected to be the focus of specific questions in the household survey. Column 1 displays treatment effect estimates as per the specification used in Table 3. The remaining columns in the table display the results of regressions of children's test scores on treatment dummies interacted with dummies for each of the two groups. See Appendix C for full estimating equation. Regressions include all control variables used in Table 3. Regressions include 6510 observations, of which 3094 were selected as the focus for specific survey questions and 3416 are other children in that age group. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

71

Appendix Table 14: Attrition Across Treatment Groups Mean Relative to Control P-value: All Control ML CHAMP ML-CHAMP coeffs 0 N (1) (2) (3) (4) (5) (6) Mother attritted: survey 0.0308 0.00700 0.00327 0.00593 0.756 8888 [0.173] (0.00530) (0.00525) (0.00523) Mother attritted: test 0.0313 0.00647 0.00152 0.00647 0.538 8857 [0.174] (0.00536) (0.00517) (0.00527) Child attritted: test 0.0488 0.00879 0.0190*** 0.0213*** 0.135 15502 [0.215] (0.00650) (0.00683) (0.00752) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display the differences in means between each treatment group and the control group. Column 5 displays the p-value of the F-test that the differences in means between the treatment groups and control group are all zero. Differences in means are computed by OLS regression, controlling for stratum dummies. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

72

Appendix Table 15: Characteristics of Attriters by Treatment Group: Mothers Dependent Variable: Mother Attrited (Either Survey or Test) Interaction with Treatment Groups Baseline P-value: Variable (Control) ML CHAMP ML-CHAMP All coeffs. 0 N (1) (2) (3) (4) (5) (6) Panel A. Household-Level Variables First principal component of -0.00320* 0.00118 0.00170 0.00562 0.816 8859 durables ownership (0.00192) (0.00274) (0.00255) (0.00526) Main source of income: farming

-0.0252*** (0.00785)

0.00453 (0.0117)

0.00642 (0.0106)

0.00212 (0.00923)

0.841

8856

Number of children 0-4

0.00215 (0.00478)

0.00593 (0.00675)

0.000356 (0.00593)

0.00542 (0.00763)

0.758

8859

Number of children 5-8

-0.00119 (0.00674)

0.000161 (0.00921)

-0.0116 (0.00908)

0.0191 (0.0148)

0.423

8859

Number of children 9-14

-0.00821** (0.00372)

-0.00274 (0.00574)

0.000140 (0.00550)

0.00599 (0.00916)

0.957

8859

Number of children 15-17

-0.0219*** (0.00638)

0.0205** (0.00964)

0.0173** (0.00865)

0.00260 (0.00634)

0.122

8859

Number of adults 18 and over

-0.00175 (0.00239)

0.00736* (0.00403)

0.00482 (0.00389)

-0.00288 (0.0109)

0.294

8859

Father's education (years)

-0.0197** (0.00930)

0.0211 (0.0146)

0.0117 (0.0133)

0.00454 (0.00885)

0.368

7556

Mother's education (years)

-0.0115 (0.00990)

0.0255 (0.0162)

0.00599 (0.0133)

0.00650 (0.00725)

0.241

7559

Mother's age

-0.00117 (0.00145)

0.00391 (0.00243)

0.00370 (0.00259)

0.00297 (0.00568)

0.283

8835

Mother has past experience with literacy classes

-0.000732 (0.000780)

0.00231* (0.00133)

0.00138 (0.00122)

0.00630 (0.00769)

0.333

8154

Fraction of household (15 and over) -0.000838* that can read (0.000479)

-0.000937 (0.000761)

-0.000458 (0.000747)

0.00986 (0.0253)

0.639

8859

Fraction of household (15 and over) -0.0259*** that can do math (0.00832)

0.0196 (0.0140)

0.0260* (0.0137)

0.00437 (0.00605)

0.201

8630

Mother's language score (fraction)

-0.00755 (0.0148)

0.0343 (0.0232)

0.0320 (0.0236)

0.00581 (0.00909)

