A Multifaceted Intervention with Savings Incentives to Reduce Multidimensional Child Poverty: Evidence from the Bridges Study (2012–2018) in Rural Uganda

Abstract

Using a randomized controlled trial design, we examine the effects of savings incentives (match rate 1:1 versus 1:2) with mentorship and financial trainings on child poverty among 1383 orphaned children (mean age 12.7 years at baseline) in rural Uganda. Given the difficulty to capture child poverty using monetary measures, we use a multidimensional class of poverty that captures four dimensions: health, assets, housing, and behavioral risks. Results show that children in treatment groups experienced reductions in poverty incidence by 10 percentage points (or deprivation score by 8 percent) relative to control group counterparts at four years post-baseline, and a higher savings incentive led to stronger effects. Further, children in treatment groups were more likely to escape the poverty trap. Finally, we assess the robustness of these results to various weighting structures. This study offers a unique evidence on effectiveness of a multifaceted intervention targeting children in alleviating poverty.

1. Introduction

Recent statistics indicate that around 767 million people, or 10% of world’s population lives in extreme poverty, defined as living on less than US$1.90 a day, and half of the extreme poor live in sub-Saharan Africa (SSA) (World Bank, 2016). If poverty is defined in the context of a multidimensional construct, including the domains of health, education, and living standard that captures an individual’s capacity to achieve functioning (Sen, 1999), studies show that 48% of the multidimensionally poor people are children, and 43% of the poor children come from SSA (Alkire et al., 2017). Growing up in extreme poverty stunts children’s growth and development and hinders their human and social capital accumulation, which further contributes to the intergenerational transmission of poverty and suboptimal adult wellbeing (UNICEF and the World Bank Group, 2016). Against this backdrop, fighting poverty has become the world’s top priority in the Sustainable Development Goals (United Nations, 2018). Recent evidence suggests the potential of multifaceted interventions in alleviating deep poverty (Bandiera et al., 2017; Banerjee et al., 2015, 2016; Blattman et al., 2016). Such interventions provide financial resources (e.g., through savings accounts or grants) and support services (e.g., through mentorship and/or skill trainings) to poor individuals or families. However, little evidence exists on the effectiveness of this approach in alleviating the interlinked dimensions of poverty among children, especially when such an approach incorporates varying levels of savings incentives.

This study evaluates the effectiveness of the Bridges to the Future trial, a multifaceted intervention composed of an incentivized savings account (with a savings match), mentorship, and training, on alleviating multidimensional poverty among orphaned children, a population that is more vulnerable to risks and sub-optimal developmental outcomes (UNICEF, 2006). The study contributes to the ongoing debate and the literature regarding addressing multidimensional poverty in four specific ways. First, we adopt a rigorous experimental design to assess the effectiveness of a multifaceted intervention approach on poverty reduction among children, a less-examined population in the existing literature and debate. Specifically, we conducted a cluster randomized controlled trial, which is ideal for establishing causality. Second, we examined whether, coupled with support services, a very high savings incentive rate (1:2 match) is more effective than an already-high savings incentive rate (1:1 match) in alleviating poverty in a resource-deprived setting. In particular, the experimental design contained three study arms: a control group, a Bridges intervention group, and a Bridges PLUS group (an intervention group with higher savings incentives). Third, we followed these children annually from baseline for four years (see Fig. 1). This time frame contributes to the growing body of literature that examines long-term effects of multifaceted interventions (e.g. Bandiera et al., 2017; Banerjee et al., 2016; Misha et al., 2018), allowing us to test long-term intervention effects. Fourth, following the recent consensus in international practice and academic communities (Atkinson, 2017; de Milliano & Plavg, 2014; Roelen, 2017), we conceptualized child poverty as a multidimensional construct and operationalized it using a multidimensional poverty index. This conceptualization allows us to study intervention effects on poverty among children, a population group that does not have regular income.

Fig. 1.

Study timeframe

To our knowledge, this paper is among the first applications of the multidimensional poverty measure in studying the effectiveness of poverty alleviation programs (Malaeb & Uzor, 2017; Robano & Smith, 2014). We therefore utilized three approaches that can enrich the literature and debate about the application of a multidimensional poverty index in intervention evaluation. First, we compared results from using a multidimensional poverty index as an outcome to results from applying a conventional approach that examines intervention effects on various single indicators as separate outcomes—a dashboard approach, and then we discussed the added value of studying multidimensional poverty. Second, we experimented with an approach in testing the sensitivity of findings to the choice of weighting structures across dimensions. Third, we took advantage of the longitudinal nature of data and tested whether the intervention was effective in helping children break the poverty trap.

Results show that this multifaceted intervention, composed of an incentivized savings account, mentorship, and training, alleviated the incidence of child multidimensional poverty and reduced the extent of deprivation. Specifically, although the intervention effects were stronger at Year 2 and Year 3 follow-ups (right after intervention completion), the intervention still reduced the incidence of child multidimensional poverty by more than 10 percentage points and reduced the deprivation score by more than 8 percent (0.25 standard deviations) at Year 4 (two years after intervention completion). Overall, the magnitude of poverty reduction was higher for Bridges PLUS children. Furthermore, we perform a series of robustness checks on the structure of our child multidimensional poverty measure, and the effect of the program on deprivation scores remains robust to such changes. We also find that using a multidimensional poverty index to study intervention effects can uncover the extent of joint deprivation that the conventional analyses by each indicator are unable to capture. Finally, we test whether the intervention provides a pathway for poverty exit. We find that, on average, the intervention was effective in breaking the poverty trap. Findings from this study suggest that implementing multifaceted combination interventions is feasible in a poor country like Uganda and has long-term effects on reducing child poverty.

2. Literature Review

2.1. Multifaceted Poverty Alleviation Programs

Recent empirical studies point to the potential of multifaceted (or “big push”) interventions in lifting families in deep poverty out of the poverty trap (Bandiera et al., 2017; Banerjee et al., 2015, 2016; Blattman et al., 2016; Malaeb & Uzor, 2017; Robano & Smith, 2014). These multifaceted interventions, usually providing grants and/or savings products as well as education and/or services over a relatively short period of time, have demonstrated effectiveness in raising income and consumption levels (Schaner, 2016).

For example, Banerjee et al. (2015), through a randomized controlled trial design, tested the effectiveness of a multifaceted “graduation” program aiming to improve gains in self-employment activities among the very poor in six developing countries (Ethiopia, Ghana, Honduras, India, Pakistan, and Peru). This program offered productive assets, training and support, life skills coaching, access to savings accounts, and health information or services. Two years after baseline, the intervention, regardless of geographical contexts, significantly increased household consumption, food security, income and revenue, and household assets, and improved financial conditions and mental health. The results were sustained one year after the end of the intervention (36 months after baseline), and this intervention was proven to be cost-effective (Banerjee et al., 2015). A recent follow-up of this intervention in the state of West Bengal in India found that the intervention effects persisted or continued to grow even seven years after baseline (5.5 years after intervention completion; Banerjee et al., 2016). A similar intervention in Bangladesh, Targeting the Ultra-Poor (TUP), that transferred assets, trainings, and cash allowance to low-income women also demonstrated sustained positive effects on labor supply and earnings at seven-year follow-up (Bandiera et al., 2017). The same intervention in Bangladesh also reduced the incidence of multidimensional poverty (Robano & Smith, 2014). However, the nine-year follow-up of the TUP program began to show diminishing employment effects, where some participants switched back to their lower-income baseline occupations (Misha et al., 2018).

Another example took place in post-war northern Uganda, where Blattman et al. (2016) studied the effectiveness of the WINGS program that targeted war-affected, low-income women. The WINGS program provided them with a 5-day training course in business skills, a grant of 300,000 UGX (or $150 in 2009), support to form self-help groups (offered to half of the treatment group), and one-on-one supervision (offered to half of the control group who received delayed treatment). The authors found that 16 months after receiving the grants, women who received the treatment doubled their microenterprise ownership (from 39 to 80%) and incomes (from $1 to $2) and had increased social support and community participation (Blattman et al., 2016). The intervention also reduced incidence of multidimensional poverty (Malaeb & Uzor, 2017). Furthermore, group formation resulted in higher earnings, and supervision improved business survival but not consumption (Blattman et al., 2016). No effect was observed in areas of physical health, mental health, or domestic violence (Blattman et al., 2016).

The above studies point to the potential of multifaceted intervention approaches in reducing poverty and improving multidimensional well-being. However, little research has been conducted adopting rigorous methodology to examine effectiveness of this approach in alleviating poverty among children. This multifaceted intervention approach is not new to service providers serving the child and youth population. Variants of this approach have been adopted by many non-governmental organizations, from PLAN International in West Africa and CARE in Burundi to Padakhep in Bangladesh, to name a few (Deshpande & Zimmerman, 2010). Similar to the examples from Bandiera et al. (2017), Banerjee et al. (2015, 2016) and Blattman et al. (2016), multifaceted programs for children provided non-financial support (training on income-generating skills, training on sexual reproductive health, life skills training, mentorship, etc.) alongside financial support such as savings and credits (Deshpande & Zimmerman, 2010). Nonetheless, evaluations of these initiatives targeting children and youth are scarce. Among the few interventions that conducted evaluations, the design was often qualitative (Ahammed, 2009) or quasi-experimental (Erulkar & Chong, 2005). Among the evaluations that adopted an experimental design (e.g., Ssewamala et al., 2016), poverty was not the major outcome examined. It is crucial to build an evidence base to help improve strategies to support children in resource-constrained contexts. Therefore, we studied the effectiveness of this multifaceted intervention model in alleviating multidimensional poverty among disadvantaged orphan children using an experimental design.

2.2. Intervention Components

In this section, we review literature on the effectiveness of components in the intervention under examination, including incentivized savings accounts, financial education, mentorship, and income generating activities training. Savings accounts are vital to cash-flow management and consumption smoothing among the poor (Karlan et al., 2014). Accumulated savings can contribute to investment in human capital development, including education and job training, and business (Steinert et al., 2018). A systematic review covering 27 randomized controlled trials on savings promotion interventions in sub-Saharan Africa found that the effects of such programs on poverty reduction were statistically significant but small (Steinert et al., 2018). These poverty reduction effects manifested in modest increases in household expenditures and incomes, improved food security, and higher gains from family businesses. When it comes to more distal outcomes of well-being, such as education and health, these savings interventions showed no effect (Steinert et al., 2018). Findings from this systematic review suggest that effects of interventions solely focusing on savings promotion are only modest and signify the importance of incorporating alternative intervention approaches in fighting poverty.

Providing financial incentives to save is a design feature that may promote savings and, in turn, alleviate poverty. Prior studies suggest that incentives can increase the likelihood of saving (Madrian, 2012), but whether the levels of incentives matter for savings performance or economic well-being is contested in prior literature (Kast et al., 2012; Karlan & Zinman, 2014; Schaner, 2016; Schreiner, 2005; Wang et al., 2018). In rural Kenya, providing financial incentives in the form of interest rates that were higher than market rates to low-income individuals, even though temporary (within a 6-month window), has long-run effects (3 years after baseline or 2.5 years after the subsidy expired) on increasing household incomes and business profits (Schaner, 2016). Specifically, interest rates of 12% and 20% had stronger effects than a 4% interest rate. In contrast, providing unconditional payment in cash had no long-run impact (Schaner, 2016). Two other studies based in Chile and the Philippines also found that lower incentives in the form of 3–5% interest rates had limited effects on savings (Kast et al., 2012; Karlan & Zinman, 2014). Limited research contrasts the effectiveness of high levels of savings incentives in promoting savings or alleviating poverty, and little research targets children and examines how families with children respond to such incentives. One recent study targeting orphaned children in Uganda found that a savings match rate as large as 1:1 increased savings when compared to those children in the control group that received no savings incentives. However, when receiving an even larger savings incentive of a 1:2 match rate, although those received higher savings incentive demonstrated a higher probability of saving any money and a higher deposit frequency, the reported total amount of savings was similar to those receiving a 1:1 match rate immediately after the intervention (two years following baseline; Wang et al., 2018). This study only examined the effects of incentives at the time of intervention completion without a longer follow-up period. Studies to date demonstrate that high levels of incentives are more effective than low levels of incentives, although it is unknown whether there is a critical threshold after which the return to incentives may diminish. There is much to be known about the effects of savings incentives on poverty alleviation among the child population and how families with children respond to varying levels of savings incentives.

