MID-TERM IMPACT ASSESSMENT REPORT Odisha Particularly Vulnerable Tribal Group Empowerment and Livelihood Improvement Program (OPELIP) Alliance Bioversity-CIAT VALUE CHAIN DEVELOPMENT PROGRAMME (PHASE II) – PRODEFI-II Jonathan Mockshell, Sylvester Ogutu, James Garrett, Thea Ritter, Andrew Kennedy, Andrea Fongar Alliance of Bioversity and CIAT/CGIAR Ally July 2022 2 Acknowledgements The authors would like to thank all IFAD, PMU, and ICRISAT staff who assisted with the design and implementation of this impact assessment and provided inputs for this report. This especially includes Frew Behabtu, Romina Cavatassi, Ulac Demirag, Tisorn Songsermsawas, and Rasha Omar from IFAD. The continued support from OPELIP PMU staff is also greatly appreciated. Particularly, we would like to thank PMU Director Arthanari P. and Goutam Mohanty for their continuous support. We express our thanks to all the PMU focal points for their support during the fieldwork. We also would like to thank Swamikannu Nedumaran, Nandi Ravi, and Ravichand K. from ICRISAT-India and Abilash Ananthula (consultant) for providing technical support and coordinating the data collection. We also acknowledge the HDI consulting company hired to collect the data, specifically Braj Das and the HDI team for data collection. We thank Ricardo Labarta, Carolina Gonzalez, Edward Martey, and Collins Asante- Addo for reviewing the study materials. Excellent research assistance from Diego Alverez and Eli Akorsikumah is appreciated. Finally, we thank all the households that were interviewed for their time and patience. This work was funded through a large grant from IFAD for the project “Linking Research to Impact: Increasing the Effectiveness of Agriculture and Food Systems in Improving Nutrition” and with additional support from the CGIAR Research Program on Agriculture for Nutrition and Health. The support of IFAD and the CGIAR is gratefully acknowledged. 3 Table of contents Acknowledgements ................................................................................................................................. 2 Table of contents .................................................................................................................................... 3 List of abbreviations ................................................................................................................................ 5 Summary ................................................................................................................................................. 6 1. Introduction ........................................................................................................................................ 8 2. Theory of change and main impact assessment questions ................................................................ 9 2.1 Development problem and theory of change............................................................................... 9 2.2 Impact assessment questions ..................................................................................................... 13 2.3 Unintended program impacts ..................................................................................................... 14 2.4 Target population ....................................................................................................................... 14 3. Impact assessment design ................................................................................................................ 16 3.1 Construction of counterfactual groups ....................................................................................... 16 3.2 Sample size calculations .............................................................................................................. 17 3.3 Sampling strategy ........................................................................................................................ 18 3.4 Data collection ............................................................................................................................ 19 3.5 Identification strategy ................................................................................................................. 20 4. Results ............................................................................................................................................... 21 4.1 Descriptive results ....................................................................................................................... 21 4.1.1 Distribution of tribal groups across the treatment and control samples ............................ 21 4.1.2 Household demographics .................................................................................................... 22 4.1.3 Housing characteristics ........................................................................................................ 23 4.1.4 Schooling outcomes ............................................................................................................. 25 4.1.5 Migration and remittances .................................................................................................. 26 4.1.6 Access to credit .................................................................................................................... 28 4.1.7 Land characteristics and farming practices.......................................................................... 30 4.1.8 Women’s empowerment ..................................................................................................... 34 4.1.9 Agricultural yield and sales .................................................................................................. 36 4.1.10 Livestock ownership ........................................................................................................... 40 4.1.11 Well-being indicators ......................................................................................................... 42 4.1.12 Income/livelihood diversification ...................................................................................... 44 4.1.13 Food security outcomes ..................................................................................................... 46 4.1.14 Dietary quality .................................................................................................................... 48 4.1.15 Anthropometric indicators ................................................................................................. 50 4.2 Estimation results ....................................................................................................................... 55 4.2.1 Effects of OPELIP on agricultural inputs, value of production, and sales revenue .............. 56 4 4.2.2 Effects of OPELIP on well-being indicators .......................................................................... 58 4.2.3 Effect of OPELIP on income and livelihood diversification .................................................. 59 4.2.4 Effect of OPELIP on food security indicators........................................................................ 60 4.2.5 Effect of OPELIP on child anthropometric indicators .......................................................... 61 4.2.6 Effect of OPELIP on dietary quality and adult anthropometric indicators ........................... 62 4.2.7 Effect of OPELIP on other selected indicators ..................................................................... 63 5. Conclusions and recommendations .................................................................................................. 65 References ............................................................................................................................................ 68 Appendix ............................................................................................................................................... 72 5 List of abbreviations BMI Body Mass Index CPI Consumer Price Index DID Difference-in-Differences FAO Food and Agriculture Organization of the United Nations FIES Food Insecurity Experience Scale GP Gram Panchayat HDDS Household Dietary Diversity Score IFAD International Fund for Agricultural Development MDD-Ch Minimum Dietary Diversity for Children MDD-M Minimum Dietary Diversity for Men MDD-W Minimum Dietary Diversity for Women MPA Micro-project Agency OPELIP Odisha Particularly Vulnerable Tribal Group Empowerment and Livelihoods Improvement Program OTELP Odisha Tribal Empowerment and Livelihoods Project PMU Program Management Unit PSM Propensity Score Matching PVTG Particularly Vulnerable Tribal Group SC Scheduled Caste STs Scheduled Tribes TLUs Tropical Livestock Units VDAs Village Development Associations 6 Summary Although there has been a huge investment in agricultural development interventions, few rigorous evaluations of such interventions exist and little is known about their impact, particularly among the most vulnerable communities. In recent years, governments and development agencies have also paid greater attention and made significant investments in shaping these projects to be nutrition sensitive so they contribute to improving nutrition among beneficiary households as well as the general population. Despite knowing the links between agriculture and nutrition conceptually, few evaluations of the impacts of such nutrition- sensitive projects on nutrition or other aspects of livelihoods exist. This study reports on the findings of a midline evaluation of the effects of the Odisha Particularly Vulnerable Tribal Group Empowerment and Livelihoods Improvement Program (OPELIP) on various indicators. OPELIP is an IFAD-funded project carried out by the Government of Odisha State and the national Government of India, in partnership with micro- project agencies (MPA) and villages. OPELIP is being implemented in 12 districts of Odisha, India, and aims to reduce poverty and improve the well-being and nutritional status of about 62,356 households, including PVTGs, scheduled castes, and scheduled tribes, largely through support to improvements in agriculture-based production, marketing, and livelihoods. The midline assessment aims to provide useful mid-program implementation insights and feedback to improve ongoing program implementation and design of new aspects for the second half of the program, from 2021 to 2024, as well as for future programs. While we expected OPELIP to have positive effects on beneficiary households, as outlined in the project design documents (e.g., increased adoption of yield-enhancing inputs and nutrient- dense crop varieties to improve dietary diversity and dietary quality; more secure land tenure; improved agricultural productivity to contribute to income growth; greater empowerment of women and an improved gender balance; and improved child nutrition and food and nutrition security), the results from a sample of about 2,000 households analyzed using difference-in- differences methods to account for potential confounding factors do not show significant effects for most of the indicators. Positive effects are only observed for use of chemical fertilizer, the value of production, monthly per adult-equivalent incomes, consumption expenditures, and poverty reduction. The insignificant effects of OPELIP on many of the outcomes defined during project design are possibly due to delays in program implementation or slow uptake of certain OPELIP interventions by beneficiaries, such as access to irrigation, adoption of yield or soil fertility improvement innovations, adoption of kitchen gardens to promote more diverse diets, land titling, and women’s empowerment training. Complementary studies were also undertaken to examine these operational issues in more depth and are reported elsewhere. The results from this quantitative evaluation show that OPELIP is contributing to improved value of production, monthly per adult equivalent incomes, consumption expenditures, and poverty reduction. However, more attention should be given to the implementation of all three program components ((1) community empowerment; (2) natural resource management, food and nutrition security, and livelihood improvement; and (3) community infrastructure and drudgery reduction) since the effects of the program on most of the indicators under the program components were insignificant. We recommend paying attention to the promotion and training on kitchen gardens, crop diversification, and nutrient-rich local crops. Increasing agricultural productivity should also be coupled with increased market access through 7 infrastructure development. Supporting women’s participation in self-help groups and the acquisition of land titles is critical to increasing the impact of the program. 