0.276

8857

Mother's math score (fraction)

-0.00555 (0.0155)

0.0522** (0.0251)

0.0316 (0.0232)

0.00513 (0.00774)

0.128

8857

Mother's total score (fraction)

-0.00700 (0.0158)

0.0479* (0.0251)

0.0342 (0.0244)

0.00530 (0.00849)

0.155

8857

Panel B. Average of Tested Children Child is male -0.00376 (0.00974)

0.0112 (0.0141)

0.000709 (0.0135)

0.00334 (0.00943)

0.850

8750

Child attends school / aanganwadi

0.00469 (0.0116)

-0.0127 (0.0178)

-0.0560*** (0.0216)

0.0181 (0.0168)

0.081

8750

Child's language score (fraction)

-0.0140 (0.0143)

0.00115 (0.0221)

-0.000977 (0.0198)

0.00930 (0.0102)

0.973

8750

Child's math score (fraction)

-0.0150 (0.0162)

-0.000908 (0.0237)

0.00126 (0.0212)

0.00587 (0.00901)

1.000

8750

Child's total score (fraction)

-0.0155 -0.0000238 0.000347 0.00724 0.998 8750 (0.0159) (0.0239) (0.0212) (0.00981) Notes: Standard deviations in parentheses. This table presents the results of regressions of mother attrition (either in the survey or test) on the baseline variable indicated, the three treatment groups, interactions between the variable and each treatment group. Column 1 represents the relationship between the variable and attrition in the control group, while Columns 2-4 represent the difference in this relationship between the treatment group indicated and the control group. Regressions control for stratum dummies. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

73

Appendix Table 16: Characteristics of Child Attriters by Treatment Group Dependent Variable: Child Attrited Interaction with Treatment Groups Baseline P-value: Variable ML CHAMP ML-CHAMP All coeffs. 0 (Control) (1) (2) (3) (4) (5)

N (6)

First principal component of durables ownership

-0.00249 (0.00194)

0.00146 (0.00289)

-0.00238 (0.00252)

-0.00194 (0.00300)

0.520

15502

Main source of income: farming

0.00159 (0.00890)

-0.0153 (0.0124)

-0.0244* (0.0130)

-0.0203 (0.0137)

0.243

15498

Number of children 0-4

0.000848 (0.00608)

0.0118 (0.00806)

0.000419 (0.00724)

-0.00572 (0.00774)

0.105

15502

Number of children 5-8

0.00518 (0.00768)

-0.0117 (0.00980)

-0.0222** (0.0103)

-0.00948 (0.0108)

0.196

15502

Number of children 9-14

-0.00850** (0.00402)

-0.00245 (0.00565)

-0.00597 (0.00604)

0.00362 (0.00619)

0.504

15502

Number of children 15-17

-0.0314*** (0.00622)

0.0324*** (0.00960)

0.0227** (0.00971)

0.0185* (0.00985)

0.006

15502

Number of adults 18 and over

0.00101 (0.00294)

0.00269 (0.00448)

-0.000901 (0.00372)

-0.00260 (0.00373)

0.652

15502

Father's education (years)

-0.00203 (0.0120)

0.00108 (0.0178)

-0.00447 (0.0156)

-0.0185 (0.0187)

0.724

13289

Mother's education (years)

0.00607 (0.0129)

-0.00263 (0.0194)

-0.0242 (0.0163)

-0.0229 (0.0182)

0.344

13294

Mother's age

0.00413** (0.00207)

-0.00168 (0.00286)

-0.00151 (0.00314)

-0.000879 (0.00327)

0.940

15464

Mother has past experience with literacy classes

0.00131 (0.000906)

-0.000736 (0.00145)

-0.00119 (0.00132)

-0.000998 (0.00155)

0.824

14364

Fraction of household (15 and over) -0.000510 that can read (0.000648)

-0.00123 (0.000819)

-0.00110 (0.000900)

-0.000988 (0.000951)

0.476

15502

Fraction of household (15 and over) that can do math

-0.0104 (0.0109)