Prior studies documented the effectiveness of financial literacy training in improving financial capability (Shephard et al., 2017), financial knowledge (Kaiser & Menkhoff, 2016; O’Prey & Shephard, 2014), self-efficacy (Shephard et al., 2017), and savings behavior (Kaiser & Menkhoff, 2016; Miller et al., 2015; O’Prey & Shephard, 2014; Shephard et al., 2017). To promote savings and earnings among youths, Jamison et al. (2014) found that providing financial education can be as effective as offering a savings account to youths (Jamison et al., 2014). In addition, incorporating mentorship, an approach that offers guidance from an individual more similar to their peers, into interventions targeting children and youths can promote positive behavioral, social, emotional, and academic developmental trajectories (DuBois et al., 2011). Specifically, mentorship programs enhanced youths’ social capital, emotional well-being, cognitive skills, and positive identity development, and reduced their risk-taking behaviors not only in western developed countries, but also in developing countries in sub-Saharan Africa (DuBois et al., 2011; Harrison et al., 2010; LoScuito et al., 1996; Moodie & Fisher, 2009; Nabunya et al., 2015; Ssewamala et al., 2014). Furthermore, income-generating activity (IGA) training aims to help poor individuals improve their technical skills and productivity in the non-agricultural or agricultural domains (e.g., crop production, poultry rearing, cattle rearing, agribusiness, etc.) (Hilton et al., 2016). Studies have shown that IGA training increased farmers’ technological competency (Jothilakshmi et al., 2009; Noor & Dola, 2010), capacities in livestock rearing, networking capacities (Jothilakshmi et al., 2009), and income or expenditure (Hilton et al., 2016; Mahmud et al., 2012).

Overall, prior research points to the fact that offering a savings account by itself only has small effects in promoting savings. Therefore, augmenting the savings account with special design features (e.g., financial incentives and commitment devices) and support services, such as financial and IGA training and mentorship, may have more promise in reducing poverty. The intervention under study evaluates the poverty alleviation effects of a multifaceted intervention incorporating the above components. Furthermore, we extend prior literature by focusing on a less-examined population in multifaceted intervention in prior literature—families with children and by delineating the effects of various levels of savings incentives on poverty alleviation. Children are often not a target group of financial interventions, yet, especially for low-income children, accumulating financial resources designated toward their growth may prevent early school dropout and encourage long-term planning (Curley et al., 2010). Therefore, this study examines whether a multifaceted intervention targeting children may chart a path for reducing intergenerational transmission of poverty beyond a short term.

3. The Intervention and Design: Bridges to the Future

Bridges to the Future is a study supported by the United States National Institute of Child Health and Development (NICHD) (2012–2016). AIDS-affected children from 48 primary schools in southwest districts of Uganda, a region heavily affected by HIV/AIDS and poverty, were recruited into the study through the help of Masaka Diocese parish priests. The inclusion criteria for participants were: (a) an HIV/AIDS-orphaned child—defined as a child who lost one or both parent(s) to AIDS (based on self-reports and school records); (b) in the last two grades of primary school¹; and (c) living within a family (broadly defined and not an institution or orphanage). Children in family care have different needs compared to those in institutions. This study initially recruited 1,410 participants for baseline assessment, among which 27 participants were identified during the first year of the study as not meeting the inclusion criteria of being an orphan. Hence, our final study sample size was reduced from 1,410 to 1,383 (slightly over 98% of the original sample).

The Bridges to the Future study used a three-group cluster randomized controlled trial design. The 48 rural public primary schools included in the study were randomly assigned to each of the three study conditions. Specifically, 16 schools were randomly assigned the control condition (n = 487 students), 16 schools to the Bridges condition (n = 396 students), and the final 16 schools to the Bridges PLUS condition (n = 500 students).² All children from the same school who met the inclusion criteria and assented to study participation received the same intervention to prevent contamination. The study received ethics approval from Columbia University (Protocol #AAAK3852) and the Uganda National Council for Science and Technology (Protocol #IRB00011353). The study’s protocol is registered in the clinicaltrial.gov database (ID#NCT01790373).

All children in the study, including those in control and intervention arms, received bolstered standards of care (SOC) for school-age AIDS-orphaned children in southwestern districts of Uganda. The SOC included counseling (provided by priests in the community), school lunches, and scholastic materials (textbooks and notebooks). In addition to the SOC, children in Bridges and Bridges PLUS arms received several intervention components: (a) workshops that focused on asset building, family microenterprise development, future planning, and how to protect oneself from risk. Children’s caregivers were invited to join their children in the workshops on family microenterprise development; (b) mentors were assigned to children to reinforce learning and to build optimism; and (c) a Child Development Account (CDA),³ from which matched savings could only be used for either secondary education or microenterprise development.

The only difference between the two interventions arms, Bridges and Bridges PLUS, was the rate of the financial incentive for saving (match rate). Children in the Bridges condition received a 1:1 match rate (each UGX $1,000 deposited into the savings account was matched by UGX $1,000, while children in the Bridges PLUS condition received a 1:2 match rate (each UGX $1,000 saving was matched by UGX $2,000). Participants in either intervention group, Bridges or Bridges PLUS were subjected to a match cap of UGX $20,000 a month (approximately US$10 in 2012), meaning that if a child saved more than UGX $20,000 in a month, they would not receive additional matching for any amount of savings above UGX $20,000. Children received a bank statement from the research team periodically which indicates the amounts of savings and matches. The comparison between intervention and control groups reveal the effectiveness of this multifaceted intervention model in alleviating multidimensional poverty. The varying match rates allow this study to investigate whether a higher level of savings incentives can alleviate multidimensional poverty to a further extent. All the above-mentioned intervention components were provided for two years post-baseline assessment. This study surveyed their outcomes for five time points, from baseline to Year 4 (two years after intervention completion).

Comparing the differences in intervention take-up rates, the account opening rate was 90% for Bridges PLUS children and 82% for Bridges children. Bridges PLUS children accumulated significantly more self-reported savings (UGX $89,751) at the end of the intervention period when compared to Bridges children (UGX $67,330; see Appendix Table 7). This amount of financial transfer is equivalent to more than half of the median monthly nominal wages for paid employees in Uganda in 2012/2013 (UGX $110,000; Uganda Bureau of Statistics, 2014). Although Bridges PLUS children accumulated more self-reported savings (these savings can be stored anywhere, including under a pillow or in a piggy bank), their actual savings deposited into the bank account did not significantly differ from those of Bridges children (Wang et al., 2018). The average monthly match amount provided by the program to children who opened a savings account during the two intervention years was UGX $1,172 for Bridges children and UGX $2,057 for Bridges PLUS children. In addition, Bridges PLUS children had a significantly higher attendance rate in mentorship sessions (67%) than Bridges children (61%), but their attendance rate in micro-enterprise development workshops did not differ significantly (61–64%; see Appendix Table 8).

Table 7.

Poverty dynamics by study groups

	Year 1		Year 2		Year 3		Year 4
	Control	Treated	Control	Treated	Control	Treated	Control	Treated
Panel 1: Comparisons between control and treatment groups
Upward mobile (%)	19.18	26.19	20.86	30.75	20.82	33.29	22.97	29.94
	0.00		0.00		0.00		0.01
Downward mobile (%)	14.87	9.88	15.11	8.91	13.08	7.58	10.81	7.58
	0.01		0.00		0.00		0.05
Structurally poor (%)	44.83	32.86	39.09	25.84	32.93	18.25	25.00	17.76
	0.00		0.00		0.00		0.00
Structurally non-poor (%)	14.87	25.36	15.11	26.10	15.01	25.19	14.86	23.35
	0.00		0.00		0.00		0.00
N	464	840	417	774	413	778	444	805
	Bridges	Bridges plus	Bridges	Bridges plus	Bridges	Bridges plus	Bridges	Bridges plus
Panel 2: Comparisons between bridges and bridges PLUS groups
Upward mobile (%)	27.55	25.16	26.71	33.87	32.56	33.87	28.25	31.31
	0.44		0.03		0.70		0.35
Downward mobile (%)	9.92	9.85	8.90	8.92	7.27	7.83	8.86	6.53
	0.98		0.99		0.77		0.21
Structurally poor (%)	29.75	35.22	27.89	24.26	17.44	18.89	18.56	17.12
	0.09		0.25		0.60		0.59
Structurally non-poor (%)	27.00	24.11	27.89	24.71	26.74	23.96	22.44	24.10
	0.34		0.32		0.38		0.58
N	363	All	337	437	344	434	361	444

Numbers in Italics represent the p-value for the two-tailed test of equality between the two groups. Figures are based on poverty cutoff of 1/4. Upwardly mobile individuals: Those who were multidimensionally poor at baseline and no longer multidimensionally poor at time t (t = Year 1, Year 2, Year 3, and Year 4). Downwardly mobile individuals: Those who were not multidimensionally poor at baseline but multidimensionally poor at time t. Structurally poor individuals: Those who were poor at baseline and at time t. Structurally nonpoor individuals: Those who were poor neither at baseline nor at time t

With regard to the external validity of our study population, although the intervention took place among AIDS-affected orphans in one African country, Uganda, the reported findings can have relevant implications for the design of interventions in other sub-Saharan African countries that have been devastated by a combination of HIV/AIDS and poverty. To be specific, AIDS-affected households containing orphaned children are likely to be more financially unstable due to loss of providers or the need to devote often scarce family resources to the care of ill family members. In addition, AIDS orphaned and vulnerable children often suffer recurrent trauma, starting with the illness and deaths of their first line of defense: the parents. This is usually followed by cycles of poverty, exploitation, and often, sexual abuse (Matshalaga, 2002). As this study targets non-institutionalized orphans, we may not be able to generalize findings from this study to institutionalized orphans who lack family support. With regard to generalizability of findings from this study to non-orphans, on one hand, orphans are likely more disadvantaged than non-orphans prior to receiving the intervention; hence, orphans may have more room for growth in outcomes post-intervention. On the other hand, one can hypothesize that if the intervention can mobilize the most vulnerable children and their caregiving families, in this case, AIDS-orphaned children, to yield from the intervention, it is likely the intervention will be applicable to children who are less vulnerable or more fortunate as well. Preliminary outcomes from a recent experiment targeting non-orphans suggests that the latter hypothesis may prevail.⁴

This multifaceted Bridges to the Future intervention aimed at promoting savings for secondary education, promoting microenterprise development to generate family income, and providing support programs to protect children from future risks (see Table 1 for the conceptual framework). Receiving a child savings account improves children’s financial inclusion into formal financial institutions and helps them accumulate resources for education pursuits (Karimli et al., 2014; Ssewamala et al., 2016). The savings accounts and mentorship, along with IGA training, strengthen the support for the child, caregiver, and the family. This support fosters a brighter prospect for the future and improves the child’s sense of self-worth (Han et al., 2013; Rohe et al., 2017). These benefits provide monetary as well as psychological assets to help a child attain better education outcomes (Ssewamala et al., 2012, 2016), improve mental health (Han et al., 2013; Ssewamala et al., 2012), and reduce the likelihood of risk-taking behavior (Jennings et al., 2016; Nabunya et al., 2015).

Table 1.