8 1. Introduction The Odisha Particularly Vulnerable Tribal Group Empowerment and Livelihoods Improvement Program (OPELIP) is a comprehensive program that aims to empower and improve the livelihoods and food and nutrition security of vulnerable tribal communities in the state of Odisha (previously known as Orissa), India. Scheduled Tribes (STs) are among the poorest population groups in rural India (IFAD, 2014). Among the 705 STs, the Government of India further defines 75 as Particularly Vulnerable Tribal Groups (PVTGs). PVTGs often have a relatively small but homogenous population, are relatively more isolated physically and socially, and are even further deprived of key development indicators, such as food and nutrition security, education, and health (IFAD, 2014). OPELIP commenced in mid-2017 and is planned to end in 2024. The program covers 12 districts in the state of Odisha and aims to improve the living standards of at least 62,356 beneficiary households, namely 32,090 PVTGs, 13,970 other ST households, 5,486 Scheduled Caste (SC) households, and 10,810 other households (IFAD, 2014). OPELIP was rolled out as a scale-up following the successful implementation of a similar IFAD-funded program called the Odisha Tribal Empowerment and Livelihoods Project (OTELP), which was implemented from 2004 to 2015 (OTELP, 2016). OPELIP uses a comprehensive approach that includes support for improved access to land and land titles, capacity building to improve gender balance, access to credit and markets, improvement of community infrastructure, and essential social services (IFAD, 2014). OPELIP is also defined by IFAD as a “nutrition- sensitive” project, meaning that its design integrated elements with the specific intent of improving the nutritional status of beneficiary households. The midline assessment was conducted in 2021 and follows the baseline assessment of the program conducted in 2017 (the same baseline households are revisited). The midline assessment aims to provide useful mid-program implementation insights and feedback to improve ongoing program implementation and design of new aspects for the second half of the OPELIP, from 2021 to 2024. Although there has been a huge investment in agricultural development interventions, little is known about their impact, especially among vulnerable tribal groups and rural populations in South Asia. In addition, in recent years, governments and development agencies have been paying greater attention and making significant investments in shaping agricultural projects to be nutrition sensitive, so they contribute to improving nutrition among beneficiary households as well as the general population. Despite knowing the links between agriculture and nutrition conceptually, few evaluations of the impact of such nutrition-sensitive projects on nutrition or other aspects of livelihoods exist (Bird et al., 2019; Pandey et al. 2016; Ruel et al., 2018; Winters et al., 2010; World Bank, 2011). This midline assessment contributes towards filling this knowledge gap and supports IFAD’s efforts in generating a large body of evidence on the impacts of agricultural and rural development interventions on economic development and improved living standards of poor and vulnerable populations. It also adds to the impact assessment portfolio used to assess the overall impact of IFAD’s projects across the globe. Generally, assessing the impact of agricultural interventions should rank high on the research and development policy agenda out of the need to generate concrete evidence for effective policymaking and/or to achieve inclusive rural transformation through agricultural development (Gertler et al., 2016). 9 The goal of OPELIP is to reduce poverty and improve the wellbeing of PVTGs through income growth and improved food and nutrition security for about 62,356 households, including other tribal and poor households in the intervention areas. The program aims to achieve this goal through the following four components: (1) community empowerment; (2) natural resource management, food and nutrition security, and livelihood improvement; (3) community infrastructure and drudgery reduction, and (4) program management. These four components (explained in more detail below) are expected to play a crucial role in improving the well-being of PVTG households by providing solutions to their weak community-based institutions, infrastructure, and overdependence on subsistence farming and forest gathering for their livelihoods, in addition to shedding light on the program’s overall management. This report presents the results of the midline assessment of the impacts of OPELIP on program beneficiaries using the same set of indicators in both baseline and midline survey rounds. It also includes some additional indicators, particularly concerning nutrition outcomes. The rest of this report is organized as follows: Section 2 describes the program’s theory of change (impact pathways), impact assessment questions, unintended program effects, and coverage or target population; Section 3 provides an overview of the sampling strategy and sample size across districts and villages, as well as a description of the impact assessment design and identification strategy; Section 4 presents descriptive statistics by program participation and estimation results of selected outcome indicators; Section 5 concludes. 2. Theory of change and main impact assessment questions 2.1 Development problem and theory of change While India has achieved the status of a middle-income economy, rural poverty is still widespread. Malnutrition is also a public health concern in India with almost one-half of children below three years of age being underweight (low weight-for-age) or stunted (low height-for-age) (IFAD, 2014). Poverty varies by tribe, caste, occupation, etc. For instance, STs are among the poorest, followed by SCs. Among STs, PVTGs are the poorest and most vulnerable. India’s 75 PVTGs are distributed across more than 14 states. The state of Odisha has the largest number of PVTGs (13 out of 75), with about 58,708 PVTG households (IFAD, 2020a; Rao et al., 2020). These PVTG households live in remote areas, are socially marginalized, have limited employment and income-generating opportunities, are extremely poor, and are vulnerable to the adverse effects of climate change, which affects their primary source of livelihood (subsistence farming and forest gathering). PVTGs also experience challenges with food and nutrition security, literacy, and health (IFAD, 2014). As result, there is a need to promote and support community-based institutions and infrastructure for PVTGs to reduce their vulnerability and improve their well-being. OPELIP aims to contribute to the improved well-being of PVTGs through four program components. We briefly highlight activities, expected outcomes, and impacts of each of the four program components with the aim of showing the program’s theory of change (ToC). The ToC shows the causal mechanisms (pathways) through which the program’s inputs and activities are expected to lead to outcomes and impacts. According to the program’s design (IFAD, 2014), the four program components are: (1) Community empowerment. This entails the promotion of self-help groups (SHGs) primarily for rural finance and savings and village development associations (VDAs) to plan need- 10 based activities. Under this component, individuals through SHGs receive training on rural finance and savings. The promotion of financial skills through SHGs is expected to enable the social development of SHG members by facilitating group savings and access to credit. Under this component, the program also aims to build the capacity of VDAs by providing training to plan and execute the needs of the community, which may include natural resource management, community-based paralegal services, healthcare, nutrition education, and community infrastructure. We would expect the increased capacity from these trainings along with additional support services to enhance the formation of and membership in women’s SHGs, increase savings, strengthen rural finance systems, and improve the gender balance, food and nutrition security, and health. (2) Natural resource management, food and nutrition security, and livelihood improvement. This component entails issuing landless households with land title certificates for their households and for the land they have been cultivating. This component also aims to improve land and water resource management through the construction of irrigation and storage facilities. In addition, it aims to improve food and nutrition security through support for crop improvement activities (on-farm demonstrations, quality seed production, and farmer field schools); promotion and training on kitchen gardens, crop diversification and nutrient-rich local crops (pulses, millets, vegetables), nutritious food items (meat, eggs, fish, milk, and milk products), and fruits and spices (mango, cashew, litchi, lime, orange, turmeric, ginger, chilies, etc.); livelihood improvement activities (poultry and goat rearing); and training on the development of grain banks and local market access infrastructure. Activities under this component are expected to lead to improvements in agricultural productivity, incomes, market access, and food and nutrition security. (3) Community infrastructure and drudgery reduction. Activities under this component include the installation of rice hullers, drying yards, milling units, and threshing floors; provision of weighing scales, household storage bins, and food processing facilities for promoting value-addition and fair trade in villages; supply of smokeless stoves and support to operations that maintain village fuel-wood reserves; and the construction and upgrading of village link roads, SHG work sheds, toilets, and gravity-fed clean drinking water supply facilities. Sufficient supply of these community infrastructure and drudgery reduction activities is expected to increase the use of clean fuel and improve sanitation and access to clean drinking water and thereby improving food security and the nutritional status of beneficiaries, as well as increasing agricultural productivity in program villages. (4) Program management. This component involves program administration and budget management, program operations, and implementation. It aims to improve program implementation performance by ensuring that the program’s goals are achieved through sound decision-making, budgeting, and effective implementation of program activities. OPELIP program management is done through the Program Management Unit (PMU) which is based in Bhubaneswar, Odisha and is responsible for the day-to-day management of the program. The PMU houses the Program Monitoring and Evaluation and Knowledge Management Unit. Overall, we expect the program to have positive impacts on beneficiary households and villages if the interventions and services are offered as planned and if beneficiary households adopt and use the interventions and services offered. In particular, among beneficiaries, we 11 expect the program to lead to the adoption of improved crop varieties (improved nutrient-dense millets, pulses, cereals, fruits, and vegetables), improved agricultural productivity, income growth, income diversification, improved food and nutrition security, increased resilience to shocks, improved women’s empowerment and gender balance, enhanced child nutrition and health, and increased school enrolment. At the village- or community-level, we expect the program to improve the capacity of village and community groups to identify and mobilize resources, reduce gender gaps, and lower food prices at local markets through improved access to markets and availability of food items (i.e., an enhanced local food economy). We also expect intervention villages to have better access to community resources (e.g., health centers) than comparison villages. Figure 1 summarizes the Theory of Change for OPELIP. As mentioned, the ToC shows the causal mechanisms (pathways) through which the program’s inputs and activities lead to outcomes and impacts. The ToC closely follows the program’s logical framework developed by IFAD and also includes the IFAD11 indicators shown in Table A1 in the appendix. All the IFAD11 indicators are analyzed, except for indicators on youth, persons with a disability, climate change, and job creation which were not included in the baseline assessment. 