0.000652 (0.0168)

-0.00355 (0.0198)

-0.0164 (0.0175)

0.780

15117

Mother's language score (fraction)

0.0120 (0.0155)

0.0139 (0.0238)

0.00418 (0.0253)

-0.0164 (0.0260)

0.742

15460

Mother's math score (fraction)

0.0169 (0.0164)

0.0212 (0.0247)

0.00780 (0.0259)

-0.0252 (0.0275)

0.436

15460

Mother's total score (fraction)

0.0161 (0.0168)

0.0194 (0.0244)

0.00658 (0.0267)

-0.0233 (0.0282)

0.512

15460

Child is male

0.00604 (0.00660)

0.00338 (0.00988)

0.0145 (0.00993)

0.00397 (0.0102)

0.518

15502

Child attends school / aanganwadi

0.00491 (0.00839)

-0.00650 (0.0124)

-0.0379** (0.0175)

-0.0139 (0.0143)

0.172

15501

Child's language score (fraction)

0.0348*** (0.0121)

-0.0413** (0.0161)

-0.0389** (0.0192)

-0.0436** (0.0186)

0.036

15502

Child's math score (fraction)

0.0402*** (0.0138)

-0.0414** (0.0181)

-0.0449** (0.0197)

-0.0394** (0.0201)

0.069

15502

Child's total score (fraction)

0.0404*** -0.0441** -0.0451** -0.0439** 0.047 15502 (0.0136) (0.0179) (0.0202) (0.0201) Notes: Standard deviations in parentheses. This table presents the results of regressions of child attrition on the baseline variable indicated, the three treatment groups, interactions between the variable and each treatment group. Column 1 represents the relationship between the variable and attrition in the control group, while Columns 2-4 represent the difference in this relationship between the treatment group indicated and the control group. Regressions control for stratum dummies. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

74

Appendix Table 17: Balance After Attrition: Child-level Data Mean Relative to Control Control (1)

P-value: CHAMP ML-CHAMP All coeffs. 0 (3) (4) (5) 0.138 0.0417 0.516 (0.104) (0.0958) 0.00342 0.0576** 0.051 (0.0258) (0.0234) 0.0672* 0.117*** 0.002 (0.0407) (0.0354) 0.0142 -0.00587 0.655 (0.0266) (0.0236) 0.0537 0.0777** 0.040 (0.0349) (0.0318) 0.0111 -0.00762 0.773 (0.0184) (0.0171) 0.197** 0.186*** 0.007 (0.0775) (0.0682) 0.0219 0.00517 0.234 (0.0201) (0.0191) 0.0279 0.0207 0.220 (0.0187) (0.0164) 0.136 0.0452 0.566 (0.115) (0.102) 0.0704 0.199 0.330 (0.231) (0.213) -0.178 -0.143 0.825 (0.299) (0.271) -0.00919 -0.0228* 0.364 (0.0141) (0.0133) 0.0112 0.00446 0.722 (0.0132) (0.0119) 0.00881 0.00464 0.900 (0.0132) (0.0120) 0.00981 0.00457 0.833 (0.0130) (0.0118)

-0.0726 [2.282] 0.479 [0.500] 1.074 [0.926] 1.591 [0.668] 1.033 [0.955] 0.273 [0.502] 2.956 [1.608] 0.381 [0.426] 0.242 [0.375] 0.708 [2.154] 3.759 [4.376] 32.35 [6.950] 0.114 [0.309] 0.294 [0.241] 0.208 [0.234] 0.244 [0.227]

ML (2) 0.00492 (0.0882) 0.0227 (0.0235) -0.00342 (0.0355) -0.0194 (0.0248) 0.0739** (0.0312) 0.00460 (0.0173) 0.0281 (0.0659) -0.0192 (0.0184) -0.00257 (0.0159) -0.00500 (0.0957) -0.182 (0.208) -0.254 (0.270) -0.00994 (0.0134) -0.00227 (0.0108) 0.000837 (0.0109) -0.000460 (0.0106)