Conceptual framework for the bridges to the future study

Input	Mechanisms of change	Proximal outcomes	Distal outcome
Intervention	Self-efficacy and selfesteem Educational plans and aspirations Social support Family support and family stability	Savings and asset accumulation Education Risks (e.g., sexual risk-taking behavior, substance use) Mental health functioning (depression)	Poverty reduction

4. Measuring Child Poverty in Developing Countries

The most widely used poverty measures focus on monetary deprivation, either through assessing income or consumption levels, as they reflect the purchasing power available to fulfill basic needs (Gordon, 2003). However, when measuring child poverty, monetary poverty measures are unable to capture household resources, infrastructure and services allocated to the child (Gordon, 2003). Also, many children have not yet started working, so the lack of their own earnings also limits the possibility to measure monetary poverty among the child population. Following Amartya Sen’s work on the capability approach,poverty is increasingly reconceptualized from material deprivation to individuals’ capabilities and freedom to achieve valuable functioning and living conditions (Roelen, 2017; Roelen et al., 2010; Sen, 1999). This is particularly true when measuring poverty among groups of the population that do not produce income, and their effective share of household consumption is challenging to measure. The capability approach of measuring child poverty can be operationalized through a multidimensional poverty index (MPI), which can capture the capabilities children need to survive, develop, and thrive by incorporating child-specific indicators, such as health, education, and child labor (Alkire et al., 2016). The MPI is an absolute poverty measure assessing the extent of deprivation or inability to meet standards of adequate functioning (Alkire & Santos, 2014). This approach is also particularly suitable in the developing world and effectively measures the outcomes that are most essential to a child’s wellbeing. Furthermore, the methodology allows for thedisentangling of the intervention effects, both individually and across the joint deprivations. Sir Anthony Atkinson (2017) along with an advisory board has also particularly recommended the World Bank to measure and monitor nonmonetary poverty using complementary indicators (including a multidimensional poverty indicator based on the counting approach) (Atkinson, 2017).

We use the Alkire and Foster (2011) methodology to measure multidimensional poverty, which is consistent with the recommendations to the World Bank by Atkinson (2017). The method consists of a dual cut-off strategy. After selecting the appropriate indicators, the first cut-off (z) is applied to obtain the dichotomous deprivation in each indicator. The second cut-off is the poverty cut-off (k) which is the sum of weighted deprivations an individual must attain in order to be considered poor. The sum of weighted deprivations is referred to as the deprivations’ score, or the C vector. After the poverty cut-off (k) is applied, the headcount (or incidence) is calculated as the proportion of people who are considered poor (i.e. having a C vector ≥ k). Next, the deprivations of non-poor individuals are censored and the intensity of poverty (A) is calculated as the average C vectors of poor individuals. Finally, the adjusted headcount ratio of poverty (M₀) is calculated as the product of H and A. M₀ is “the sum of the weighted deprivations that the poor (and only the poor) experience, divided by the total population” (Alkire & Santos, 2014). M₀ satisfies desirable axiomatic properties including sub-group decomposability (making it useful to assess the impact by treatment arms) and dimensional monotonicity (deprivation increases if a poor person is deprived in one additional indicator).

Many studies have investigated the construct and patterns of child multidimensional poverty in the developing world (Alkire et al., 2016; Bastos & Machado, 2009; de Neubourg et al., 2012; Roelen et al., 2010; Trani & Cannings, 2013); however, little evidence exists examining the effectiveness of poverty alleviation approaches on addressing multidimensional poverty. Given that various dimensions of deprivation are detrimental to child development and their future success, understanding the effects of poverty alleviation interventions through a multidimensional lens is critical. To this end, we will assess how the multifaceted Bridges to the Future intervention affects multidimensional poverty as evaluated by a measure relevant to the Ugandan context.

5. Data and Methods

This study uses data from five time points: baseline, Year 1 (12-month; the middle of the intervention), Year 2 (24-month; at the end of the intervention), Year 3, and Year 4 (36- and 48-month: after the intervention) post intervention initiation. Data were collected between 2012 and 2016. The interviews in Year 3 and 4 allow us to assess outcomes at 1 and 2 years after intervention completion (see Fig. 1). The attrition rate at Year 4 was 9.7%. We cannot reject the hypothesis that attrition rates across three study groups were equal, suggesting that attrition rates across three study groups did not differ.⁵

All data and variables used in this study were collected through annual interviews. These interviews were conducted either at the child’s school or at home in a private room. The interview duration was 60–70 min on average; the interviews were conducted in Luganda, the language spoken in the study area, and the responses were recorded in English. All interviewers received structured and intensive interview training led by the Principal Investigator of the study, including Good Clinical Practice. Prior to data collection, all interview questions were translated from English to Luganda and then back-translated from Luganda to English. All questions in the questionnaire had been pre-tested to ensure cultural appropriateness in the Ugandan context.

5.1. A Measure of Child Poverty

Constructing an MPI involves selecting a set of indicators, setting the deprivation cutoffs for each indicator, choosing weights for each indicator, and then determining a poverty cutoff (Alkire & Santos, 2014). We develop an MPI relevant to the Ugandan context through balancing theoretical justifications, empirical evidence, and data availability (Alkire et al., 2016; Bastos & Machado, 2009; de Neubourg et al., 2012; Roelen et al., 2010; Trani & Cannings, 2013). The list of selected indicators is presented in Table 2. The dimensions we include are health, assets, housing, and behavioral risks. These dimensions cover a rich set of indicators, including (a) malnutrition, sexual risk, and mental health under the dimension of health, (b) savings, clothing and shoes, and means of communication or transportation under the dimension of assets, (c) water source, type of housing, and access to electricity under the dimension of housing, and (d) child labor, alcohol use, and school dropout under the dimension of behavioral risk. Most indicators are at the level of the individual child, except for certain indicators related to housing and assets, which are only available at the household level (Roelen et al., 2010). Including some household level characteristics into a child’s MPI is necessary to capture the degree to which the household environment adequately offers basic levels of infrastructure for the child’s development. Also, the intervention may affect households resources and behaviors, as some if not most of children’s savings come from caregivers, and the intervention components, those involving (savings account opening and depositing and family micro-enterprise training) and not involving (mentorship) caregivers, may affect caregivers’ attention and responses to households environment. One of the main challenges in constructing a multidimensional measure is data availability, the lack of which constrains the dimensions and indicators that could be included in the MPI (Alkire & Santos, 2014). This study is no exception. For instance, immunization, physical health,⁶ quality of water, exposure to violence, leisure activities, and social participation⁷ were not included due to the unavailability of data.

Table 2.

Multidimensional child poverty index in Uganda

Dimension	Indicator	Deprived if…	Weight
Health	Malnutrition (I)	The child has less than two meals a day OR the child did not eat any meat/fish last week	1/12
	Sexual risk (I)	The child ever had sex unwillingly or the child aged below 18 had unprotected sex	1/12
	Mental health: depression (I)	The child has moderate to severe depression as measured by the child depression inventory	1/12
Assets	Savings (I)	The child does not have money saved anywhere (bank, savings and credit cooperative, with parents/caregivers)	1/12
	Clothing and shoes (I)	The child has less than two pieces of clothing or no shoes	1/12
	Means of communication or transportation (H)	The household does not own any of the following communication means: television, radio, and cell phone OR any of the transportation means: bicycle, motorcycle, and car	1/12
Housing	Water source (H)a	The water source is more than 1 km from the household	1/12
	Type of housing (H)	The house the household resides in is a muzigo, a hut, or a mud house / is not a brick house	1/12
	Electricity (H)	The household does not have electricity	1/12
Behavioral risks	Child labor (I)b	The child aged below 18 engages in work	1/12
	Drinking (I)	The child has drunk more than a few sips of alcohol in the past month	1/12
	School Dropout (I)	The school-aged child is not currently enrolled in school	1/12

(I) individual child level indicator; (H) household level indicator

^aA total of 88 children responded “don’t know” to the water source question at baseline; hence the sample size was lowered due to this variable. The three groups did not differ in their likelihoods to respond “don’t know” to this question

^bChildren aged above 18 are classified as not child laborers

In Table 2, we also list the deprivation cutoff and weight for each indicator. Similar to the Global MPI by Alkire and Santos (2014), this study adopts equal weight across dimensions and equal weight across indicators in each dimension. In sensitivity analyses, we test the robustness of our results to various weight combinations. We also adopt a poverty cutoff mirroring deprivation in one whole dimension, totaling 1/4 of all indicators (Alkire & Santos, 2014).⁸ After applying the 1/4 poverty cutoff, we can compute the proportion of children who are multidimensionally poor (H: headcount ratio or poverty incidence), the intensity of poverty (A),⁹ and the adjusted headcount ratio (M₀).¹⁰ The three indicators (H, A, M₀) are used to depict multidimensional poverty patterns across groups and time points. Since these three measures are at the aggregate level (e.g. control vs. treatment groups, etc.), we use their analogous micro-components in our regression analysis. Specifically, we examine the impact of intervention on two outcomes, the multidimensional poverty status (H) and the C vector. The C vector is a simultaneous deprivation score for each child. It is derived through summing the deprivation status of each indicator multiplied by weight (1/12), which is a weighted sum of overlapping deprivations ranging between 0 and 1.

5.2. Empirical Strategy

To evaluate the impact of the Bridges intervention on multidimensional poverty, we focus on the intent-to-treat (ITT) effects. This means that we do not assess the impact by program take-up but concentrate on the impact of program assignment status. We adopt a difference-in-difference (DD) model¹¹ to track the changes in outcomes from baseline to a given time point:

$${\text{Y}}_t-{\text{Y}}_{Baseline}={\alpha }_0+{\alpha }_B\text{Bridges}+{\alpha }_{BL}\text{BridgesPLUS}+\epsilon$$

where Y represents two outcomes, multidimensional poverty status (1 is poor, and 0 otherwise) and the C vector (ranging from 0 to 1, which is the weighted sum of overlapping deprivations). The outcomes are the difference in poverty status and deprivation score outcomes between one time-point (Year 1, 2, 3, and 4 follow-ups) and the baseline.¹² Bridges and Bridges PLUS denote dummy indicators of two intervention groups, and the control group serves as the reference group. We estimate the program effect by ordinary least squares regression models, and we cluster the standard errors at the school level to adjust for intra-school correlations. We also conducted sensitivity analyses to bootstrap standard errors, and the results remain largely the same as our main findings with school-level clustered standard errors. After each regression model, we conduct Wald tests to test two null hypotheses: (a) the coefficients from Bridges and Bridges PLUS are jointly equal to zero, and (b) the Bridges coefficient equals the Bridges PLUS coefficient.

6. Results

6.1. Baseline Comparisons

We report descriptive statistics of selected characteristics at baseline in Table 3, including demographic characteristics and MPI indicators.¹³ We also checked the balance across the three groups by testing the null hypothesis that the coefficients from three groups are jointly equal. Across 19 of the 21 demographic and outcome variables tested, we found that the three study groups did not differ at the 0.05 significance level at baseline.¹⁴ The results also show that children across the three groups do not differ in any of the MPI indicators or the C vector. The two exceptions are orphan status and relationships to the primary caregiver. Children in the control arm are 5–7 percentage points more likely to be a double orphan (losing both parents) and to have grandparents as their primary caregiver. In the sensitivity analysis (presented in Appendix Table 8), we control for these two characteristics in our analytic models, and the results do not differ from the main results without these controls.

Table 3.

Baseline summary statistics and randomization tests

	Control (N = 487)		Bridges (N = 396)		Bridges PLUS (N = 500)		Total (N = 1383)		Wald-tests
	Mean/%	SD	Mean/%	SD	Mean/%	SD	Mean/%	SD
Demographic characteristics
Age	12.76	1.23	12.56	1.31	12.71	1.25	12.68	1.26	2.74
Female	0.55		0.57		0.56		0.56		0.21
Household size	6.43	2.97	6.29	2.62	6.32	2.74	6.35		1.33
Number of children	3.18	2.32	3.14	2.08	3.23	2.18	3.19		0.62
Years living in the households	7.12	4.41	7.19	4.44	7.44	4.54	7.26	4.46	0.88
Double orphan	0.25		0.18		0.20		0.21		6.84*
Primary caregiver									18.96**
Parents	0.37		0.41		0.44		0.41
Grandparents	0.40		0.35		0.36		0.37
Other relatives	0.23		0.24		0.21		0.23
Caregiver: employed	0.31		0.34		0.24		0.29		0.59
MPI indicators
Health
Malnutrition	0.85		0.86		0.85		0.85		0.06
Sexual risk	0.03		0.04		0.02		0.03		2.97
Depression	0.05		0.03		0.05		0.05		1.63
Assets
No savings	0.71		0.70		0.66		0.69		1.66
Few clothing and shoes	0.44		0.30		0.35		0.37		5.90+
Lack of communication or transportation	0.44		0.46		0.43		0.44		0.24
Housing
Distant water sources	0.31		0.28		0.36		0.32		4.30
No brick house	0.28		0.25		0.30		0.28		0.76
Electricity	0.94		0.87		0.91		0.91		2.24
Behavioral risks
Child labor	0.10		0.10		0.09		0.09		0.49
Drinking	0.02		0.04		0.03		0.03		1.74
School dropout	0.00		0.00		0.00		0.00		n/a
C vector a	0.35	0.11	0.33	0.12	0.34	0.11	0.34	0.11	3.46

⁺p < 0.10,

^*p < 0.05,

^**p < 0.01.