12 Figure 1: Theory of change Source: Modified from IFAD (2017). Community Level • Communities can identify and mobilize additional resources • Gender gap reduced • Local markets developed • Enhanced local food economy • Community health improved OUTCOMES IMPACTS OUTPUTS INPUTS AND ACTIVITIES A S S U M P T IO N S • There is sufficient demand for training and support for village development association promotion • There is a willingness to accept land titles and develop land upon title receipt • There is sufficient demand for drudgery reduction activities • Program activities are implemented as planned • Shocks do not adversely affect program Natural Resource Management, Food and Nutrition Security, and Livelihood Improvement • Land titles to landless, land/water management • Kitchen gardens, growing fruits, spices, and nutrient- rich crops, and poultry and goat rearing • Irrigation and storage • Crop improvement, grain banks, and market access Community Empowerment • Capacity building through self-help groups, rural finance, training on gender, nutrition (causes, impacts, and remedies of child malnutrition, infant feeding, etc.), hygiene, and health • Promoting village development associations to plan need-based activities Household Level • Improved women’s empowerment • Increased value of production, agricultural productivity • Increased agricultural commercialization and value of sales • Increased household incomes • Enhanced wellbeing (asset ownership, consumption, and household income) • Incomes diversified • Food and nutrition security improved • Increased resilience • Increased school enrolment • Improved hygiene and health • Improved maternal and child nutrition (anthropometry) and health Household Level • Improved women’s access to finances, control over resources • Secured land tenure for agriculture and homestead • Increased agriculture area, input use, and household spending on inputs • Increased land under irrigation • Increased asset ownership and access to credit • Increased agronomic skills • Improved caregivers’ knowledge and feeding practices • Reduced malnutrition and improved child health • Improved sanitation and access to clean drinking water • Diversified diets (pulses, cereals, milk, fish, eggs, fruits, vegetables) for children, men, and women • Reduced drudgery (e.g., in millet processing) Community Level • Village development associations are capacitated to provide basic services • Distance to markets reduced • Increased involvement of women and youth in community activities • Irrigated area increased • Increased use of clean energy fuels • Increased land development • Reduced drudgery for women Household Level • Markets for inputs, credit, output, etc., exist and function well • Beneficiaries face no other barriers to improving productivity, such as land access, soil quality, weather conditions Community Level • Reasonable support to community groups by state and local governments • Women’s self-help groups formed, village development associations are functional and efficient, and training on gender, nutrition, reproductive health, marriage, etc. provided • Landless households provided land titles or rights • Water supply schemes and rain-water harvest constructed and operational • Gravity fed clean drinking water structures built • Farmer Field Schools conducted to provide knowledge and skills on post-harvest management, nutrition (training on nutrient- rich local foods, causes, effects, and remedies of malnutrition), home gardens promoted • Production and consumption of nutrient-dense local crops (pulses, millets, vegetables), fruits, and food items promoted • Beneficiaries trained on poultry and goat rearing, and on other means of income generation • Rice hullers, drying yards, grain (e.g., lentil) banks, and food (e.g. millet) processing facilities constructed • Smokeless stoves and solar lanterns installed and fuel wood reserves established • Beneficiaries respond to program interventions by taking up offered services • There is a positive response to opportunities to improve food and nutrition security • Training and Farmer Field Schools are effective and will lead to adoption of recommended practices Community Infrastructure and Drudgery Reduction • Installing rice hullers, drying yards, food (e.g., millet) processing facilities, smokeless stoves, fuel wood reserves • Construction of toilets and drinking water structures • Easy year-round access to drinking water 13 2.2 Impact assessment questions Following the ToC, the baseline assessment questions and indicators (IFAD, 2017), and the IFAD11 indicators (such as land tenure, value of production, value of market sales, household income, consumption, asset ownership, nutrition or food security, gender (e.g., women’s empowerment), and resilience), the midline assessment aims to answer the following questions: (1) Does the program through its activities, such as the provision of irrigation and seed/lentil storage infrastructure, on-farm demonstrations, and training on improved agricultural production techniques, increase the value of agricultural production or productivity? (2) Does the program increase access to secure agricultural and homestead land tenure among PVTGs and other landless households? (3) Does the program contribute to higher household income, per capita expenditure, asset accumulation, and reduced poverty among beneficiaries? (4) Does the program improve child nutrition, contribute to diversified nutritious diets (pulses, milk, fish, eggs, fruits, and vegetables), and improve food security among beneficiaries? (5) Does the program increase child school enrolment among beneficiary households? (6) Does the program contribute to higher market sales and better market access? (7) Does the program contribute to improvements in gender balance, women’s empowerment, and intra-household decision-making? (8) Does the program improve technology adoption and use of complementary inputs (fertilizer, pesticide, and other improved crops and agricultural practices) among program beneficiaries? (9) Does the program increase income diversification, livelihood opportunities, and resilience to negative exogenous shocks among beneficiaries? A large body of literature has attempted to answer these questions in different geographies and contexts, yet rigorous evidence on the impacts of agricultural development interventions on livelihood improvements of tribal communities hardly exists. While some studies have rigorously assessed the impacts of agricultural interventions on nutritional outcomes (Pandey et al., 2016; Ruel et al., 2018; Sharma et al., 2021), agricultural productivity (Davis et al., 2012; Martey et al., 2021), food and nutrition security (Banerjee et al., 2015; Fiorella et al., 2016; Jodlowski et al., 2016), women’s empowerment and intra-household decision-making (Alkire et al., 2013; de Brauw et., 2014; Kafle et., 2016), poverty reduction (Cahyadi & Waibel, 2016; Ogutu et al., 2020a), asset accumulation (Dillon, 2011; Muriithi & Matz, 2015), and the adoption of improved agricultural technologies (Nakano et al., 2018; Ogutu et al., 2020b), the evidence base on the impacts of agricultural interventions on these outcomes is still thin in general and non-existent for tribal communities. In addition, the existing literature shows limited evidence of the impacts of agricultural interventions on child school enrolment, women’s empowerment and intra-household decision-making, resilience, and income diversification (IFAD, 2017). Concrete evidence regarding the impact of land titling certificates, agricultural productivity, and income growth is also mixed. Lawry et al. (2017) find a positive causal impact between the land tenure system and agricultural productivity and income, but the impacts differed significantly by region. The midline assessment of OPELIP aims to address these knowledge gaps in the literature by answering the above questions. 14 2.3 Unintended program impacts As mentioned, MPAs are the implementing units for OPELIP. OPELIP is designed to cover all villages within MPAs as well as all other villages outside MPA areas, but within the same Gram Panchayat (GP)1 – a GP saturation approach (IFAD, 2017). Since OPELIP’s targeting includes all households within a GP with PVTGs, the program may lead to spillover or unintended effects on non-target households. For instance, positive spillover effects may occur if program beneficiaries share agricultural knowledge received from the program with non-beneficiaries. This could lead to increased adoption of superior technologies and higher yields among both beneficiaries and non-beneficiaries. Higher yields may also result in lower local food prices and improved food security for beneficiaries and non-beneficiaries. Such spillover effects, if present, can be very difficult to measure and would lead to a downward bias (underestimation) of the estimated program effects (comparing intervention versus non-intervention villages or groups). Reliable program estimates can only be obtained by selecting a control group that is not contaminated or affected by the program (Angelucci & Di Maro, 2016). Although we expect positive impacts from the program, some undesirable impacts may also arise. For instance, increased agricultural productivity and farm incomes may increase the demand for farm labor. This may, in turn, lead to increased demand for on-farm child labor, particularly in settings where labor markets are imperfect. This implies that agricultural transformation efforts in such contexts may be achieved at the expense of children’s schooling. A study by Kafle et al. (2018) showed a negative (adverse) association between increases in agricultural assets and child education through increased demand for child labor. Hence, it may be possible that children in beneficiary households may drop out of school or register poor grades as a result of increased demand for child labor. 