Child is male

0.511 [0.500]

-0.00735 -0.00305 (0.00979) (0.0103)

-0.0126 (0.0101)

0.628

18283

Child attends school / aanganwadi

0.798 [0.401]

0.00895 0.0323** (0.0137) (0.0136)

0.0181 (0.0139)

0.090

15018

Child's language score (fraction)

0.348 [0.296]

0.0109 (0.0112)

0.0130 (0.0116)

0.0134 (0.0112)

0.599

14576

Child's math score (fraction)

0.274 [0.302]

0.0149 (0.0113)

0.0178 (0.0120)

0.0114 (0.0112)

0.447

14576

First principal component of durables ownership Main source of income: farming Number of children 0-4 Number of children 5-8 Number of children 9-14 Number of children 15-17 Number of adults 18 and over Fraction of household (15 and over) that can read Fraction of household (15 and over) that can do math Mother's education (years) Father's education (years) Mother's age Mother has past experience with literacy classes Mother's language score (fraction) Mother's math score (fraction) Mother's total score (fraction)

Child's total score (fraction)

N (6) 18283 18282 18283 18283 18283 18283 18283 15670 15672 18235 17024 18283 17831 18221 18221 18221

0.307 0.0131 0.0157 0.0123 0.509 14576 [0.290] (0.0110) (0.0116) (0.0109) Notes: Standard deviations in square brackets, standard errors in parentheses. The sample in each row includes all children tested at endline for whom baseline values of the indicated variable are available. Columns 2, 3, and 4 display the differences in means between each treatment group and the control group. Column 5 displays the p-value of the F-test that the differences in means between the treatment groups and control group are all zero. Differences in means are computed by OLS regression, controlling for stratum dummies. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

75

76

0.0417** (0.0207)

0.00976 (0.0183) 0.0325 0.0504** (0.0207) (0.0207)

-0.000793 0.0245 (0.0183) (0.0183)

0.0355* (0.0190)

0.0630*** (0.0190)

0.0624*** 0.0519** 0.0710*** (0.0207) (0.0207) (0.0207)

0.0472** (0.0190)

0.0139 0.00885 0.0173 0.0282 0.0184 0.0373* 0.0575*** 0.0462*** 0.0672*** (0.0186) (0.0186) (0.0186) (0.0199) (0.0199) (0.0199) (0.0171) (0.0171) (0.0171) Notes: Standard errors in parentheses. Standard errors computed from 500 bootstrap draws, sampling by village. Sample for regressions is restricted children who took the baseline test. Columns 1, 4, and 7 display estimates from Appendix Table 4, Panel A, Columns 3, 6, and 9. Columns 2-3, 5-6, and 8-9 display the bounds of the estimates using Lee's (2009) trimming method. Number of observations equals 14,576 for main estimates and 14,438 for upper and lower bound estimates. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

0.0364* 0.0320 0.0414** (0.0200) (0.0200) (0.0200)

Math

Total

-0.0142 -0.0196 -0.00777 (0.0190) (0.0190) (0.0190)

Language

Appendix Table 18: Children's Test Scores, using Lee (2009) bounds Impact of Treatment in Endline ML CHAMP ML-CHAMP Estimate Lower Upper Estimate Lower Upper Estimate Lower Upper (1) (2) (3) (4) (5) (6) (7) (8) (9)

Appendix Table 19: Tests for ML Spillovers to CHAMP and Control Hamlets in Bihar Type of Independent Variables Number of close villages Dummy for close village in in each category each category (1) (2) (3) (4) Panel A. Dependent Variable: Mother Knew of Classes in Village Hamlets within 1km 0.0267 0.0641* (0.0217) (0.0327) Hamlets in village Constant

-0.00317 (0.0185) 0.290*** (0.0215)

R-Squared 0.002 N 2145 Panel B. Dependent Variable: Mother Attended Hamlets within 1km 0.00366 (0.0132) Hamlets in village Constant

-0.0196 (0.0349)

0.307*** (0.0255)