To examine whether children across three groups are different by their demographic characteristics and indicators at baseline, we employ a multilevel model for each characteristic and use the group membership variables to predict that characteristic. In each multilevel model, we include school-level random intercepts to account for clustering at the school level. We report the test statistic and p-value from the Wald-test after each model testing whether coefficients for the group membership are jointly equal to zero

^aC vector is a simultaneous deprivation score for each child. It is derived through summing deprivation status of each indicator multiplied by weight (1/12), which is a weighted sum of overlapping deprivations

6.2. Poverty Profile: Moments of Aggregate Measures

First, we study the effects of the intervention at the mean by comparing the means of aggregate measures. Table 4 presents the estimates for H (the headcount ratio), A (the intensity), and M₀ (the adjusted headcount ratio: product of H and A). With respect to H, the proportion of children who were MPI poor at baseline was 59.3% for the Bridges group, 64.7% for the Bridges PLUS group, and 68.4% for the control group. These differences across the three groups were not statistically significant at the 5% level. From baseline to Year 1, when children received mentorship and matched CDA, the poverty incidence declined by 17 percentage points (p.p.) for both Bridges and Bridges PLUS children, while the decline in the control group was less than 5 p.p. From the Year 1 to Year 3 (one year after the completion of the intervention), the headcount ratio dropped continuously by 10 p.p for Bridges, 16 p.p for Bridges PLUS, and 8 p.p for control group. At the Year 4 follow-up (two years after intervention completion), H bounced back for the Bridges children but remained stable for both the Bridges PLUS group and the control group. Nevertheless, the poverty incidence was much lower for children in intervention groups than the control group. With respect to poverty intensity (A) among those who were poor, we observed that the poverty intensity was similar for children in all groups at baseline. From baseline to the Year 2, the poverty incidence dropped slightly by 1 to 1.5 percentage points for children in the intervention group and remained the same for control group children.

Table 4.

MPI results by study arms and time points

	Poverty incidence (H)				Poverty intensity (A)				Adjusted headcount ratio (M0)
	Bridges (n = 396)	Bridges PLUS (n = 500)	Control (n = 487)	Overall (n = 1383)	Bridges (n = 396)	Bridges PLUS (n = 500)	Control (n = 487)	Overall (n = 1383)	Bridges (n = 396)	Bridges PLUS (n = 500)	Control (n = 487)	Overall (n= 1383)
Baseline	59.30	64.68	68.35	64.43	40.34	40.46	40.94	40.61	0.24	0.26	0.28	0.26
s.e	2.55	2.21	2.18	1.33	0.43	0.35	0.36	0.22	0.01	0.01	0.01	0.01
P-value	0.07				0.62				0.10
Year 1	41.87	47.80	63.58	51.76	38.16	39.29	41.16	39.85	0.16	0.19	0.26	0.21
s.e	2.59	2.29	2.24	1.38	0.36	0.34	0.38	0.21	0.01	0.01	0.01	0.01
Year 2	40.06	36.61	60.67	46.01	38.64	38.96	40.28	39.49	0.15	0.14	0.24	0.18
s.e	2.67	2.31	2.39	1.44	0.37	0.36	0.38	0.22	0.01	0.01	0.01	0.01
Year 3	32.56	32.26	55.93	40.55	37.43	38.21	41.63	39.67	0.12	0.12	0.23	0.16
s.e	2.53	2.25	2.45	1.42	0.34	0.34	0.46	0.24	0.01	0.01	0.01	0.01
Year 4	37.67	32.43	49.10	39.87	39.28	38.89	40.52	39.71	0.15	0.13	0.20	0.16
s.e	2.55	2.22	2.38	1.39	0.43	0.32	0.35	0.21	0.01	0.01	0.01	0.01

Poverty cutoff is 1/4

P-values P-values derived from Wald tests examining the null hypothesis that Control = Bridges = BridgesPlus

H: Headcount ratio or incidence of multidimensional poverty (MPI)

A: Intensity of multidimensional poverty: A is calculated as total deprivation scores (summation of weighted deprivation status across indicators) among the MPI poor divided by the total number of MPI poor people

M₀: Adjusted headcount ratio: calculated as H x A. M₀ is the sum of the weighted deprivations that the poor (and only the poor) experience, divided by the total population (Alkire & Santos, 2014)

Next, we move on to M₀, the adjusted headcount ratio that captures “the proportion of weighted deprivations that the poor experience out of all the total potential deprivations in the society” (Alkire & Santos, 2014). The results show that, from baseline to the Year 4, Bridges PLUS children experienced a larger decrease in M₀ (by 0.13) than did Bridges children (by 0.09) or control group children (by 0.08). The trend over time is illustrated in Appendix Fig. 3 (cutoff = 1/4). Overall, we observe a steeper decline in M₀ among Bridges and Bridges PLUS children than among control group children. Bridges group children experienced an increase in M₀ from the 36-month to the 48-month follow-up, but no such trend was shown among Bridges PLUS or control group children.

To delineate what indicators drove the changes in multidimensional poverty incidence across the intervention groups, we further investigated the censored headcount ratio for each indicator by study group and time (see Table 5). Censored headcount ratios are the poverty incidence rates (%) for each indicator among those who are poor. For brevity, we discuss the findings on changes from baseline to the Year 4 follow-up.

Table 5.

Censored Headcount Ratio (%) at Poverty Cutoff 1/4 for each indicator across study groups and time

		Bridges						Bridges PLUS						Control
		Baseline	Year 1	Year 2	Year 3	Year 4	Δ Year 4 – Year 1	Baseline	Year 1	Year 2	Year 3	Year 4	Δ Year 4 – Year 1	Baseline	Year 1	Year 2	Year 3	Year 4	Δ Year 4 – Year 1
Health
Malnutrition	Mean	57.68	37.47	37.39	30.81	34.35	−23.33	61.49	45.28	34.55	29.72	29.73	−31.76	63.96	60.78	56.83	51.33	45.27	−18.69
	s.e	2.57	2.54	2.64	2.49	2.5	3.59	2.25	2.28	2.28	2.2	2.17	3.13	2.25	2.27	2.43	2.46	2.36	3.26
Sexual risk	Mean	3.23	1.65	2.37	3.49	6.65	3.42	1.91	2.1	2.52	3.46	3.15	1.24	2.86	2.8	5.04	6.78	7.43	4.57
	s.e	0.92	0.67	0.83	0.99	1.31	1.60	0.63	0.66	0.75	0.88	0.83	1.04	0.78	0.77	1.07	1.24	1.25	1.47
Depression	Mean	2.96	4.13	0.89	2.91	3.32	0.36	4.47	1.47	0.46	2.76	3.15	−1.32	4.84	3.23	1.92	3.87	2.93	−1.91
	s.e	0.88	1.05	0.51	0.91	0.94	1.29	0.95	0.55	0.32	0.79	0.83	1.26	1.01	0.82	0.67	0.95	0.8	1.29
Assets
No savings	Mean	46.9	23.69	23.15	16.57	24.1	−22.8	49.15	20.55	14.65	14.75	20.05	−29.1	55.6	53.23	50.12	41.16	36.71	−18.89
	s.e	2.59	2.23	2.3	2.01	2.25	3.43	2.31	1.85	1.69	1.7	1.9	2.99	2.33	2.32	2.45	2.42	2.29	3.27
Lack of communication or transportation	Mean	38.81	28.37	25.52	19.19	23.55	−15.26	37.45	30.4	23.34	19.59	21.4	−16.05	40	40.52	36.93	37.53	33.11	−6.89
	s.e	2.53	2.37	2.38	2.13	2.24	3.38	2.23	2.11	2.03	1.91	1.95	2.96	2.3	2.28	2.37	2.39	2.24	3.21
Few clothes and shoes	Mean	25.07	13.5	10.39	4.94	7.2	−17.87	30.64	18.03	8.24	7.14	5.86	−24.78	39.34	32.54	22.06	15.01	9.91	−29.43
	s.e	2.25	1.8	1.66	1.17	1.36	2.63	2.13	1.76	1.32	1.24	1.12	2.41	2.29	2.18	2.03	1.76	1.42	2.69
Housing
Distant water sources	Mean	23.45	17.36	14.24	8.14	10.53	−12.92	29.79	21.8	11.21	6.91	5.86	−23.93	26.37	23.28	16.31	14.77	11.94	−14.43
	s.e	2.2	1.99	1.91	1.48	1.62	2.73	2.11	1.89	1.51	1.22	1.12	2.39	2.07	1.96	1.81	1.75	1.54	2.58
Non-brick house	Mean	21.02	14.33	13.35	13.95	14.68	−6.34	27.02	22.01	16.48	16.13	17.12	−9.9	25.93	23.28	21.58	21.07	20.05	−5.88
	s.e	2.12	1.84	1.86	1.87	1.87	2.83	2.05	1.9	1.78	1.77	1.79	2.72	2.06	1.96	2.02	2.01	1.9	2.80
No electricity	Mean	2.58	2.57	2.63	2.34	2.43	−0.15	2.24	2.29	2.27	2.06	2.06	−0.18	2.21	2.25	2.43	2.46	2.33	0.12
	s.e	2.58	2.57	2.63	2.34	2.43	3.54	2.24	2.29	2.27	2.06	2.06	3.04	2.21	2.25	2.43	2.46	2.33	3.21
Behavioral risks
Child labor	Mean	8.36	4.96	8.31	7.85	7.76	−0.6	7.66	11.11	11.21	8.06	7.66	0	8.35	5.6	8.87	12.83	8.56	0.21
	s.e	1.44	1.14	1.51	1.45	1.41	2.02	1.23	1.44	1.51	1.31	1.26	1.76	1.3	1.07	1.39	1.65	1.33	1.86
Drinking	Mean	3.23	1.38	0.59	1.45	1.66	−1.57	2.34	1.05	0.46	0.69	0.45	−1.89	1.98	0.86	1.2	0.48	0.9	−1.08
	s.e	0.92	0.61	0.42	0.65	0.67	1.14	0.7	0.47	0.32	0.4	0.32	0.77	0.65	0.43	0.53	0.34	0.45	0.79
School dropout	Mean	0	4.96	12.76	11.92	13.02	13.02	0	5.03	14.19	14.52	11.71	11.71	0	5.6	15.35	25.67	22.07	22.07
	s.e	0	1.14	1.82	1.75	1.77	1.77	0	1	1.67	1.69	1.53	1.53	0	1.07	1.77	2.15	1.97	1.97

Censored headcount ratios are the poverty incidence rates (%) for each indicator among those who are multidimensionally poor

Significant changes across several indicators from baseline to Year 4 follow-up contribute to the observed patterns in Table 4. With respect to the health dimension, the malnutrition rates decreased more and sexual risk-taking behaviors increased less for the intervention group children. No significant effect of intervention on depression was found. In the dimension of assets, we found that rates of not having any savings or lacking means of communication or transportation decreased more for children in either of the treatment arms when compared to control group children. When comparing Bridges and Bridges PLUS children in both the health and asset dimensions, Bridges PLUS children experienced more positive changes than Bridges children. This trend was also observed in the dimension of housing. Bridges PLUS children showed greater improvement in their housing conditions, including being less likely to live far away from water sources, to live in a non-brick house, or to have no electricity than both control group and Bridges children. Lastly, in the dimension of behavioral risk, we found that the increases in school dropout and child labor were lower and the decrease in the likelihood to drink alcohol was greater for children in the intervention groups. Overall, Bridges PLUS children showed a larger decrease in deprivation across indicators, followed by Bridges children. Control group children are the most disadvantaged overall across time.