2.4 Target population OPELIP is implemented in 12 districts in the state of Odisha, which has the largest number of PVTGs in India. In contrast to OTELP, which targeted tribal groups, OPELIP targets the most vulnerable and ultra-poor tribal groups, who generally live in remote areas with a high prevalence of poverty (IFAD, 2014). Within the 12 districts of Odisha, OPELIP is implemented in 17 micro-project areas, which are covered by a micro-project agency (MPA). MPAs are government entities that were formed in the late 1970s. They are mandated with the responsibility of implementing special programs targeting PVTGs based on their cultural practices and beliefs. The 17 MPAs cover all 13 PVTGs in Odisha and serve as the implementing units for OPELIP. Across the 17 MPAs, OPELIP covers 84 GPs (basic village- governing or administrative units in India), 1,243 villages, and 62,356 households. The program is implemented in all villages covered by the MPAs (664 villages) and in all other villages that are outside MPA areas (579 villages) but within the same GP. There are no additional eligibility criteria, apart from the requirement to be in a GP with PVTGs. The 12 districts covered by the program are depicted in Figure 2, and the administrative hierarchy of the program coverage area is shown in Figure 3. 1 A Gram Panchayat (GP), also called village council, is a local self-governing body formed by local residents. 15 Figure 2: The 12 districts covered by OPELIP Source: IFAD (2014, p. vi) Figure 3: Diagram of program coverage areas Source: Modified from IFAD (2017). Notes: MPA is a Micro-project agency and GP is Gram Panchayat. 19 Blocks Program coverage 12 Districts 17 MPAs 84 GPs 1,243 Villages 664 MPA Villages 579 Non-MPA Villages MPA Households Non-MPA Households 16 3. Impact assessment design 3.1 Construction of counterfactual groups We use the same counterfactual groups/villages that were used during the baseline survey, which was undertaken in 2017 since we are revisiting the same households (IFAD, 2017). Finding a suitable counterfactual group can be very challenging. For instance, using villages or households outside the program coverage area may be problematic as the villages may be more likely to differ from program villages in observed and unobserved characteristics. On the other hand, using households within a program area may lead to using counterfactual households that are also affected by the program. To address this challenge, the baseline survey used propensity score matching to construct counterfactual groups/villages. The approach involved finding a pool of ineligible villages outside the program areas and “matching” them on similar observed characteristics as program villages. Using propensity score matching with the three nearest neighbors matched with replacement and the caliper length set at 0.1, counterfactual villages were drawn from outside the program area but within the same program district and Block to ensure geographical similarities between program and comparison villages.2 This was done using 2011 census data. In particular, program and non-program villages were matched based on 20 village-level variables, which included household demographics, asset ownership, and living conditions as shown in Table 1. Due to data limitations, agricultural variables were not used. Agricultural data were only available at a higher level (the Block level) which could not be used for matching since the program is implemented at the GP level and mostly covers no more than one Block. All the variables shown in Table 1 are village-level averages. Table 1. Village-level variables used in the propensity score matching of non-program villages Matching variables Household size Source of drinking water Home ownership Source of energy Living conditions of the dwelling Access to sanitation facility Quality of roof materials Use of cooking fuel Quality of wall materials Access to banking services Quality of floor materials Asset ownership (transportation) Number of rooms in the house Asset ownership (information) Quality of drinking water Bicycle ownership Number of farming households in the village Literacy rate Proportion of agricultural labor in the household Unemployment ratio Notes: IFAD (2017). Data were obtained from a publicly available database from the 2011 India Census. 2 A Block or Tehsil is an administrative unit that is smaller than a district, but larger than a GP. 17 3.2 Sample size calculations We revisit the same households which were interviewed at the baseline survey conducted in 2017 (IFAD, 2017). Since the selection of an adequate sample size is a very crucial part of a successful impact assessment, the sample size was carefully selected at the baseline to ensure that the assessment could reliably measure the impacts of OPELIP. At the baseline, the expected change in the outcome of interest was considered more important than other factors in the sample size determination. The sample size determination strategy relied on a method developed by the World Bank that incorporates the expected minimum change in the outcome variable, its standard deviation, the critical values of the confidence interval, statistical power, and the minimum number of units to be sampled within each cluster (World Bank, 2007; Winters et al., 2010). Following this method, the sample size was determined as follows: 𝑁 = 4 𝜎2(𝑍𝛼 + 𝑍𝛽)2 𝐷2 [1 + 𝜌(𝑚 − 1)] , (1) where 𝜎 is the standard deviation of the outcome variable, 𝑍𝛼 is the critical value of the confidence interval, 𝑍𝛽 is the critical value of the statistical power, 𝐷 is the minimum expected change in the average outcome variable, 𝜌 is the intra-cluster correlation of the unit of analysis, and 𝑚 is the number of units to be sampled within each cluster. The standard deviation (SD), and minimum expected change (𝐷) are presented in Table 4. It was assumed that the analysis would have 80% statistical power and a 95% confidence level, so 𝑍𝛽 = 1.96, and 𝑍𝛼 = 1.28. At least 15 units of observation per cluster (𝑚) and an intra-cluster correlation of 0.05 was also assumed for the baseline sample computation (IFAD, 2017). We summarize the details of the sample size calculation at the baseline for this impact assessment in Table 2. While there are several outcome variables of interest for OPELIP, the sample size calculation at baseline was based on only four variables – food security index, cereal productivity, rice productivity, and proportion of irrigated area – due to data limitations. Again, due to data limitations, the sample size computations were done based on district-level averages of the outcome variables rather than at the program implementation unit (GP) or village-level. Therefore, the minimum expected changes are relatively small as these changes are for the whole district. Using equation (1) and adjusting the sample size for a 10% margin of error, the largest sample size required to achieve the minimum expected change in each of the outcome variables is presented in Table 2. The largest sample size (2,096) among the four indicators was chosen to guarantee sufficient power for all outcome variables. The total sample was split equally into two to obtain a sample size of 1,048 households each for the program beneficiaries and non-beneficiary households. 18 Table 2. Sample size calculation Outcome Average of 12 districts SD Minimum expected change D N 1.1*N Food security index 0.39 0.10 5% 0.02 1,905 2,096 Cereal productivity (kg/ha) 1,404 555.69 10% 140.4 1,118 1,230 Rice productivity (kg/ha) 1,439.58 642.13 10% 143.9 1,420 1,562 Irrigated area (%) 27.72 12.86 10% 2.8 1,537 1,691 Calculated sample size 1,905 2,096 Notes: SD is the standard deviation, D is the minimum expected change in the average of the outcome variable, and N is the sample size. The minimum expected changes are relatively small because we use district averages, but the program coverage area is only a fraction of each district. Because these expected changes are for the districts, we expect a higher level of minimum changes in the project area. Source: IFAD (2017, p.18). 3.3 Sampling strategy As mentioned, we interviewed the same households surveyed in 2017 in the baseline survey. The sampling strategy used for the baseline survey is as follows. At the baseline, a two-stage proportional sampling procedure was used to generate the sample. In the first stage, proportional stratified sampling was used to determine the number of households to be selected from each of the 17 MPAs covered by the program. This aimed at ensuring that there is a proportional representation of households from all the 17 MPAs and also sought to capture variations in geographical and agro-climatic conditions in each of the MPAs. Sample size calculation showed that 1,048 program households – 1.68% of 62,356 beneficiary households – would be required for the baseline. Thus, 1.68% of the households in each of the 17 MPAs covered by the program were randomly selected. In the second stage, the sample of 1,048 households to be covered by the program was divided by 12 (the minimum number of households set to be interviewed in each village) generating a total of 87 treatment villages that were randomly selected for the survey. Details of OPELIP’s sampling frame are shown in Table 3. The final sample includes 1.68% of all beneficiaries from each of the 17 MPAs to sufficiently capture differences in MPA characteristics. As mentioned, there were no MPA areas associated with the comparison group (non-program villages). Hence, 1,048 comparison/control households were selected from outside program areas, but within the same District and Block. Control households were sampled from control villages that were selected based on a propensity score matching procedure described in subsection 3.1.3 A total of 87 villages were also sampled for the comparison group. Therefore, treatment households are drawn from 87 treated villages (villages covered by OPELIP), while control households are drawn from 87 control villages (villages not covered by the program), according to the baseline survey. 3 From each MPA, three sets of program villages were generated by randomly selecting the required number of villages three times. Program villages in each random set were accompanied by three potential control villages as determined in subsection 3.1, using the propensity score matching approach. For each randomly selected program village, the best matched control village was identified and validated with help from PMU and MPA-level staff. 19 Table 3. Sampling frame for program beneficiaries: Districts, MPAs, villages, and households District MPA No. of GPs in MPA No. of Villages in MPA No. of HHs in MPA Final sample villages and HHs 1. Anugul PBDA, Jamardihi 4 43 2,482 4 42 2. Debgarh PBDA, Rugudakudar 8 94 7,300 10 123 3. Gajapati LSDA, Serango 3 34 1,925 3 32 SDA, Chandragiri 10 121 5,012 7 84 4. Ganjam TDA, Tumba 3 110 2,088 3 35 5. Kalahandi KKDA, Lanjigarh 3 62 3,092 4 52 6. Kandhamal KKDA, Belghar 3 63 1,568 2 26 7. Keonjhor JDA, Gonasika 6 57 3,500 5 59 8. Malkangiri BDA, Mudulipada 4 56 2,297 3 39 DDA, Kudumuluguma 4 79 3,462 5 58 9. Mayuribhanj HKMDA, Jashipur 12 156 9,907 14 167 LDA, Morada 8 105 7,993 11 134 10. Nuapada CBDA, Sunabeda 3 31 1,621 2 27 11. Rayagda DKDA, Chatikona 5 124 4,637 7 78 DKDA, Parsali 2 48 1,276 2 22 LSDA, Puttasing 1 8 720 1 12 12. Sundargarh PBDA, Khuntugaon 5 52 3,476 5 58 Total: 12 Districts 17 MPAs 84 GPs 1,243 villages 62,356 HHs 88 villages 1,048 HHs Notes: There are equal number of counterfactual villages and households in each district. PBDA, Poudi Bhuyan Development Agency; LSDA, Lanjia Saora Development Agency; SDA, Saora Development Agency; TDA, Tumba Development Agency; KKDA, Kutiakandha Development Agency; JDA, Juang Development Agency; BDA, Banda Development Agency; DDA, Didayi Development Agency; HKMDA, Hill Khadia & Mankirdia Development Agency; LDA, Lodha Development Agency; CBDA, Chuktia Bhujia Development Agency; and DKDA, Dangria Kandha Development Agency. Source: IFAD (2017, p.19). 