0.279*** (0.0230)

0.315*** (0.0292)

0.000 2145

0.005 2145

0.000 2145

0.0248 (0.0249) 0.00188 (0.0139)

0.130*** (0.0154)

0.131*** (0.0179)

-0.0126 (0.0254) 0.123*** (0.0165)

0.139*** (0.0201)

R-Squared 0.000 0.000 0.001 0.000 N 2145 2145 2145 2145 Notes: Each column displays estimated coefficients of an OLS regression of self-reported mother attendance on the variables listed. Sample restricted to Bihar control and CHAMP villages. The independent variables in Columns 1 and 2 represent the number of nearby ML and ML-CHAMP villages in each category, while the independent variables in Columns 3 and 4 are dummies representing the presence of each type of nearby village. Out of 120 CHAMP and Control hamlets, 50 have a an ML or ML-CHAMP hamlet within 1 km (average 0.56), and 68 have an ML or ML-CHAMP hamlet in the larger village (average 0.83). Standard errors, in parentheses, are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

77

Appendix Table 20: Determinants of Mothers' Take-up Intervention ML Pct. Attended Attended Attended (Survey) (Admin) (Admin) (1) (2) (3)

CHAMP Pct. Attended (Admin) (4)

First principal component of durables ownership

-0.00504 (0.00500)

-0.00403 (0.00487)

-0.00291 (0.00253)

-0.000330 (0.00215)

Number of children 5-14

0.0118 (0.00758)

0.00377 (0.00626)

0.000355 (0.00393)

0.00696** (0.00305)

Fraction of children 5-8 in school or aanganwadi

0.0474* (0.0242)

0.0540* (0.0280)

0.0141 (0.0141)

0.0457*** (0.0137)

Average children's total test score (normalized)

-0.00982 (0.0373)

-0.0569* (0.0306)

0.0151 (0.0201)

0.0540*** (0.0150)

Father's education (years)

-0.00374* (0.00219)

-0.00192 (0.00188)

0.000313 (0.00125)

-0.000142 (0.000916)

Mother's education > 0

0.144*** (0.0432)

0.0336 (0.0361)

0.0461* (0.0240)

0.0137 (0.0217)

Mother's education (years)

-0.0466*** -0.0289*** -0.0241*** (0.00744) (0.00713) (0.00357)

-0.0109*** (0.00334)

Mother's age

-0.00130 (0.00127)

0.000688 (0.00105)

-0.000877 (0.000605)

0.00101** (0.000482)

Mother's total test score (normalized)

0.0566*** (0.0152)

0.0330*** (0.0125)

0.0312*** (0.00870)

0.00795 (0.00574)

Mother has past experience with literacy classes

0.0674** (0.0286)

-0.00318 (0.0257)

0.0292* (0.0162)

0.00810 (0.00962)

Mother is member of self-help group

0.101*** (0.0253)

0.0790*** (0.0254)

0.0471*** (0.0142)

0.0204** (0.00873)

Baseline mothers' participation index

0.0248*** (0.00914)

0.00117 (0.00727)

0.00718 (0.00468)

0.00341 (0.00327)

Total hours worked (in and out of home) per week

0.000341 (0.000385)

0.000211 -0.0000158 (0.000316) (0.000169)

State = Bihar

0.120*** (0.0297)

0.0193 (0.0367)

-0.0231 (0.0164)

-0.0000363 (0.000154) -0.0703*** (0.0128)

CHAMP & ML-CHAMP Mean of Dep. Var. 0.425 0.810 0.254 0.852 R-Squared 0.0572 0.0274 0.0265 0.0538 N 3795 3914 3870 3875 Notes: Each column displays estimated coefficients of an OLS regression of the mother attendance measure indicated on the variables listed. The dependent variables in Columns 1 and 2 are dummies for attendance in ML classes from the survey and administrative data, respectively. The dependent variables in Columns 3 and 4 are the percentage of ML classes or CHAMP sessions attended, respectively. The mothers' participation index is defined in Table 6. A self-help group is a group of villagers that pools savings and provides loans to members of the group. Standard errors, in parentheses, are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01. Sample