6.3. Intervention Effects: Regression Results

Table 6 (Panel 1) shows the effects of the intervention on multidimensional poverty at each follow-up time point. First, from baseline to the Year 1, one year after intervention initiation (when the mentorship programs were completed, savings accounts had been opened, and the children were saving more actively), the decline in poverty incidence rates was 14.1 p.p. and 11.6 p.p. more for Bridges and Bridges PLUS children, respectively, than for the change shown among control group. The results for the C vector (simultaneous deprivation scores for each child) show similar patterns to the poverty status results. The test results suggest that these effect sizes are significantly different from zero, indicating that changes in outcomes among children in the intervention groups are significant. However, the differences in coefficients between the Bridges and Bridges PLUS groups are not statistically significant.

Table 6.

Regression results on multidimensional poverty incidence and deprivation score

Outcome	MPI Poor (poverty incidence)				C vector (deprivation score)
Time points	Year 1 – Baseline	Year 2 – Baseline	Year 3 – Baseline	Year 4 – Baseline	Year 1 – Baseline	Year 2 – Baseline	Year 3 – Baseline	Year 4 – Baseline
Panel 1: All sample
Bridges	−0.141** (0.053)	−0.136** (0.052)	−0.169** (0.064)	−0.028 (0.045)	−0.048*** (0.014)	−0.042*** (0.011)	−0.055*** (0.014)	−0.010(0.011)
Bridges PLUS	−0.116** (0.049)	−0.217*** (0.048)	−0.207*** (0.061)	−0.107*** (0.034)	−0.033** (0.013)	−0.055*** (0.011)	−0.061*** (0.014)	−0.028** (0.011)
Constant	−0.046 (0.038)	−0.055 (0.036)	−0.096* (0.050)	−0.191*** (0.029)	−0.014(0.011)	−0.019** (0.007)	−0.023** (0.011)	−0.060*** (0.007)
N	1227	1112	1118	1169	1227	1112	1118	1169
Bridges = BridgesPlus (p-value)	0.614	0.109	0.484	0.054	0.192	0.282	0.614	0.124
Bridges = BridgesPlus = 0 (p-value)	0.023	0.000	0.005	0.007	0.003	0.000	0.000	0.053
Panel 2: Sample aged < 18 at a given time point
Bridges	−0.142** (0.053)	−0.134** (0.054)	−0.186*** (0.061)	−0.031 (0.052)	−0.048*** (0.014)	−0.041*** (0.011)	−0.059*** (0.014)	−0.011 (0.013)
Bridges PLUS	−0.115** (0.049)	−0.213*** (0.050)	−0.221*** (0.059)	−0.146*** (0.039)	−0.03** (0.013)	−0.054*** (0.012)	−0.068*** (0.014)	−0.039*** (0.013)
Constant	−0.046 (0.038)	−0.055 (0.038)	−0.086* (0.048)	−0.152*** (0.029)	−0.014(0.011)	−0.019** (0.007)	−0.019* (0.011)	−0.049*** (0.009)
N	1224	1098	1047	866	1224	1098	1047	866
Bridges = BridgesPlus (p-value)	0.587	0.128	0.486	0.026	0.176	0.307	0.442	0.033
Bridges = BridgesPlus = 0 (p-value)	0.023	0.000	0.002	0.001	0.003	0.000	0.000	0.014

School-level cluster robust standard errors in parentheses.

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

MPI poor = H: Headcount ratio or incidence of multidimensional poverty (MPI); C vector = C vector is a simultaneous deprivation score for each child. It is derived through summing deprivation status of each indicator multiplied by weight (1/12), which is a weighted sum of overlapping deprivations. All models cluster the standard errors at the school level to adjust for intra-school correlations

Second, from baseline to the Year 2, after which the intervention activities were completed and additional savings were no longer matched by the study, the results suggest that Bridges PLUS children showed a larger decline in the likelihood of being poor (−21.7 p.p.) and the deprivation score (−0.055) when compared to Bridges children (−13.6 p.p.; −0.042), but the difference between Bridges and Bridges PLUS children was not statistically different, suggesting that there was no additional benefit for higher savings incentives (i.e., a 1:2 match rate) at Year 2.

Third, from baseline to the Year 3, one year after intervention completion, the results suggest that Bridges and Bridges PLUS children still exhibited larger declines in the likelihood of being poor (by 16.7–20.7 p.p.) and deprivation scores (by −0.055 to −0.061) compared to the control group children. This finding supports that the persistence of the intervention effect at Year 3. Between the two intervention groups, the differences in their changes in outcomes are not statistically different, signifying limited benefits of offering additional savings incentives at Year 3.

Fourth, from baseline to the Year 4, two years after intervention completion, we find that although incidence of poverty and deprivation scores declined on average for both Bridges and Bridges PLUS children, the declines were only statistically significant for the Bridges PLUS children. This indicates that, compared to baseline, the intervention at Year 4 significantly reduced poverty incidence (by 10.7 p.p.) and deprivation scores (by 0.028) only among Bridges PLUS children, not Bridges children.¹⁵

Some indicators are only applicable to children aged below 18, and some children turned age 18 in follow-up years. To understand the extent to which this phenomenon may influence our estimates of intervention effects, we further conduct analyses on samples aged below 18 at a given time point, because two indicators in the index are only relevant to children aged below 18 (sexual risk and child labor). The results presented in Panel 2 of Table 6 use the same model as the one in Panel 1. Overall, although the effect estimates are slightly different from those presented in Panel 1 (particularly in later time points), the results reveal a similar pattern, suggesting that our findings are not biased by the reduction in applicable population in indicators at later time points.

We explored the potential mechanisms of change listed in Table 1 through examining whether the intervention affects each mechanism at Year 2 (at the end of the intervention and before measuring the long-term outcome at Year 4), and we present the results in Appendix Table 13. We found that the intervention improved children’s self-concept and family relationship, and, to a moderate extent, improved overall social support. We further found that this poverty reduction effect was not brought forth by children’s change in educational aspiration.

6.4. Robustness Analysis

We conducted two sets of robustness analysis. First, we investigated the degree to which our findings were sensitive to the weighting structure. Second, we tested the sensitivity of our analysis to the selection of indicators. Third, we studied how our findings on multidimensional poverty incidence might differ if we had adopted different poverty cutoffs.

While the choice of indicators inherently involves a set of normative decisions based on the previous literature, local context, and policy reports, the choice of weights on each dimension of multidimensional measures is often a point of contention. We overcame this limitation of index building by piloting a robustness check to assess the degree to which the impact of our intervention would be affected by the choice of weights. In our main analysis, we chose equal nested weights for each dimension. That is, we gave all the dimensions equal weights and divided the weights equally across indicators in each dimension. Given that we had equal number of indicators in each dimension, all the indicators in our multidimensional measure had equal weights. Therefore, an inevitable question is whether our results were sensitive to our choice of weights.

In order to test the robustness of our results, we first created a matrix of all possible weighting structures in increments of 5% (i.e., 15%, 15%, 40%, 30% for each of the four dimensions, respectively), and this led us to 1,771 different weighting structures. To maintain the multidimensionality of our measure, we ruled out any combination that gave a single dimension less than 15% or more than 55% weight, which left us with 165 different weighting structures. We maintained the equal weighting structure across the three indicators of each dimension. We repeated our main analysis (Panel 1 in Table 6) using all these weight combinations. For illustration purposes, we randomly chose 45 weighting structures to plot the coefficients of Bridges and Bridges Plus along with their standard errors.¹⁶ The results are shown in Fig. 2.

Fig. 2.

Robustness Check: Effect sizes across various random weighting structure (15–55). Note: The X-axis for the MPI poor outcome: differences in MPI poverty incidence from baseline to a given time point (range: −1 to 1). The X-axis for the c-vector outcome: differences in deprivation scores from baseline to a given time point (range: −1 to 1)

We found that our results for the C vector outcome, the linear deprivation score, were more robust to changes in the specification of weights in the multidimensional poverty index, compared to the incidence of poverty. From baseline to the Year 2 or Year 3 follow-up, the effects of intervention on deprivation scores among children in Bridges and Bridges PLUS groups were statistically significant regardless of weight structures. From baseline to the Year 4 follow-up, the intervention effects on deprivation scores were only predominantly statistically significant for Bridges PLUS children. In contrast, for the poverty status as an outcome, the intervention effects were more sensitive to the choice of weights. That is, under some alternative weighting structures, the intervention effects were statistically significant in some models but not others. This shows that the intervention was effective in reducing multidimensional poverty at the intensive margin but not concentrated around the poverty cutoff of 0.25.

In addition to conducting robustness analysis regarding weighting structure, we also tested the extent to which our findings are sensitive to the selection of indicators. One may concern whether the intervention effect on multidimensional poverty is largely driven by one strong indicator in the index. Therefore, we further conducted sensitivity analyses to construct alternative MPIs which exclude one indicator at a time (leaving the remaining two indicators in the same dimension sharing equal weights, which is 1.5/12 for each indicator). We used these MPIs as outcomes and ran regression models similar to those in Table 6 to examine intervention effects at Year 4. The results are presented in Appendix Table 12. Results suggest that, the intervention effects on C vector remain robust and similar to results in Table 6 regardless of the indicator excluded from the index. In models using multidimensional poverty incidence as the outcome, the intervention effects for Bridges PLUS are more sensitive to the exclusion of indicators, but this sensitivity in findings is driven by multiple indicators (e.g., sexual risk, depression, no savings, no brick house) rather than one particular strong indicator. This finding again offers support that the intervention did not exert its effects through one single channel, suggesting the importance of measuring joint deprivation.

Furthermore, we conducted additional analyses using alternative poverty cutoffs for the multidimensional poverty incidence. In the main analysis, the adopted poverty cutoff was 1/4 (deprived in a total of one dimension or three indicators). We further explored poverty patterns when the cutoffs were 0/4 and 2/4. When the poverty cutoff is 0/4 (deprived in any indicator), the average proportion of poor children is 99.8% at baseline and 97.4% at Year 4, suggesting that almost all children are poor under this definition. When the poverty cutoff is 2/4 (deprived in a total of two dimensions or six indicators), the average proportion of poor children is 3.8% at baseline and 2.1% at Year 4, suggesting that very few children are in deep poverty. We also conducted regression analysis to examine the effects of the intervention on multidimensional poverty incidence when poverty cutoffs were 0/4 and 2/4, and the results are presented in Appendix Table 10. This additional set of analyses can only be used to examine the effects of the intervention on reducing deprivation levels to zero or reducing extreme deprivation. The results show that the Bridges PLUS intervention was effective in reducing extreme poverty at Year 4, but the intervention was not particularly effective in reducing the level of deprivation to zero.

6.5. Do Multifaceted Intervention and Savings Incentives Break the Poverty Trap?

The central question is whether such interventions and savings incentives provide a robust pathway of upward mobility for poor individuals to follow to exit poverty. The experimental setting in this paper based on four-year longitudinal data (five time-points) provides a unique opportunity to test this hypothesis. Adato et al. (2006) used quantitative and qualitative evidence from South Africa to substantiate the hypothesis that polarized societies leave poor people trapped in a vicious poverty cycle. Similarly, Radeny et al. (2012) explored the poverty dynamics of Kenyan households and decomposed their dynamics into structural declines and stochastic escapes. In this section, we test two hypotheses: (a) whether multifaceted interventions (Bridges and Bridges Plus) provide a pathway for upward mobility, and (b) whether offering a higher incentive (a higher match rate) is more likely to break the poverty trap than offering a lower one.

1. We decompose our sample at each time period into four categories:

2. Upwardly mobile individuals: Those who were multidimensionally poor at baseline and no longer multidimensionally poor at time t (t = Year 1, Year 2, Year 3, and Year 4).

3. Downwardly mobile individuals: Those who were not multidimensionally poor at baseline but multidimensionally poor at time t.