3.4 Data collection Two rounds of quantitative survey data were collected from smallholder farm households in Odisha, India. The baseline round was collected by the Academy of Management Studies (AMS) in 2017 before the roll-out of the program and covers 2,099 farm households. The midline round was collected by HDI in collaboration with the Alliance of Bioversity International and CIAT, and ICRISAT between October 2021 and January 2022, about four years into the program. To address possible issues of seasonality, the surveys covered the one-year period preceding the surveys in both survey rounds, so that all 12 months in a year were covered by the data. Due to sample attrition (of about 8.5% of households, evenly distributed between treatment and control groups), the follow-up survey round includes observations from 1,921 farm households. In the two survey rounds, data were collected from the sampled households on household demographics, migration, access to credit, group membership, asset ownership, agricultural production, marketing, off-farm income sources, food consumption, anthropometric indicators, and broader socioeconomic and institutional characteristics through face-to-face interviews conducted in the local language by trained enumerators. In 20 the midline round, additional data were collected on dietary quality indicators and adult anthropometry. 3.5 Identification strategy This study aims to evaluate the impacts of OPELIP on key development indicators. A successful impact assessment requires a good identification strategy to identify the causal effects of the intervention by eliminating potential confounding factors. Since OPELIP’s placement is not random, identification of its effects can be more challenging due to potential endogeneity arising from its purposive placement. We use panel data which allows us to observe the study households across time and evaluate the program’s effects using panel data regression models. In particular, we use the difference- in-differences (DID) estimator to account for time-invariant confounding factors (Greene, 2012). However, since we introduced some new indicators in the midline (midline) survey, we only have cross-sectional data for new indicators. Under such circumstances, the DID approach cannot be used. Hence, we use the Propensity Score Matching (PSM) to analyze the impacts of the program on the new indicators for which panel data is not available. We now describe the DID approach, which is our main estimation approach. Assuming that 𝑌 is the outcome of interest and 𝑋 is the vector of control variables, we can estimate the impact of the project (𝑃) using the DID approach as follows: 𝑌𝑖𝑡 = 𝛽0 + 𝛽1𝑇𝑖𝑚𝑒𝑡 + 𝛽2𝑃𝑡 + 𝛽3(𝑇𝑖𝑚𝑒𝑡 ∗ 𝑃𝑡) + 𝜗𝑋 + 𝛼𝐷 + 𝜀𝑔𝑡 , (1) where 𝛽0 denotes the constant, 𝛽1 captures the time trend or unobserved time effects, and 𝛽2 represents the group effects or unobserved effects among program beneficiaries relative to non-beneficiaries in absence of the program. The coefficient on the interaction term 𝛽3 is our parameter of interest, which estimates the effects of OPELIP. The DID estimator 𝛽3 relies on the parallel or common trends assumption, which postulates that in the absence of an intervention (treatment), the difference between the 'treatment' and 'control' group is constant over time (Angrist & Pischke, 2008; Cameron & Trivedi, 2005). 𝜗 is the coefficient for the vector of control variables, 𝛼 is the coefficient for the vector of district dummy variables interacted with time to control for possible unobserved time-varying district differences, and 𝜀𝑔𝑡 is the error term. Subscripts 𝑔 and 𝑡 denote group and time, respectively. As described above, we employ PSM to evaluate the impact of the program on new indicators, such as adult anthropometric indicators, minimum dietary diversity (MDD), MDD for women (MDD-W) and MDD for men (MDD-M). The PSM method is a non-parametric estimation technique that does not depend on the functional form or distributional assumptions. The basic idea of the PSM method is to match observations of program participants and non-participants according to the predicted probability of program participation. The PSM assumes the common support condition, which requires substantial overlap in covariates between program participants and non-participants, such that households that are being compared have a common probability of being both participants and non-participants. Hence, PSM allows for the comparison of outcomes of participants and non-participants with similar observed characteristics (for more detail, see Rosebaum and Rubin (1983)). For the purpose of robustness and comparison, we employ nearest neighbor and kernel matching methods. 21 4. Results In this section, we present descriptive statistics of various outcome variables, differentiated by program beneficiary (treatment) and non-beneficiary (control) household, as well as PVTG and non-PVTG households, following the same format as the baseline evaluation. We also present and discuss the estimation results of selected outcome variables. As mentioned, the selected indicators are informed by the impact assessment questions identified in subsection 2.2, IFAD 11 indicators shown in Table A1 in the Appendix, and the baseline’s Impact Assessment Report (IFAD, 2017). All results, except indicated, are based on panel data from the baseline and midline surveys. Incomes, production, sales values, and remittances reported in the midline survey were deflated to baseline (the year 2017) figures using the consumer price index (CPI) for India for the period 2017-2021 for appropriate comparison with baseline figures.4 4.1 Descriptive results 4.1.1 Distribution of tribal groups across the treatment and control samples Table 4 presents the distribution of tribal groups across the treatment and control samples at baseline and midline. As mentioned, the midline sample is slightly smaller (N=1,921) than the baseline (N=2,099) sample due to attrition, which is about 8.5% overall, and distributed evenly between treatment and control groups. Results show that PVTG households, which are of particular interest, make up about one-third of the total baseline and midline samples. Larger numbers (shares) of PVTG households are found in the treatment group in both survey rounds, while there are fewer numbers (shares) of ST households. This is expected since OPELIP targets PVTG households. However, in spite of the priority given to PVTGs under this program, more than one-third of PVTG households remain outside OPELIP program coverage areas, potentially due to variation in the households’ own classification of their tribal grouping from official records as reported by the baseline survey (IFAD, 2017). To prevent inconsistencies across the surveys, the same classification of the household’s tribal group or caste used at baseline was used for this midline survey. The distribution of the tribal groups in our sample differs significantly from the Odisha state-level distribution because OPELIP focused on the 12 districts with the largest population of PVTGs. Overall, in both survey rounds, other STs comprise the largest share of the total sample, followed by PVTGs, SCs, and other tribes. 4 Year 2021 figures are deflated using CPI deflator (0.824) computed using the CPI from the World Bank (2022). 22 Table 4. Distribution of tribal groups across treatment and control samples Tribal group 2017 2021 Treatment Control Treatment Control Particularly vulnerable tribes (PVTG) 449 255 408 228 (42.8) (24.3) (42.4) (23.8) Other Scheduled tribes (STs) 399 589 365 533 (38.0) (56.2) (37.9) (55.6) Scheduled castes (SC) 46 71 44 67 (4.38) (6.77) (4.57) (6.99) Other tribes 156 134 145 131 (14.9) (12.8) (15.1) (13.7) Total 1,050 1,049 962 959 (100) (100) (100) (100) Notes: Numbers represent the sample size. Numbers in parentheses are column percentages. 4.1.2 Household demographics Table 5 presents summary statistics of selected variables covering household demographics of the sample households. Panel A of Table 5 shows that, on average, household heads were about 46 years old at baseline and 48 years old in the midline survey. Household heads also had about 2.5 years of education at baseline and three years of education in the midline survey, with the changes possibly occurring due to changes in the composition of household heads across the two survey rounds. About 10% and 14% of the households are headed by women at baseline and midline samples, respectively, and over 80% of household heads are married in both survey rounds. A comparison of these variables between the treatment and control groups does not reveal any significant differences between the groups in the two survey rounds. Panel B of Table 5 presents summary statistics of household membership and composition. In both survey rounds, the results show that, on average, households have about five members, between one and two members of the household are children (less than 14 years old) or youth (15–29 years old),5 or belong to the working-age group (30–64 years old). Households have relatively fewer members that are at least 65 years of age, with an average of 0.23 individuals in this age group per household. This is consistent with the low dependency ratio that shows that, on average, each household has about one dependent (dependency ratio of 0.6). The results also show that at least 50% of the adult population is literate, i.e., able to read and write in at least one language, in both survey rounds. The level of literacy actually increased in the period between the survey rounds (about 7 percentage points), although the difference in literacy rates between the treatment and control groups remained about the same (1.5–1.7%). 5 This definition is according to the definition of the government of Odisha. However, it is noteworthy that according to the United Nations, “youth” consists of individuals between the age of 15 and 24 years (United Nations, 2004). 23 Table 5. Statistics of selected household variables by treatment status 2017 2021 Panel A: Household head characteristics Treatment Control Treatment Control Age of head (years) 46.2 46.4 48.3 48.8 (13.1) (13.0) (13.0) (12.8) Education of head (years) 2.58 2.53 3.00 2.99 (3.89) (3.78) (4.01) (3.96) Female head (%) 10.2 9.63 13.6 13.6 (30.3) (29.5) (34.3) (34.2) Married head (%) 86.7 85.8 82.2 82.3 (34.0) (34.9) (38.3) (38.2) Panel B: Household characteristics Household size (count) 4.83 4.78 4.69 4.58 (1.87) (1.75) (1.85) (1.76) Number of children below 14 1.49 1.43 1.36** 1.20 (1.37) (1.34) (1.31) (1.26) Number of youth 15-29 1.40 1.39 1.33 1.38 (1.20) (1.18) (1.19) (1.23) Number of adults 30-64 1.73 1.75 1.76 1.76 (0.85) (0.86) (0.84) (0.81) Number of adults 65 and over 0.20 0.22 0.25 0.24 (0.46) (0.48) (0.52) (0.52) Dependency ratio 0.66 0.65 0.63*** 0.56 (0.64) (0.64) (0.65) (0.61) Literacy rate (%) 50.9 52.4 57.2 59.0 (30.8) (29.5) (30.2) (29.0) Observations 1,050 1,049 962 959 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p <0.10, **p < 0.05, and ***p <0.01. 