ML & ML-CHAMP

78

Appendix Table 21: Mothers' Aspirations and Perceptions Impact of Treatment in Endline Endline Mean Control ML CHAMP ML-CHAMP (1) (2) (3) (4) Panel A. Perceptions of parental involvement 0.727 0.0408*** 0.0296** 0.0396*** Believes mother should be responsible for children's education [0.446] (0.0129) (0.0133) (0.0129)

N (5) 8467

Number of educational activities for which parents are responsible

1.393 [1.035]

0.0613 (0.0418)

0.0439 (0.0406)

0.0749** (0.0367)

8581

Number of things mother can do to help child improve studies

1.902 [1.120]

0.0689* (0.0412)

0.0121 (0.0444)

0.0937** (0.0397)

8513

Identifies home help as reason for children's achievement Panel B. Expectations/aspirations for child's attainm Believes child will pass 8th grade

0.210 [0.407]

0.00777 (0.0141)

0.00437 (0.0132)

-0.00508 (0.0138)

8581

0.851 [0.357]

-0.00255 (0.0118)

-0.00540 (0.0115)

-0.000294 (0.0110)

8224

Believes child will pass 12th grade

0.649 [0.477]

0.000714 (0.0162)

-0.00183 (0.0162)

-0.00312 (0.0163)

7978

Highest grade to which mother aspires for child to study

10.13 [3.051]

-0.0132 (0.107)

0.124 (0.108)

0.245** (0.116)

4841

Highest grade to which mother aspires for child to study: child's wish Panel C. Mother's perceptions of child's reading/ma Perception of child's language ability (out of 4)

0.432 [0.495]

0.0110 (0.0125)

-0.00929 (0.0145)

0.00416 (0.0137)

8513

2.458 [1.586]

-0.0286 (0.0460)

0.154*** (0.0472)

0.118** (0.0472)

8028

Perception of child's math ability (out of 4)

2.569 [1.610]

0.00978 (0.0479)

0.211*** (0.0494)

0.154*** (0.0530)

8101

Absolute value of difference between language perception and child's score

1.603 [1.365]

-0.0157 (0.0428)

0.133*** (0.0446)

0.0714 (0.0447)

7822

Absolute value of difference between math 1.480 -0.0315 0.156*** 0.0937** 7897 perception and child's score [1.270] (0.0431) (0.0464) (0.0433) Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. "Child" refers to one randomly selected child per household aged 5-8 at baseline. "Number of educational activities for which parents are responsible" is constructed as the number of responses the mother gave to the open-ended question, "When it comes to children's education, what are the responsibilities parents have other than making sure their children attend school?" Possible responses include: helping with homework, checking homework, telling the child to study, meeting with teachers, buying educational materials, sending the child for tuitions, sending the child to a better school, and spending money on the child's education. "Number of things mother can do to help child improve studies" is constructed as the number of responses the mother gave to the open-ended question "What can you do to help your child do better in his/her studies?" Possible responses include those listed above in addition to spending more time with the child, playing with the child, telling stories to the child, and sending the child to school. The regression using "highest grade to which mother aspires for child to study" as an outcome has missing observations due to mothers indicating that grade attainment would be the child's decision. Regressions using beliefs of 8th and 12th grade passing and perceptions of child ability have fewer observations due to responses of "don't know" to the survey questions. Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01.