4. Structurally poor individuals: Those who were poor at baseline and at time t.

5. Structurally nonpoor individuals: Those who were poor neither at baseline nor at time t.

In Panel 1 in Table 7, we test whether there are significant differences in mobility between the control and treatment (both Bridges and Bridges Plus) arms of the study. We find that children in the treatment groups were more likely to be upwardly mobile and (remained) structurally nonpoor at each time period, and this difference is statistically significant. We also find that children in the treatment groups were less likely to fall into poverty (downward mobility) or to be structurally poor. This suggests that the multifaceted economic empowering intervention can provide an effective mechanism in breaking the vicious cycle of poverty and provide a pathway out of the poverty trap.

We then asked whether children in the Bridges Plus group—who received higher savings incentives at a 2:1 match rate—have a greater likelihood of escaping poverty than those in Bridges (1:1 match rate). In Panel 2 of Table 7, we present the results of the poverty dynamics between the two treatment arms. In most cases, we could not reject the hypothesis that the effects of Bridges and Bridges Plus were equal. This is consistent with our main findings in Table 6.

7. Discussion and Conclusion

This study has examined the effects of a multifaceted intervention, comprising a savings account with matched savings, mentorship, and trainings, in reducing multidimensional poverty among orphan children in rural Uganda. We constructed a multidimensional child poverty measure that is relevant to the local Ugandan context. This measure included four dimensions: health, assets, housing, and behavioral risks. We examined the intervention effects through OLS regression and difference-in-difference models. We found that this multifaceted economic empowering intervention alleviated the incidence of multidimensional poverty by more than 10 percentage points, with the intervention effects being stronger right after intervention completion (Year 2 and 3) than two years after intervention completion (Year 4).

This study extends the body of literature examining the effectiveness of multifaceted interventions in poverty alleviation by extending the scope of examination to children and youth populations. The findings from this study suggest that a multifaceted intervention approach can be effective in alleviating multidimensional poverty in the short term as well as the long term (two years after intervention completion) for children. However, the intervention effects exhibited a declining trend, with the poverty alleviation effect being stronger at the end of the intervention period and at one year after intervention completion. Furthermore, from a poverty dynamic perspective, our findings revealed that the multifaceted intervention was effective in avoiding the poverty trap or facilitating exiting poverty.

This study is among the first studies to examine the effects of a poverty alleviation intervention using a multidimensional poverty outcome measure (Malaeb & Uzor, 2017; Robano & Smith, 2014). We have illustrated that multidimensional impact evaluation is necessary and adequate for two reasons. First, a multidimensional measure can capture the overlapping deprivation that an individual experiences, which conventional analyses by various single outcome measures are unable to uncover. We argue that a multidimensional poverty measure is a useful construct for capturing joint deprivation and can be an essential measure to complement with analyses by indicator. Second, a multidimensional poverty measure is particularly adequate in studying child poverty. Income poverty measures are unable to precisely capture the household resources available to children, and researchers have long argued that child poverty should be conceptualized not only as deprivation in the monetary aspect but also in terms of the resources available for them to survive and realize their capacity. Furthermore, we contribute to the literature on multidimensional poverty measures by administering robustness analyses to test the sensitivity of results to various weighting structures for poverty dimensions. Most studies on multidimensional poverty have adopted an equal weighting structure across dimensions due to the difficulty of deriving and justifying alternative theory-driven weighting structures. We propose a new approach in examining how the findings may be sensitive to the range of weighting structures, and we have exhausted the potential weighting combinations across dimensions and visually illustrated the sensitivity of the findings. This approach enables critically investigating the influence of weighting choices on results.

This study has further examined the effectiveness of a higher level of savings incentives in alleviating multidimensional poverty. We compared a bundled intervention with a 1:1 match rate (100%) versus a 1:2 match rate (200%). Both study groups showed similar levels of intervention effects across most time points, and the effect estimates at Year 4 were stronger for the Bridges PLUS children. We found that Bridges children (1:1 match rate) showed some signs of increase in MPI incidence and deprivation score at the Year 4 follow-up (two years post intervention completion) but Bridges PLUS children did not (1:2 match rate).¹⁷ Past studies have shown that low interest rates (3–5%) had a smaller effect in reducing poverty than higher interest rates (e.g., 12–20%; Karlan & Zinman, 2014; Kast et al., 2012; Schaner, 2016). Findings from our study suggest that an already-high match rate (1:1; 100%) may suffice to reduce multidimensional poverty, but an even higher match rate (1:2; 200%) may lead to more sustained positive outcomes. This finding can also be explained by prior studies on savings outcomes in relation to savings incentives. Wang et al. (2018), based on the same intervention, examined whether 1:2 and 1:1 savings match rates led to different savings outcomes among orphan children in Uganda. They found that two years after baseline, children receiving a higher savings incentive had a higher level of self-reported savings than those receiving a lower incentive, and such effect on self-reported savings was sustained until four years after baseline (see Appendix Table 14). Additionally, the take-up rates of other intervention components are higher for Bridges PLUS children (see Appendix Table 15). These differences suggest that children in the Bridges PLUS group accumulated more resources for development; hence, children in the Bridges PLUS group exhibited lower multidimensional poverty incidence and deprivation levels at Year 4.

Another key finding was the observed abrupt decline of the intervention effect at the Year 4 follow-up (two years after intervention completion). The intervention effects were strong at the Year 2 and Year 3 follow-ups (right after intervention completion), yet at Year 4, the effect of intervention in alleviating multidimensional poverty reduced to statistical insignificance for children in the Bridges group. For children in the Bridges PLUS group, the effect of the intervention in reducing poverty incidence diminished by 10 percentage points. Had this study only observed children’s outcomes until intervention completion or one year after intervention completion, we would have drawn a conclusion about strong and persistent intervention effects. Through following children for a longer time frame, however, we uncovered a trajectory of diminishing intervention effects over time. This finding is an important reminder to researchers and practitioners regarding the needs to inquire or investigate intervention effects for a longer time frame. Understanding the longer-term effects of an intervention or the evolution of intervention effects post intervention completion is crucial for informing the improvement of intervention design or the creation of booster follow-up components. Further follow-up studies are warranted because it is still premature to conclude the effectiveness of multifaceted interventions or higher savings incentives. For example, although Bridges and Bridges PLUS children deposited similar amounts of savings into bank accounts (Wang et al., 2018), Bridges PLUS children received a higher amount of matched savings and consequently accumulated more financial resources. As participants in this multifaceted program are only allowed to spend their accumulated matched savings for the purpose of paying educational expenses or starting a new business, the benefits of higher savings incentives may take longer to realize. A future study should follow these children for a longer period (e.g., five to 10 years after intervention completion) to acquire a firmer understanding of the effectiveness of multifaceted interventions and higher savings incentives in the longer term.

Appendix

See Tables 8, 9, 10, 11, 12, 13, 14, 15, 16 and 17.

See Figs. 3 and 4

Table 8.

Regression results on multidimensional poverty incidence and deprivation score with control variables

Time points	Year 1 – Baseline			Year 2 – Baseline			Year 3 – Baseline			Year 4 – Baseline
Models	(1)	(2)	(3)	(1)	(2)	(3)	(1)	(2)	(3)	(1)	(2)	(3)
Panel 1: MPI
Bridges	−0.142** (0.053)	−0.142** (0.054)	−0.143*** (0.050)	−0.134** (0.053)	−0.126** (0.053)	−0.121** (0.049)	−0.168** (0.063)	−0.168** (0.064)	−0.165*** (0.060)	−0.028 (0.046)	−0.033 (0.046)	−0.029 (0.046)
Bridges PLUS	−0.114** (0.049)	−0.114** (0.051)	−0.125** (0.047)	−0.214*** (0.047)	−0.213*** (0.046)	−0.216*** (0.043)	−0.204*** (0.060)	−0.203*** (0.060)	−0.207*** (0.057)	−0.106*** (0.034)	−0.105*** (0.035)	−0.103*** (0.036)
Double orphan	−0.030 (0.051)	−0.044 (0.050)	−0.039 (0.051)	−0.010 (0.054)	−0.024 (0.053)	−0.023 (0.053)	−0.081 (0.059)	−0.093 (0.059)	−0.090 (0.059)	−0.060 (0.047)	−0.057 (0.047)	−0.057 (0.047)
Primary caregiver (Ref: Parent)
Grandparent	0.010 (0.040)	0.026 (0.042)	0.040 (0.042)	0.024 (0.046)	0.028 (0.047)	0.026 (0.049)	0.091* (0.047)	0.100** (0.048)	0.091* (0.049)	0.056 (0.047)	0.045 (0.049)	0.031 (0.049)
Other relative	0.114** (0.049)	0.137** (0.056)	0.143** (0.057)	0.071 (0.062)	0.073 (0.067)	0.071 (0.071)	0.090* (0.050)	0.103* (0.056)	0.097 (0.060)	0.048 (0.048)	0.034 (0.052)	0.023 (0.055)
Age		−0.018 (0.013)	−0.020 (0.013)		0.016 (0.014)	0.012 (0.014)		−0.003 (0.016)	−0.008 (0.016)		−0.025 (0.016)	−0.026 (0.016)
Female		−0.125*** (0.036)	−0.118*** (0.037)		−0.115** (0.052)	−0.116** (0.053)		−0.136*** (0.045)	−0.131*** (0.046)		−0.048 (0.045)	−0.046 (0.045)
Years living in the households		0.004 (0.004)	0.004 (0.004)		−0.000 (0.004)	−0.001 (0.004)		0.002 (0.004)	0.002 (0.004)		−0.005 (0.004)	−0.006 (0.004)
Household size			−0.024*** (0.009)			0.001 (0.014)			0.017 (0.014)			0.020 (0.014)
Number of children			0.037*** (0.013)			0.014 (0.015)			−0.007 (0.018)			−0.015 (0.015)
Caregiver: employed			−0.098** (0.046)			−0.054 (0.053)			−0.098* (0.054)			−0.021 (0.048)
Constant	−0.069 (0.045)	0.195 (0.172)	0.276 (0.173)	−0.078 (0.048)	−0.221 (0.194)	−0.198 (0.183)	−0.133** (0.057)	−0.040 (0.234)	−0.036 (0.227)	−0.209*** (0.035)	0.180 (0.203)	0.128 (0.211)
N	1227	1225	1221	1112	1111	1107	1118	1115	1111	1169	1166	1162
Bridges =BridgesPlus (p-value)	0.580	0.580	0.723	0.111	0.079	0.048	0.490	0.511	0.430	0.054	0.069	0.065
Bridges = BridgesPlus = 0 (p-value)	0.022	0.029	0.010	0.000	0.000	0.000	0.005	0.005	0.002	0.006	0.009	0.012
Panel 2: C vector
Bridges	−0.048*** (0.013)	−0.047*** (0.014)	−0.047*** (0.012)	−0.041*** (0.011)	−0.038*** (0.010)	−0.037*** (0.010)	−0.054*** (0.014)	−0.052*** (0.014)	−0.051*** (0.013)	−0.009 (0.011)	−0.010 (0.011)	−0.010 (0.011)
Bridges PLUS	−0.033** (0.013)	−0.033** (0.013)	−0.036*** (0.012)	−0.054*** (0.011)	−0.054*** (0.011)	−0.054*** (0.011)	−0.060*** (0.014)	−0.059*** (0.014)	−0.060*** (0.014)	−0.027** (0.011)	−0.027** (0.011)	−0.026** (0.011)
Double orphan	−0.001 (0.009)	−0.004 (0.009)	−0.003 (0.010)	0.009 (0.011)	0.004 (0.010)	0.004 (0.010)	−0.013 (0.014)	−0.016 (0.014)	−0.016 (0.014)	−0.005 (0.013)	−0.005 (0.013)	−0.005 (0.013)
Primary caregiver (Ref: Parent)
Grandparent	0.003 (0.011)	0.006 (0.011)	0.008 (0.011)	0.004 (0.012)	0.006 (0.012)	0.006 (0.012)	0.022* (0.012)	0.023* (0.012)	0.021 (0.013)	0.009 (0.013)	0.006 (0.014)	0.003 (0.014)
Other relative	0.024** (0.012)	0.028** (0.013)	0.028** (0.013)	0.016 (0.013)	0.019 (0.014)	0.017 (0.014)	0.023* (0.013)	0.023 (0.015)	0.021 (0.015)	0.012 (0.012)	0.010 (0.013)	0.006 (0.014)
Age		0.002 (0.003)	0.002 (0.003)		0.008** (0.003)	0.008** (0.004)		0.006* (0.003)	0.006* (0.003)		−0.006 (0.004)	−0.006 (0.004)
Female		−0.021** (0.009)	−0.018** (0.009)		−0.029*** (0.011)	−0.030*** (0.011)		−0.033*** (0.011)	−0.031*** (0.011)		−0.016 (0.011)	−0.016 (0.011)
Years living in the households		0.001 (0.001)	0.001 (0.001)		0.001 (0.001)	0.001 (0.001)		−0.000 (0.001)	−0.000 (0.001)		−0.001 (0.001)	−0.001 (0.001)
Household size			−0.004** (0.002)			0.000 (0.003)			0.002 (0.003)			0.004 (0.003)
Number of children			0.007*** (0.003)			0.003 (0.004)			0.002 (0.004)			−0.002 (0.004)
Caregiver: employed			−0.025** (0.011)			−0.002 (0.010)			−0.018 (0.011)			−0.000 (0.012)
Constant	−0.020 (0.013)	−0.041 (0.039)	−0.026 (0.040)	−0.026*** (0.009)	−0.117*** (0.043)	−0.122*** (0.042)	−0.034*** (0.012)	−0.098** (0.047)	−0.105** (0.046)	−0.065*** (0.010)	0.025 (0.051)	0.005 (0.052)
Bridges = BridgesPlus = 0 (p-value)	0.022	0.029	0.010	0.000	0.000	0.000	0.005	0.005	0.002	0.006	0.009	0.012
Panel 2: C vector
Bridges	−0.048*** (0.013)	−0.047*** (0.014)	−0.047*** (0.012)	−0.041*** (0.011)	−0.038*** (0.010)	−0.037*** (0.010)	−0.054*** (0.014)	−0.052*** (0.014)	−0.051*** (0.013)	−0.009 (0.011)	−0.010 (0.011)	−0.010 (0.011)
Bridges PLUS	−0.033** (0.013)	−0.033** (0.013)	−0.036*** (0.012)	−0.054*** (0.011)	−0.054*** (0.011)	−0.054*** (0.011)	−0.060*** (0.014)	−0.059*** (0.014)	−0.060*** (0.014)	−0.027** (0.011)	−0.027** (0.011)	−0.026** (0.011)
Double orphan	−0.001 (0.009)	−0.004 (0.009)	−0.003 (0.010)	0.009 (0.011)	0.004 (0.010)	0.004 (0.010)	−0.013 (0.014)	−0.016 (0.014)	−0.016 (0.014)	−0.005 (0.013)	−0.005 (0.013)	−0.005 (0.013)
Primary caregiver (Ref: Parent)
Grandparent	0.003 (0.011)	0.006 (0.011)	0.008 (0.011)	0.004 (0.012)	0.006 (0.012)	0.006 (0.012)	0.022* (0.012)	0.023* (0.012)	0.021 (0.013)	0.009 (0.013)	0.006 (0.014)	0.003 (0.014)
Other relative	0.024** (0.012)	0.028** (0.013)	0.028** (0.013)	0.016 (0.013)	0.019 (0.014)	0.017 (0.014)	0.023* (0.013)	0.023 (0.015)	0.021 (0.015)	0.012 (0.012)	0.010 (0.013)	0.006 (0.014)
Age		0.002 (0.003)	0.002 (0.003)		0.008** (0.003)	0.008** (0.004)		0.006* (0.003)	0.006* (0.003)		−0.006 (0.004)	−0.006 (0.004)
Female		−0.021** (0.009)	−0.018** (0.009)		−0.029*** (0.011)	−0.030*** (0.011)		−0.033*** (0.011)	−0.031*** (0.011)		−0.016 (0.011)	−0.016 (0.011)
Years living in the households		0.001 (0.001)	0.001 (0.001)		0.001 (0.001)	0.001 (0.001)		−0.000 (0.001)	−0.000 (0.001)		−0.001 (0.001)	−0.001 (0.001)
Household size			−0.004** (0.002)			0.000 (0.003)			0.002 (0.003)			0.004 (0.003)
Number of children			0.007*** (0.003)			0.003 (0.004)			0.002 (0.004)			−0.002 (0.004)
Caregiver: employed			−0.025** (0.011)			−0.002 (0.010)			−0.018 (0.011)			−0.000 (0.012)
Constant	−0.020 (0.013)	−0.041 (0.039)	−0.026 (0.040)	−0.026*** (0.009)	−0.117*** (0.043)	−0.122*** (0.042)	−0.034*** (0.012)	−0.098** (0.047)	−0.105** (0.046)	−0.065*** (0.010)	0.025 (0.051)	0.005 (0.052)
N	1227	1225	1221	1112	1111	1107	1118	1115	1111	1169	1166	1162
Bridges = BridgesPlus (p-value)	0.174	0.205	0.293	0.274	0.160	0.124	0.617	0.555	0.433	0.121	0.142	0.149
Bridges = BridgesPlus = 0 (p-value)	0.003	0.005	0.001	0.000	0.000	0.000	0.000	0.000	0.000	0.059	0.075	0.078