4.1.3 Housing characteristics Table 6 presents a summary of sample households’ dwelling characteristics. In the baseline, almost all respondents (> 98%) owned their homes, which, consistently, had an average of about two rooms. This number increased to > 99% in the midline. A bit over half of the households’ dwellings had improved roofing material, meaning they had corrugated iron sheet, concrete or cement, or a brick or asbestos roof at the baseline. This number increased to around 60 percent in the midline, with a larger share (65.6%) of households in the treatment group having improved roofing materials compared to the control group (61.6%). Around 35% of both treatment and control households had improved wall material (cement, bricks, or blocks) at baseline and around 20–25% had improved floor material (cement, tiles, or wood planks). The share of households with improved walls and improved floor materials increased substantially, to around 43–46% for both groups, by midline. 24 Table 6. Housing characteristics by treatment status 2017 2021 Treatment Control Treatment Control Home ownership (%) 98.0* 99.0 99.9 99.8 (14.01) (10.19) (3.22) (4.56) Number of rooms 2.32 2.33 2.22 2.14 (1.140) (1.044) (1.17) (1.12) Improved roof material (%) 55.4 52.9 65.6* 61.6 (49.73) (49.94) (47.53) (48.65) Improved wall material (%) 36.1 34.6 46.0 46.3 (48.05) (47.59) (49.87) (49.89) Improved floor material (%) 25.9* 22.5 46.5 43.5 (43.83) (41.78) (49.90) (49.60) Source of water >30 minutes, roundtrip (%) 18.3 15.7 8.73** 5.94 (38.67) (36.43) (28.24) (23.66) Access to safe drinking water (%) 74.3 75.8 79.2*** 84.0 (43.73) (42.86) (40.60) (36.64) Access to improved sources of energy (%) 69.5*** 78.2 94.0 92.6 (46.1) (41.3) (23.8) (26.2) Access to improved toilet (%) 27.9 27.8 39.7 39.1 (44.87) (44.84) (48.95) (48.82) Traditional cooking fuel use (%) 96.2*** 97.7 90.1 91.1 (19.15) (14.96) (29.85) (28.44) Observations 1,050 1,049 962 959 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p <0.01. Less than 20% of households were located fairly far – more than 30 minutes (round trip) – from a source of water at baseline, while the share dropped to less than 10% at midline. This may suggest that the large majority of households now have relatively good access to water. Access improved by about 10 percentage points for both treatment and control groups and so the control group continued to have slightly better access than the treatment group. This is unexpected given that under the community empowerment component, OPELIP aims to improve access to drinking water and sanitation by connecting households through overhead tanks, either filled through gravity where a water source is available upstream or by pumping water from a tube well or stream if electricity is available (IFAD, 2014). Overall, more than 75% of the households have access to safe drinking water, which suggests that access to water is not a major problem, even though a smaller share of treatment households has access to clean drinking water compared to control households in the midline sample. The results of Table 6 also show that more than two-thirds of treatment households have access to improved sources of energy (e.g., solar, electricity) for home use (lighting and cooking) at baseline, as do more than three-quarters of control households. In the midline survey, access to improved sources of energy among both groups increased, with treatment households’ access now exceeding that of control households (94% to 93%, respectively). At 25 the same time, although many households have access to improved sources of energy, more than 90% of all households still use traditional sources of energy, such as wood and dung, although this has declined from around 96–97% of households at baseline. Lastly, only about 39% of the households in both groups have access to improved toilets, although this has increased from about 28% for both groups at baseline. This points to some improvement, but also to continuing sanitation challenges. Overall, a comparison of the households’ dwelling conditions at baseline and midline shows improvements in access to water, quality of materials of roofs, walls, and floors, as well as access to improved energy sources and toilets, although generally, the improvements are similar for both treatment and control households, with the exceptions noted above for access to safe drinking water, roofing material and improved energy sources. 4.1.4 Schooling outcomes Table 7 shows summary statistics of schooling outcomes among sample households by program participation. The outcomes mainly focus on the highest level of education attained by the household head and school-aged boys and girls currently attending school. On average, household heads have very low levels of education, about 2.5 years of education at baseline and about 3 years in the midline survey. Table 7. School enrolment by treatment status 2017 2021 Treatment Control Treatment Control Household level outcomes Highest grade completed by head 2.58 2.53 3.00 2.99 (3.89) (3.78) (4.01) (3.96) Number of school-age children currently attending school 1.02 1.08 0.80 0.78 (1.17) (1.15) (1.08) (1.06) Proportion of children (aged 5–18) attending school (%) 0.63** 0.67 0.54* 0.59 (0.40) (0.39) (0.44) (0.43) Share of boys currently attending school (%) 0.65** 0.71 0.53** 0.59 (0.43) (0.41) (0.47) (0.47) Share of girls currently attending school (%) 0.63 0.66 0.57 0.59 (0.44) (0.43) (0.47) (0.46) Observations 1,050 1,049 962 959 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p< 0.10, **p< 0.05, and ***p<0.01. Results of Table 7 also show that on average, a household has one child attending school, with between 54–67% of school-aged children attending school, with the share of boys declining proportionately more than that of girls. A significantly smaller share of children in treatment households is attending school compared to control households in the two survey rounds, with a lower share of boys than girls attending school among treatment households. 26 Apart from comparing treatment and control households, we also compare schooling outcomes by PVTG and non-PVTG households as shown in Table 8. In both survey rounds, PVTG households have significantly lower household head education and a lower share of school-aged children (both boys and girls) attending school compared to non-PVTG households. As mentioned, improvements in the household head’s education in the midline survey could be due to changes in the composition of the household head as we do not expect the program to support a return of adults to the school, but enrollment of children in school could definitely change. Overall, shares of school-attending children seem smaller in the midline sample compared to the baseline sample, perhaps due to the effects of the COVID- 19 pandemic. Table 8. Schooling outcomes by tribal group 2017 2021 PVTG Non-PVTG PVTG Non-PVTG Household level outcomes Highest grade completed by head 1.46*** 3.15 2.28*** 3.36 (2.89) (4.15) (3.51) (4.16) Number of school-age children currently attending school 1.00 1.07 0.85 0.76 (1.14) (1.17) (1.10) (1.06) Proportion of children (aged 5–18) attending school (%) 0.57*** 0.70 0.54* 0.58 (0.41) (0.38) (0.44) (0.44) Share of boys currently attending school (%) 0.58*** 0.74 0.52** 0.58 (0.45) (0.40) (0.47) (0.47) Share of girls currently attending school (%) 0.57*** 0.69 0.56 0.59 (0.44) (0.43) (0.46) (0.47) Observations 704 1,395 636 1,285 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between PVTG and non-PVTG in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p <0.01. 4.1.5 Migration and remittances PVTGs are typically characterized by a lifestyle of migration in search of better farmland and livelihoods (IFAD, 2017). Against this background, Table 9 shows a summary of migration and remittance statistics by treatment status. Migration entails both local and regional movement of individuals out of the household for more than one continuous month in the last 12 months before each survey round rather than temporary movement out of the household to hunt and gather food. Panel A shows that on, average, about 27% and 24% of households have a migrant at baseline and midline samples, respectively. At baseline, the households have an average of one migrant, but in the midline survey, the number increases to five migrants among households with migrants. Around 40% of the households received remittances ranging between 9,500 to 14,500 rupees/year (panel B), on average. Although the number of migrants is higher in the midline survey, it has not contributed to increased remittances to the households, possibly due to the COVID-19 pandemic, which significantly reduced jobs and economic opportunities for migrant workers across India’s major cities (Sahoo et al., 2022). A comparison of the migration and remittance variables between treatment and control groups 27 does not reveal any significant differences between the groups in the baseline and midline samples. Table 9. Migration and remittances by treatment 2017 2021 Panel A: Migration outcomes Treatment Control Treatment Control Household has a migrant (%) 27.2 27.4 23.8 24.3 (44.5) (44.6) (42.6) (42.9) Observations 1,050 1,049 962 959 If household has a migrant … Number of migrants 1.34 1.34 5.22 5.15 (0.63) (0.63) (1.74) (1.70) Panel B: Remittance outcomes Received remittances last 12 months (%) 42.0 40.8 41.5 44.2 (49.4) (49.2) (49.4) (49.8) Amount of remittances (rupees/year) 14,474 12,042 9,552.9 11,158.9 (36,926) (22,670) (19,906.2) (31,236.0) Observations 286 287 229 233 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p <0.01. 1 USD = 65.1 INR (Indian rupees) in 2017, the reference/base year.6 Table 10 compares migration and remittance statistics between PVTG and non-PVTG households. About 27% of PVTG households have a migrant at baseline and 24% at the midline. Out of the households with migrants, the average number of migrants is 1.34 at baseline, but the number increases to five in the midline sample, perhaps due to COVID-19 as mentioned above. The share of PVTG households with a migrant and the number of their migrants do not differ significantly from those of non-PVTG households in the two survey rounds. About 40% of the sample received remittances, ranging between 10,000 to 15,000 rupees/year, depending on the sub-sample. As shown, PVTG and non-PVTG households do not differ in their migration and remittance characteristics in the two survey periods, except for the amount of remittances received at baseline, which is significantly lower for PVTG households compared to non-PVTG households. 6 https://www.exchangerates.org.uk/USD-INR-spot-exchange-rates-history-2017.html 28 Table 10. Migration and remittances across tribal groups 2017 2021 Migration outcomes PVTG Non-PVTG PVTG Non-PVTG Household has a migrant (%) 27.9 26.9 23.3 24.4 (44.9) (44.4) (42.3) (42.9) Observations 704 1,395 636 1,285 If household has a migrant… Number of migrants 1.33 1.34 5.31 5.12 (0.60) (0.64) (1.78) (1.70) Received remittances last 12 months (%) 38.6 42.8 42.6 43.0 (48.8) (49.5) (49.6) (49.6) Amount of remittances (rupees/year) 9,980* 14,972 9,876 10,592 (19,406) (35,001) (36,530) (19,644) Observations 197 376 148 314 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between PVTG and non-PVTG in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p <0.01. 