79

Appendix Table 22: Individual Empowerment Measures from PAP Impact of Treatment in Endline Endline Mean Control ML CHAMP ML-CHAMP N (1) (2) (3) (4) (5) Index of All Measures (26 variables)

0.000 [1.000]

0.0710** 0.0545 (0.0347) (0.0340)

0.144*** (0.0362)

8581

Mobility And Networks

0.000 [1.000]

0.00450 (0.0383)

-0.0321 (0.0399)

-0.00416 (0.0412)

8581

Times left village in the past month

1.150 [1.648]

0.0366 (0.0575)

0.0687 (0.0601)

0.0992 (0.0636)

8581

Left without adult accompaniment

0.113 [0.316]

0.00279 0.0286*** (0.0101) (0.0105)

0.00828 (0.0104)

8581

Left village without permission

0.0168 [0.129]

-0.00358 0.000985 (0.00360) (0.00384)

-0.00227 (0.00367)

8581

Member of Self-Help Group

0.310 [0.463]

-0.0204 (0.0203)

-0.0251 (0.0175)

8888

How Often Attends Self-Help Group Meetings

0.289 [0.441]

-0.0001 -0.0579*** (0.0201) (0.0209)

-0.0248 (0.0212)

8578

Capability

0.000 [1.000]

0.0687** 0.0172 (0.0347) (0.0349)

0.131*** (0.0358)

8581

Considers self literate

0.233 [0.423]

0.0476*** 0.00637 (0.0126) (0.0124)

0.0691*** (0.0129)

8581

Signed name of official documents

0.561 [0.496]

0.0612*** 0.00733 (0.0134) (0.0126)

0.0765*** (0.0130)

8581

Counts change

0.867 [0.339]

0.0255** 0.0219** (0.0108) (0.0106)

0.0415*** (0.0105)

8581

Caught mistakes counting change

0.314 [0.464]

0.0147 (0.0168)

-0.00911 (0.0160)

0.0239 (0.0165)

8581

Knows NREGA wage

0.621 [0.485]

-0.0612** -0.00750 (0.0243) (0.0245)

-0.0145 (0.0242)

8356

Does not believe husband should be more educated than wife

0.127 [0.333]

-0.00327 (0.0130)

-0.0187 (0.0133)

-0.0121 (0.0131)

8581

Does not believe girls should be at home or married when 18

0.0135 [0.116]

0.00462 0.00574 (0.00372) (0.00395)

0.00365 (0.00387)

8332

Beliefs and Attitudes

0.000 [1.000]

0.0628** 0.121*** (0.0318) (0.0348)

0.135*** (0.0349)

8581

Does not believe husband should be more educated than wife

0.330 [0.470]

0.0464*** 0.0136 (0.0158) (0.0164)

0.0579*** (0.0165)

8315

Does not believe girls should be at home or married when 18

0.0449 [0.207]

-0.00286 0.00527 (0.00636) (0.00666)

0.00248 (0.00674)

8581

Believes girls should be doing further studies when 18

0.280 [0.449]

0.0642*** 0.0865*** (0.0154) (0.0154)

0.105*** (0.0158)

8581

Would have wanted to study up to (grade level)

5.799 [4.362]

-0.412*** (0.142)

0.149 (0.144)

-0.309** (0.135)

8290

Happiness

3.107 [1.443]

0.0473 (0.0493)

0.0590 (0.0502)

0.0382 (0.0464)

8581

-0.0274 (0.0172)

Notes: Standard deviations in square brackets, standard errors in parentheses. Columns 2, 3, and 4 display estimated coefficients of a regression of the outcome in each row on treatment group dummies, controlling for stratum dummies, baseline values (where available), and all variables in Table 1, using household-level averages for child-level variables. Missing values of control variables are coded as 0, with additional dummies indicating missing values. Each index is a normalized average of z-scores of the component variables of the index, using the control group means and standard deviations. Baseline indices only include indicators for which data were collected. The index of all measures includes all individual measures in this table, as well as the 9 measures of mother involvement in household decisiosn listed in Table 5. "Knows NREGA wage" indicates whether the mother answered an amount between Rs. 100 and Rs. 150 to the question "What is the per day wage fixed by the government under NREGA?" Official wage rates in Bihar and Rajasthan at the time of the survey were Rs. 122 and Rs. 124, respectively. "Happiness" is the answer to the question "Last week, how happy were you?" coded on a scale from 1 (very sad) to 5 (very happy). Standard errors are clustered at the village level. * denotes significance at 0.10; ** at 0.05; *** at 0.01

80