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

MPI poor = H: Headcount ratio or incidence of multidimensional poverty (MPI); C vector = C vector is a simultaneous deprivation score for each child. It is derived through summing deprivation status of each indicator multiplied by weight (1/12), which is a weighted sum of overlapping deprivations. All models cluster the robust errors at the school level to adjust for intra-school correlations

Table 9.

Regression results on multidimensional poverty incidence and deprivation score using multilevel models

Outcomes	MPI poor	C vector	MPI poor	C vector
Sample	All sample	All sample	Sample age < 18	Sample age < 18
Models	Logistic	Linear	Logistic	Linear
	Coefficient	Coefficient	Coefficient	Coefficient
Group (Ref: Control)
Group χ2(2)	2.53	1.78	2.55	1.80
Bridges	−0.458 (0.292)	−0.017 (0.013)	−0.456 (0.289)	−0.017 (0.013)
Bridges PLUS (Bridges +)	−0.160 (0.230)	−0.008 (0.011)	−0.160 (0.227)	−0.008 (0.011)
Time (Ref: Baseline)
Time χ2(4)	49.47***	80.63***	22.02***	34.75***
Year 1 (Y1)	−0.313 (0.230)	−0.012 (0.010)	−0.292 (0.228)	−0.012 (0.010)
Year 2 (Y2)	−0.434** (0.191)	−0.022*** (0.006)	−0.393* (0.205)	−0.021*** (0.006)
Year 3 (Y3)	−0.744** (0.305)	−0.028*** (0.010)	−0.678** (0.300)	−0.026** (0.010)
Year 4 (Y4)	−1.179*** (0.198)	−0.058*** (0.007)	−0.911*** (0.214)	−0.044*** (0.008)
Group X time
Group X Time χ2(8)	27.48***	34.38***	27.37***	33.75***
Bridges X Y1	−0.688** (0.317)	−0.045*** (0.013)	−0.691** (0.315)	−0.044*** (0.013)
Bridges + X Y1	−0.658** (0.276)	−0.033*** (0.012)	−0.654** (0.275)	−0.033*** (0.012)
Bridges X Y2	−0.616** (0.300)	−0.035*** (0.010)	−0.648** (0.311)	−0.035*** (0.010)
Bridges + X Y2	−1.176*** (0.257)	−0.051*** (0.011)	−1.197*** (0.273)	−0.051*** (0.011)
Bridges X Y3	−0.859** (0.404)	−0.048*** (0.013)	−1.026** (0.403)	−0.051*** (0.014)
Bridges + X Y3	−1.168*** (0.381)	−0.057*** (0.013)	−1.241*** (0.375)	−0.060*** (0.013)
Bridges X Y4	−0.195 (0.317)	−0.009 (0.011)	−0.330 (0.382)	−0.016 (0.013)
Bridges + X Y4	−0.888*** (0.266)	−0.032*** (0.011)	−1.167*** (0.327)	−0.044*** (0.013)
Constant	1.054*** (0.203)	0.347*** (0.009)	1.040*** (0.201)	0.347*** (0.009)
Variance of school random intercepts	0.246 (0.084)	0.001 (0.000)	0.247 (0.087)	0.001 (0.000)
Variance of child random slopes (time)	0.259 (0.047)	0.001 (0.000)	0.245 (0.053)	0.000 (0.000)
Variance of child random intercepts	3.255 (0.568)	0.007 (0.001)	2.837 (0.534)	0.007 (0.001)
Covariance of child slopes and intercepts	−0.587 (0.127)	−0.001 (0.000)	−0.471 (0.122)	−0.001 (0.000)
Variance of residual	N/A	0.008 (0.000)	N/A	0.008 (0.000)
Outcomes	MPI poor	C vector	MPI poor	C vector
Observations	6231	6231	5821	5821
N	1382	1382	1381	1381
Pairwise comparisons
Y0 bridges versus control	−0.458 (0.292)	−0.017 (0.013)	−0.456 (0.289)	−0.017 (0.013)
Y0 bridges + versus control	−0.160 (0.230)	−0.008 (0.011)	−0.160 (0.227)	−0.008 (0.011)
Y0 bridges + versus bridges	0.298 (0.247)	0.008 (0.011)	0.296 (0.244)	0.009 (0.011)
Y1 bridges versus control	−1.147*** (0.285)	−0.061*** (0.015)	−1.147*** (0.285)	−0.061*** (0.015)
Y1 bridges + versus control	−0.818** (0.265)	−0.042** (0.014)	−0.815** (0.267)	−0.042** (0.014)
Y1 bridges + versus bridges	0.328 (0.256)	0.020 (0.012)	0.333 (0.261)	0.020 (0.013)
Y2 bridges versus control	−1.074*** (0.285)	−0.052*** (0.014)	−1.104*** (0.294)	−0.052*** (0.014)
Y2 bridges + versus control	−1.336*** (0.281)	−0.059*** (0.013)	−1.357*** (0.292)	−0.060*** (0.013)
Y2 bridges + versus bridges	−0.262 (0.243)	−0.008 (0.013)	−0.254 (0.253)	−0.008 (0.013)
Y3 bridges versus control	−1.317** (0.381)	−0.065*** (0.015)	−1.482*** (0.393)	−0.068*** (0.016)
Y3 bridges + versus control	−1.328*** (0.379)	−0.066*** (0.015)	−1.401*** (0.381)	−0.068*** (0.015)
Y3 bridges + versus bridges	−0.011 (0.337)	−0.001 (0.014)	0.081 (0.354)	−0.000 (0.014)
Y4 bridges versus control	−0.654 (0.324)	−0.025 (0.013)	−0.786 (0.405)	−0.033 + (0.014)
Y4 bridges + versus control	−1.048*** (0.289)	−0.041** (0.013)	−1.328*** (0.342)	−0.052*** (0.013)
Y4 bridges + versus bridges	−0.395 (0.310)	−0.015 (0.013)	−0.541 (0.426)	−0.020 (0.015)

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

“Multilevel models have the advantage of being statistically efficient in accounting for the clustering nature of data, where multiple individuals are nested within each school, and multiple observations across time are nested within each individual. Specifically, children from the same school are likely to be correlated with each other, and the repeated measures for each individual are likely to be not independent. Multilevel models allow us to estimate the school-level random intercepts, individual-level random intercepts, and the individual random slopes across survey time points. In each model, we included study group status (Bridges and Bridges PLUS, with control group serving as the reference group), time dummy(ies), and their interactions. The interactions allowed us to identify the intervention effects accounting for any baseline differences in outcomes of interest across the three study arms before the intervention as well as accounting for changes in outcomes across time shared by the three arms (Wang et al., 2018).” The continuous outcome, C vector, were estimated using linear mixed models (LMMs) with the resulting regression coefficients estimating the mean change in the outcome per unit change in the predictor via the -mixed- command in Stata 14. The binary outcomes, MPI poverty incidence, were estimated using a logistic GLMM via the Stata -melogit- command with the coefficients being reported. We modeled the covariance structure to be unstructured, which made the least assumptions regarding the covariance structure, and we estimated the cluster-adjusted robust standard errors with school ID treated as the clustering variable

Table 10.