1 USD = 65.1 INR (Indian rupees) in 2017, the reference/base year. 4.1.6 Access to credit We measure access to credit using outcomes for loan application, approval, amount, and use. Under the community empowerment component, OPELIP aims to promote rural finance and savings through SHGs where individuals receive training on rural finance and saving. Hence, SHGs could play a crucial role in promoting access to credit among treatment households. Table 11 presents summary statistics for access to credit by program participation at baseline and midline. Results of Table 11 show that a significantly smaller share of treatment households applied for a loan compared to control households at baseline (17% versus 20%, respectively). The initial low rates of loan application observed are expected because most of the sample households are located in remote rural areas with limited access to financial institutions. In the midline survey, the share of households that applied for loans does not differ significantly between treatment and control households. However, we see that the percentage of treatment households that applied for a loan increased by over 5 percentage points, versus a bit over 1 percentage point for the control group. While about one of every five households in the sample applied for a loan, the approval rates of applications were quite high. In fact, almost all loans were approved. Furthermore, while over 90% of applicants were approved even at baseline (94% of treatment applicants and 97% of control applicants), these numbers increased to 98% for both groups at midline. Loan amounts averaged about 17,000 rupees per year, for both groups at baseline, but increased to about 18,000 rupees for the treatment households and 19,000 rupees for the control households. SHGs were the primary source of loans for about 30% of treatment households and 20% of control households at baseline, but accounted for well over 50% of loans in the midline sample (59% and 55% for the treatment and control households, respectively), suggesting a growing role of SHGs in the provision of loans. A significantly larger share of the treatment group (33%) used loans for consumption expenditures compared to the control 29 Table 11. Access to credit by treatment status 2017 2021 In the last 12 months Treatment Control Treatment Control Household applied for loan (%) 16.6** 20.0 22.3 21.4 (37.2) (40.0) (41.6) (41.0) Observations 1050 1049 962 959 If applied for loan… Loan approved (%) 94.3 96.7 98.1 98.0 (23.3) (18.0) (13.6) (13.9) Amount of loan (rupees/year) 17,311 16,505 17,839 19,080 (18,736) (14,773) (19,097) (18,529) Primary source of loan (1 = SHG†, 0 = other) 28.2 21.4 59.3 55.1 (45.1) (41.1) (49.2) (49.9) Primary use of loan (1 = consumption, 0 = other) 33.3** 22.9 11.7 8.29 (47.3) (42.1) (32.2) (27.6) Observations 174 210 214 205 Notes: Point estimates are sample means. Standard deviations are in parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. †SHG stands for self-help groups. 1 USD = 65.1 INR (Indian rupees) in 2017, the reference/base year. group (23%) at baseline, but the difference was not significant in the midline sample (12% versus 8%, respectively). Use of loans for consumption expenditures fell quite significantly for both groups at midline. The share of the households that applied for loans, the percent of loans approved, the amounts received, and the primary source of loans do not differ significantly between treatment and control households in both survey rounds. Table 12 presents summary statistics of access to credit differentiated by tribal groups. The differences in loan application rates between PVTG and non-PVTG households are insignificant at baseline, but in the midline survey, a smaller share of PVTG households applied for loans compared to non-PVTG households. A similar pattern is also observed with loan approval rates, where at baseline, the loan approval rates between PVTGs and non- PVTGs are statistically insignificant, but in the midline survey, a smaller share of PVTG households received loan approvals compared to non-PVTG households. These midline sample results are quite unexpected since part of OPELIP's objectives was to strengthen SHGs and promote rural finance and savings among the PVTGs. Table 12 also shows that the amount of loan approved does not differ in both baseline and midline samples, a smaller share of PVTG households rely on SHGs as their primary source of credit compared to non- PVTG households in both survey rounds, while PVTG and non-PVTG households do not differ in their usage of loans. 30 Table 12. Access to credit by tribal groups 2017 2021 In the last 12 months PVTG non-PVTG PVTG non-PVTG Household applied for loan (%) 19.0 17.9 17.6*** 23.9 (39.3) (38.4) (38.1) (42.6) Observations 704 1,395 636 1,285 If applied for loan… Loan approved (%) 94.8 96.0 95.5** 99.0 (22.3) (19.6) (20.7) (9.85) Amount of loan (rupees/year) 17,648 16,452 17,539.5 18,776.4 (19,866) (14,698) (19,431.8) (18,598.0) Primary source of loan (1 = SHG†, 0 = else) 11.9*** 31.2 48.2** 60.6 (32.5) (46.4) (50.2) (48.9) Primary use of loan (1 = consumption, 0 = other) 28.4 27.2 13.4 8.79 (45.2) (44.6) (34.2) (28.4) Observations 134 250 112 307 Notes: Point estimates are sample means. Standard deviations are in the parentheses. A two-sample t-test is used for the test of mean differences between PVTG and non-PVTG in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. †SHG stands for self-help groups. 1 USD = 65.1 INR (Indian rupees) in 2017, the reference/base year. 4.1.7 Land characteristics and farming practices Under the natural resource management, and food and nutrition security component, OPELIP aims to improve food and nutrition security through various activities that include improvement of land tenure and support for crop improvement. Against that background, Table 13 presents a summary of statistics on land characteristics and farming practices. Panel A shows that at least 75% of the households owned land in both survey rounds, with some increase in landownership by midline. Less than 1% of households leased land in both survey rounds. Women own at least one plot of land in about half of the households in both survey rounds, but the share dropped to 40–41% for both groups at midline. About 17% of the households had land titles at baseline, but in the midline sample, this share more than doubled for both groups (34–37%). Panel A of Table 13 also shows that at least 90% of households cultivated their land – about 1.5 acres at baseline and 2 acres in the midline sample – in the last 12 months. Comparison of the variables in panel A between the treatment and control groups does not reveal any significant differences between groups in either survey round. 31 Table 13. Access to land, land holding size, and farming practices by treatment 2017 2021 Land characteristics: Panel A Treatment Control Treatment Control Land ownership (%) 76.1 76.9 81.0 82.6 (42.7) (42.1) (39.3) (37.9) Household leases land (%) 0.48 0.38 0.21 0.00 (6.89) (6.17) (4.56) (0.00) Share of women owning land (%) 49.0 48.6 40.9 39.6 (50.0) (50.0) (49.2) (48.9) Land title, Patta (%) 17.2 16.1 33.9 36.5 (37.8) (36.8) (47.4) (48.2) Crop cultivation in the last 12 months (%) 92.2 92.2 90.2 90.3 (26.8) (26.9) (29.7) (29.6) Land holding size (acres) 1.57 1.42 2.05 2.06 (2.20) (2.34) (2.49) (4.64) Farming practices: Panel B Slash and burn (%) 8.38 6.58 3.74 2.61 (27.7) (24.8) (19.0) (15.9) Contour farming (%) 13.5 11.9 3.53** 1.77 (34.2) (32.4) (18.5) (13.2) Regular/settled agriculture (%) 76.9 78.2 86.9* 89.5 (42.2) (41.3) (33.8) (30.7) Number of crops grown (count) 0.85 0.88 1.10 1.15 (0.35) (0.33) (0.75) (0.80) Household irrigates a plot (%) 7.81** 5.53 45.4 46.3 (26.8) (22.9) (49.8) (49.9) Observations 1,050 1,049 962 959 Plot-level input use: Panel C High yielding seed used on plot (%) 10.5** 7.80 34.7 35.4 (30.7) (26.8) (47.6) (47.8) Chemical fertilizer used on plot (%) 44.4* 48.4 75.0* 72.1 (49.7) (50.0) (43.3) (44.8) Manure used on plot (%) 35.8 36.5 13.7*** 17.9 (48.0) (48.2) (34.4) (38.3) Herbicides/pesticides used on plot (%) 17.1* 13.6 18.2* 16.1 (37.6) (34.3) (38.6) (36.7) Observations 1,508 1,418 2,223 2,242 Notes: Point estimates are sample means. Standard deviations are in the parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. 32 Panel B of Table 13 shows a summary of the farming practices households were engaged in during the 12 preceding months. Less than 10% of farm households in both groups practiced slash and burn or shifting cultivation at baseline, while about 10% practiced contour farming. Shares of both practices fell to less than 5% in the midline survey, with a fairly substantial drop in use of contour farming (around a 10 percentage point drop in both groups). Over three- quarters of households in both groups practiced regular agriculture (settled agriculture) at baseline, with a slight increase to 87–90% in the midline sample. On average, households cultivate one crop, suggesting that sample farms are less diversified in their production. Less than 10% of the households irrigated at least one of their plots at baseline, with a larger share of treated households practicing irrigation compared to control households. In the midline sample, the share of households that irrigated a plot increased to at least 45% for both groups, but the differences between treatment and control households are statistically insignificant. Except for contour farming and regular agriculture, the differences in the use of other farming practices presented in panel B of Table 13 are statistically insignificant in the midline survey. Part of OPELIP’s objectives, under Component 2, is to improve food and nutrition security through support for crop diversification. However, the low average number of crops grown by sample households in the midline survey may imply that this objective has not yet been realized. In panel C of Table 13, we show the results of input use at the plot level. High yielding seeds were used on around 11% of the plots at baseline by the treatment groups and used somewhat less, 8%, for the control group. At midline, the two groups were about the same, at around 35%. Fertilizer was less likely to be applied on treated farm household plots compared to the control households’ plots at baseline (44% and 48%, respectively), but these numbers increased greatly by midline for both groups, with treatment groups now exceeding the control group (75% and 72%). Manure was used in around 35% of the plots of both groups at baseline, but in the midline sample manure use fell substantially for both groups (from 36% to 14% for the treatment group and from 37% to 16% for the control group). In both survey rounds, herbicides and pesticides were more likely used in treatment group plots than control group plots (17% versus 14% in 2017 and 18% versus 16% in 2021). Table 14 presents summary statistics of land characteristics and farming practices differentiated between PVTG and non-PVTG households. At baseline, a significantly smaller share of the PVTG households compared to non-PVTGs own land, lease land, have a land title, and cultivated crops in the 12 months prior to the survey. However, in the midline sample, differences between PVTG and non-PVTG households with respect to these variables are statistically insignificant, suggesting an improvement in the indicators, especially among PVTG households in the midline survey given that they were worse-off at baseline. 33 Table 14. Access to land, land holding size, and farming practices by tribal group 2017 2021 Land characteristics PVTG Non-PVTG PVTG Non-PVTG Land ownership (%) 74.0* 77.8 80.0 82.6 (43.9) (41.6) (40.0) (37.9) Household leases land (%) 0.0** 0.65 0.16 0.078 (0.0) (8.01) (3.97) (2.79) Share of women owning land (%) 49.9 48.3 36.8 41.9 (50.0) (49.9) (48.3) (49.4) Land title, Patta (%) 13.5*** 18.3 37.1 34.2 (34.2) (38.7) (48.3) (47.5) Crop cultivation in the last 12 months (%) 90.5** 93.0 91.5 89.6 (29.4) (25.4) (27.9) (30.5) Land holding size (acres) 1.60 1.44 1.91 2.13 (2.87) (1.91) (2.37) (4.23) Farming practices: Panel B Slash and burn (%) 13.1*** 4.66 3.46 3.04 (33.7) (21.1) (18.3) (17.2) Contour farming (%) 21.0*** 8.53 3.93** 2.02 (40.8) (27.9) (19.4) (14.1) Regular/settled agriculture (%) 66.8*** 82.9 87.3 88.6 (47.1) (37.6) (33.4) (31.7) Number of crops grown (count) 0.85 0.87 1.12 1.13 (0.36) (0.34) (0.71) (0.72) Plot irrigated (%) 8.95*** 5.52 56.6*** 40.5 (28.6) (22.8) (49.6) (49.1) Observations 704 1,395 636 1,285 Plot-level input use: Panel C High yielding seed used on plot (%) 7.31** 10.1 31.6*** 36.6 (26.1) (30.2) (46.5) (48.2) Fertilizer used on plot (%) 25.1*** 57.2 59.3*** 79.8 (43.4) (49.5) (49.1) (40.2) Manure used on plot (%) 44.3*** 32.0 23.0*** 12.6 (49.7) (46.7) (42.1) (33.2) Herbicides/pesticides used on plot (%) 13.8 16.1 16.0 17.6 (34.5) (36.8) (36.7) (38.1) Observations 1,002 1,924 1,399 3,066 Notes: Point estimates are sample means. Standard deviations are in the parentheses. A two-sample t-test is used for the test of mean differences between PVTG and non-PVTG in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < .01. 34 Panel B of Table 14 presents a summary of the farming practices sample households were engaged in during the 12 months before the survey. At the baseline, significantly larger shares of PVTG households compared to non-PVTG households practiced slash and burn (shifting cultivation), contour farming, or had irrigated a plot. However, a significantly smaller share of PVTG households practiced regular/settled agriculture compared to non-PVTG households at baseline. These results are plausible given that PVTG households tend to practice shifting cultivation and less regular/settled agriculture compared to non-PVTG households (IFAD 2014; IFAD, 2017). In the midline sample, significantly larger shares of PVTG households compared to non-PVTG households also practice contour farming and irrigation. The share of households with irrigated plots increased significantly from less than 10% to over 40% in the midline sample. Lastly, panel C of Table 14 shows a comparison of input use at the plot level between PVTG and non-PVTG households. As shown, high yielding seeds and fertilizer were less likely to be applied on PVTG household plots compared to non-PVTG household plots in both survey rounds. For both groups, herbicides and pesticides were used in 14–18% of the plots in both survey rounds. However, larger shares of PVTG household plots had manure applied compared to non-PVTG household plots in both survey rounds. 4.1.8 Women’s empowerment Under the community empowerment component, OPELIP aims to improve women’s empowerment and gender balance through the promotion of women SHGs, gender training, and rural finance services to enable the social development of SHG members. We use the same women’s empowerment index used in the baseline survey, which is comprised of a total of 23 intra-household decision-making questions that capture the dynamics of decision making in seven decision domains (IFAD, 2017). The decision domains are agricultural input use, use of cash income, crop and livestock sales, food items’ purchase, non-food items’ purchase, child school enrolment, and another domain which includes questions related to wage employment, group membership, and loan application. Women’s empowerment is measured using the proportion of women’s independent decisions over the total decisions made in the household, with all the decisions given equal weight (IFAD, 2017). We also compute the share of men’s independent and joint decisions over the total decisions made in the household for comparison. Table 15 presents a summary of the women’s empowerment measures by treatment status. Panel A shows that at baseline for both treatment and control groups about 10% of the household decisions were made independently by women, at least 85% of household decisions were made independently by men, and less than 3% of decisions were made jointly. In the midline sample, the share of solitary decisions made by women remains the same, but the proportion of decisions made independently by men decreased to about 63%, while the share of joint decisions increased substantially to about 24%. A breakdown of the share of solitary decisions made independently by women across the seven decision domains as shown in Panel B of Table 15 reveals the same decision-making trend observed when all decisions are aggregated. In all the decision domains, women make around 11% of the decisions independently and significant differences are not observed across treatment and control households in both survey rounds. Overall, these results show that men dominate decision-making and resource-control among the sample households. 35 Table 15. Proportions of women’s and men’s decisions by treatment Decisions 2017 2021 Panel A: Decision-making in all seven decision domains combined Treatment Control Treatment Control Women's solitary decisions 0.101 0.095 0.109 0.114 (0.301) (0.293) (0.308) (0.314) Men's solitary decisions 0.876 0.885 0.625 0.632 (0.327) (0.318) (0.452) (0.459) Joint decisions 0.022 0.020 0.237 0.236 (0.141) (0.137) (0.389) (0.399) Panel B: Share of solitary decisions by women in each of the seven decision domains Treatment Control Treatment Control Agricultural input use 0.101 0.095 0.113 0.117 (0.302) (0.293) (0.312) (0.317) Cash income use 0.104 0.085 0.118 0.115 (0.304) (0.273) (0.312) (0.310) Crop and/or livestock sales 0.100 0.093 0.111 0.107 (0.298) (0.288) (0.302) (0.296) Food purchase 0.101 0.096 0.122 0.117 (0.301) (0.295) (0.326) (0.321) Non-food purchase 0.101 0.093 0.113 0.107 (0.301) (0.289) (0.316) (0.309) Child schooling 0.101 0.094 0.112 0.112 (0.301) (0.292) (0.316) (0.315) Other decisions 0.101 0.092 0.119 0.112 (0.301) (0.287) (0.313) (0.306) Observations 1,050 1,049 962 959 Notes: Point estimates are sample means. Standard deviations are in the parentheses. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. Table 16 presents a summary of empowerment measures differentiated by PVTG and non- PVTG households. The results are consistent with those of Table 15 which show that decision- making and resources are largely controlled by men. However, the results in Table 16 show some significant differences. Panel A shows that at baseline, the share of joint decisions made by PVTG households is significantly lower than the share of joint decisions made by non- PVTG households. In the midline sample, a larger share of women made solitary decisions in PVTG households compared to non-PVTG households, but the share of joint decisions made by PVTG households was lower than that of non-PVTG households. Consistent with the aggregated results in Panel A, a breakdown of the share of decisions made independently by women across the seven decision domains shown in Panel B of Table 15 reveals that a significantly larger share of women made solitary decisions in all the seven decisions domains in PVTG households compared to non-PVTG households. Since the share of solitary 36 decisions made by women is our main interest, improvements in the variable among PVTG households are certainly welcome. Table 16. Proportions of women’s and men’s decisions by tribal group Decisions 2017 2021 Panel A: Decision-making in all seven decision domains combined PVTGs non-PVTGs PVTGs non-PVTGs Women's solitary decisions 0.112 0.091 0.131* 0.103 (0.316) (0.287) (0.332) (0.300) Men's solitary decisions 0.880 0.881 0.625 0.630 (0.325) (0.321) (0.457) (0.455) Joint decisions 0.008*** 0.028 0.214* 0.248 (0.086) (0.158) (0.376) (0.402) Panel B: Share of solitary decisions by women in each of the seven decision domains PVTGs non-PVTGs PVTGs non-PVTGs Agricultural input use 0.112 0.091 0.135* 0.105 (0.316) (0.287) (0.336) (0.303) Cash income use 0.104 0.088 0.141** 0.105 (0.303) (0.279) (0.340) (0.296) Crop and/or livestock sales 0.111 0.089 0.132** 0.098 (0.312) (0.283) (0.324) (0.286) Food purchase 0.112 0.092 0.151*** 0.104 (0.316) (0.289) (0.357) (0.304) Non-food purchase 0.114* 0.088 0.141*** 0.095 (0.316) (0.283) (0.346) (0.293) Child schooling 0.112 0.090 0.138*** 0.099 (0.316) (0.287) (0.346) (0.299) Other decisions 0.113* 0.088 0.142*** 0.102 (0.316) (0.281) (0.340) (0.292) Observations 704 1,395 636 1,285 Notes: Point estimates are sample means. Standard deviations are in the parentheses. A two-sample t-test is used for the test of mean differences between PVTG and non-PVTG in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. 4.1.9 Agricultural yield and sales Table 17 presents summary statistics on the total value of production (including both crops and livestock) in rupees; selected crop yields in kilograms per hectare (kg/ha), value of sales in rupees, and crop sales in kg/ha by treatment status. Panel A shows that, on average, sample households produced output worth 48,000 rupees at baseline and 64,000 rupees at the midline survey during the Kharif season.7 7 Following the baseline analysis, we report results of Kharif season for proper comparison and to create panel data. Baseline data only covered the Kharif season; however, the midline survey includes Kharif, Rabi, and summer seasons. 37 Table 17. Agricultural yield and sales in Kharif season by treatment 2017 2021 Treatment Control Treatment Control Panel A: Produced output Household-level Value of total production (rupees) 46,798.2 49,054.2 63,598.6 63,592.8 (50,659.6) (55,372.6) (78,753.9) (81,622.2) [1,050] [1,049] [962] [959] Plot-level Paddy yield (kg/ha) 1,530.2 1,544.7 2,546.3 2,598.8 (1,409.2) (1,419.7) (1,143.8) (1,227.6) [691] [710] [1,383] [1,499] Maize yield (kg/ha) 1617.0 1,482.5 2,418.0 2,382.9 (1577.6) (1,303.8) (1,321.5) (1,271.7) [85] [68] [171] [121] Mandia (finger millet) yield (kg/ha) 1,046.4 939.9 439.3 452.0 (1,296.8) (1,170.4) (703.8) (629.6) [97] [117] [72] [61] Panel B: Sold output Household-level Value of total sales (rupees) 3,821.5 3,474.4 11,278.4 11,100.0 (9,030.2) (8,594.5) (19,845.7) (19,903.5) [1,050] [1,049] [962] [959] Plot-level Paddy sold (kg/ha) 159.5 126.0 126.0 124.9 (600.4) (582.7) (553.9) (578.7) [691] [710] [1,383] [1,499] Maize sold (kg/ha) 180.8 216.2 1511.7** 1,123.8 (886.4) (821.9) (1,619.1) (1,571.7) [85] [68] [171] [121] Mandia (finger millet) sold (kg/ha) 4.12* 74.8 141.6 72.9 (22.1) (411.9) (534.0) (401.9) [97] ]117] [72] [61] Notes: Point estimates are mean. Standard deviations are in parentheses and the numbers of observations are in brackets. A two-sample t-test is used for the test of mean differences between treatment and control groups in the same year. Level of significance *p < 0.10, **p < 0.05, and ***p < 0.01. The numbers of observations differ across crops because the number of crops grown differs by household. 1 USD = 65.1 INR (Indian rupees). Although there is an increase in the value of production in the midline survey, the value of production does not differ significantly between treatment and control households. Plot-level 38 results show that the average yield of rice increased from 1,500 kg/ha to 2,500 kg/ha in the midline survey, maize yield increased from 1,500 kg/ha to 2,500 kg/ha, but Mandia (finger millet) yield was reduced by one-half in the midline survey. We show statistics of these three crops as they are the most widely grown crops by sample households, with rice being the most widely grown crop and Mandia the least. Panel B of Table 17 shows that, on average, sample households sold output worth around 3,600 rupees at baseline and 11,000 rupees in the midline survey during the Kharif seasons. Compared to the value of production, only 8% of the farm output or value of production was sold at baseline, while around 17% of the farm output was sold in the midline survey, which suggests that sample households mainly produce for home-consumption and have low market participation. This is also supported by the plot-level results which show very low sales values per hectare, especially for rice. More maize is sold per ha compared to rice, even though more rice is sold in absolute terms, probably because the few households that grow maize sell maize, but consume rice. Given that very few households grow and sell maize, these results are only suggestive due to the small sample size. Comparison of treatment and control households sh