Regression results on multidimensional poverty incidence with alternative poverty cutoffs (poverty cutoff = 0/4 and 2/4)`

Outcome	MPI poor (poverty cutoff=0/4)				MPI poor (poverty cutoff=2/4)
Time points	Year 1 – baseline	Year 2 – baseline	Year 3 – baseline	Year 4 – baseline	Year 1 – baseline	Year 2 – baseline	Year 3 – baseline	Year 4 – baseline
Panel 1: All sample
Bridges	−0.020*** (0.007)	−0.013 (0.009)	−0.010(0.010)	0.005 (0.009)	−0.028* (0.015)	−0.029* (0.015)	−0.054*** (0.019)	−0.008 (0.015)
Bridges PLUS	−0.013* (0.007)	−0.015 (0.009)	−0.024** (0.010)	−0.021 (0.015)	−0.022 (0.016)	−0.022 (0.015)	−0.050*** (0.016)	−0.026* (0.013)
Constant	0.000 (0.003)	0.000 (0.005)	−0.003 (0.004)	−0.019** (0.007)	0.005 (0.011)	0.000 (0.010)	0.026** (0.012)	−0.007 (0.010)
N	1227	1112	1118	1169	1227	1112	1118	1169
Bridges = BridgesPlus (p-value)	0.420	0.861	0.249	0.062	0.717	0.686	0.850	0.207
Bridges = BridgesPlus = 0 (p-value)	0.010	0.198	0.057	0.169	0.156	0.142	0.004	0.141
Panel 2: Sample aged < 18
Bridges	−0.021*** (0.007)	−0.013 (0.009)	−0.004 (0.008)	−0.002 (0.009)	−0.028* (0.015)	−0.026* (0.015)	−0.063*** (0.019)	−0.011 (0.019)
Bridges PLUS	−0.013* (0.007)	−0.015 (0.009)	−0.026** (0.011)	−0.019 (0.014)	−0.020 (0.016)	−0.020 (0.016)	−0.062*** (0.017)	−0.036* (0.020)
Constant	0.000 (0.003)	0.000 (0.005)	−0.003 (0.005)	−0.013* (0.007)	0.005 (0.011)	−0.003 (0.010)	0.036*** (0.013)	0.003 (0.016)
N	1224	1098	1047	866	1224	1098	1047	866
Bridges = BridgesPlus (p-value)	0.421	0.856	0.065	0.206	0.605	0.699	0.966	0.131
Bridges = BridgesPlus = 0 (p-value)	0.010	0.197	0.061	0.382	0.165	0.192	0.001	0.151

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

Table 11.

ANCOVA results on multidimensional poverty incidence and deprivation score

	Year 1 – baseline		Year 2 – baseline		Year 3 – baseline		Year 4 – baseline
	MPI poor	C vector	MPI poor	C vector	MPI Poor	C vector	MPI Poor	C vector
Bridges	−0.197*** (0.035)	−0.059*** (0.008)	−0.189*** (0.036)	−0.051*** (0.009)	−0.242*** (0.036)	−0.071*** (0.009)	−0.101*** (0.035)	−0.027*** (0.009)
Bridges PLUS	−0.148*** (0.032)	−0.041*** (0.008)	−0.239*** (0.034)	−0.060*** (0.008)	−0.244*** (0.034)	−0.072*** (0.008)	−0.153*** (0.033)	−0.039*** (0.009)
MPI poor at baseline	0.249*** (0.029)		0.183*** (0.030)		0.121*** (0.030)		0.129*** (0.029)
C Vector at baseline		0.426*** (0.028)		0.309*** (0.030)		0.249*** (0.032)		0.225*** (0.032)
Constant	0.466*** (0.030)	0.187*** (0.011)	0.489*** (0.032)	0.220*** (0.012)	0.495*** (0.031)	0.238*** (0.013)	0.403*** (0.031)	0.211*** (0.013)
N	1227	1227	1112	1112	1118	1118	1169	1169

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

Table 12.

Regression results on deprivation score and multidimensional poverty incidence—excluding one indicator: Year 4 – baseline among sample aged < 18

Outcome	C vector (deprivation score)
Excluded indicator	Malnutrition	Sexual risk	Depression	No savings	Lack of com	Few clothing	Water sources	No brick house	Electricity	Child labor	Drinking	School dropout
Bridges	−0.009 (0.012)	−0.006 (0.014)	−0.010 (0.014)	0.001 (0.012)	−0.004 (0.013)	−0.022 (0.016)	−0.011 (0.015)	−0.003 (0.013)	−0.011 (0.014)	−0.013 (0.014)	−0.018 (0.013)	0.005 (0.013)
Bridges PLUS	−0.042*** (0.011)	−0.040*** (0.013)	−0.040*** (0.014)	−0.029** (0.012)	−0.040*** (0.013)	−0.054*** (0.016)	−0.036** (0.014)	−0.044*** (0.013)	−0.043*** (0.013)	−0.048*** (0.013)	−0.049*** (0.013)	−0.027** (0.012)
Constant	−0.040*** (0.007)	−0.056*** (0.009)	−0.048*** (0.010)	−0.047*** (0.007)	−0.064*** (0.009)	−0.032*** (0.011)	−0.050*** (0.010)	−0.066*** (0.009)	−0.028*** (0.009)	−0.037*** (0.008)	−0.027*** (0.009)	−0.079*** (0.010)
N	930	930	930	930	930	930	930	930	930	930	930	930
Outcome	MPI poor (poverty incidence)
Bridges	−0.032 (0.046)	0.008 (0.044)	−0.004 (0.046)	0.049 (0.043)	−0.024 (0.049)	−0.030 (0.048)	−0.028 (0.047)	0.042 (0.056)	−0.061 (0.053)	−0.051 (0.068)	−0.051 (0.056)	−0.011 (0.058)
Bridges PLUS	−0.125*** (0.046)	−0.069 (0.043)	−0.061 (0.043)	−0.046 (0.044)	−0.147*** (0.051)	−0.134*** (0.049)	−0.096** (0.040)	−0.066 (0.043)	−0.147*** (0.051)	−0.167*** (0.050)	−0.133*** (0.040)	−0.117*** (0.043)
Constant	−0.053 (0.032)	−0.201*** (0.023)	−0.185*** (0.023)	−0.163*** (0.025)	−0.232*** (0.033)	−0.172*** (0.027)	−0.207*** (0.019)	−0.292*** (0.024)	−0.056 (0.035)	−0.091** (0.039)	−0.088*** (0.030)	−0.213*** (0.034)
N	930	930	930	930	930	930	930	930	930	930	930	930

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

C vector = a simultaneous deprivation score for each child, which is a weighted sum of overlapping deprivations. For the dimension that the excluded indicator belongs to, the weight for each of the other two indicators is 1.5/12. For the dimensions that no indicator is excluded, the weight for each indicator remains to be 1/12. All models cluster the standard errors at the school level to adjust for intra-school correlations. MPI poor = H: Headcount ratio or incidence of multidimensional poverty (MPI). For the dimension that the excluded indicator belongs to, the weight for each of the other two indicators is 1.5/12. For the dimensions that no indicator is excluded, the weight for each indicator remains to be 1/12. All models cluster the standard errors at the school level to adjust for intra-school correlations

Table 13.

Regression results on mechanisms of change

Outcome	Educational aspiration at year 2 (Ref. dropped out before secondary school)				Self-efficacy at year 2 (standardized)	Self-concept at year 2 (standardized)	Social support at year 2 (standardized)	Family relationship at year 2 (standardized)
	Some secondary education until S4	Complete S6 or attend technical college	Complete a university degree	Pursue graduate studies
Model	Multinomial logit				OLS	OLS	OLS	OLS
Bridges	−0.016(0.300)	0.303 (0.296)	0.033 (0.303)	0.304 (0.357)	0.203+ (0.120)	0.232* (0.093)	0.200* (0.088)	0.222** (0.081)
Bridges PLUS	−0.154 (0.273)	0.059 (0.291)	−0.002 (0.303)	−0.032 (0.348)	0.114 (0.121)	0.178+ (0.097)	0.068 (0.091)	0.193* (0.086)
Constant	−0.693** (0.227)	−0.288+ (0.155)	0.314+ (0.189)	0.022 (0.213)	−0.100 (0.112)	−0.131* (0.063)	−0.081 (0.063)	−0.134* (0.062)
N	1221				1211	1219	1219	1220
Bridges = BridgesPlus (p-value)	−				0.156	0.594	0.149	0.707
Bridges = BridgesPlus = 0 (p-value)	0.586				0.149	0.039	0.079	0.022

School-level cluster robust standard errors in parentheses

⁺p < 0.10;

^*p < 0.05;

^**p < 0.01;

^***p < 0.001

Self-efficacy is measured by the 29-item Youth Self-Efficacy Survey (Earls & Visher, 1997). Self-concept is measured by the 20-item Tennessee Self-Concept Scale (Fitts & Warren, 1996). Social support is measured by the 24-item Social Support Scale for Children (Harter, 1985). Family relationship is adapted from the acceptance/involvement subscale (9-item) and psychological autonomy-granting subscale (9-item) of the Parenting Styles Index (Steinberg et al., 1992, 1994)

Table 14.

The intervention effect on self-reported savings

	Control	Bridges	Bridges PLUS	Wald test: group X time (8)	Bridges PLUS versus bridges	Control versus bridges	Control versus bridges PLUS
	Mean	Mean	Mean		P-value	P-value	P-value
Self-reported savings amount (nominal Ugandan Shillings $)
Baseline	1999.38	2097.48	2363.73	χ2 = 290.21	0.759	0.960	0.474
12 months	5588.71	20,081.64	22,736.49	p < 0.001	0.145	0.000	0.000
24 months	7251.52	31,523.92	37,323.53		0.040	0.000	0.000
36 months	19,817.86	49,315.29	66,198.08		0.041	0.000	0.000
48 months	46,812.50	67,330.06	89,751.23		0.021	0.000	0.000

Comparisons were conducted using the following model: Outcome = intercept + α₁ Bridges + α₂ BridgesPLUS + ∑ α₃ Time + ∑ α₄ Bridges x Time + ∑ α₅ BridgesPLUS x Time + ε. We administered a Wald test to test the assumption that all group by time interaction dummies are jointly equal to zero (i.e., an omnibus test for interaction). For significant omnibus interactions, we further decomposed the group-by-time point interactions by comparing the between group difference at each time point using the -margins- command with a Sidak p-value adjustment for multiple comparisons

Table 15.

Take-up rate of mentorship and financial trainings

	Bridges (%)	Bridges PLUS (%)	T-test
Mentorship	60.9	67.0	−2.86**
Workshop on micro-enterprise development	61.1	64.3	−1.25

⁺p < 0.1;

^*p < 0.05;

^**p < 0.01

Table 16.

Correlation coefficients of weighted dimensional scores

Panel 1: baseline	Health	Assets	Housing	Behavioral risks
Health	1
Assets	0.11	1
Housing	0.08	0.13	1
Behavioral risks	0.08	−0.12	0.03	1
Panel 2: year 4	Health	Assets	Housing	Behavioral risks
Health	1
Assets	0.18	1
Housing	0.14	0.24	1
Behavioral risks	−0.04	−0.01	−0.02	1

Table 17.

Results from Cramer’s V, redundancy tests, and Chi-square tests between MPI indicators within each dimension

Dimension	Indicators	Cramer’s V	Redundancy test	Chi-square (1)	P-value
Health	Malnutrition versus sexual risk	−0.02	0.77	2.62	0.11
	Malnutrition versus depression	0.04	0.90	11.12	0.00
	Sexual risk versus depression	0.03	0.08	6.84	0.01
Assets	No savings versus few clothing and shoes	0.13	0.63	109.30	0.00
	No savings versus lack of communication or transportation	0.09	0.55	48.34	0.00
	Few clothing and shoes versus lack of communication or transportation	0.11	0.51	71.78	0.00
Housing	Distant water sources versus no brick house	−0.02	0.24	1.60	0.21
	Distant water sources versus electricity	0.12	0.87	91.61	0.00
	No brick house versus electricity	0.00	0.78	0.04	0.84
Behavioral risks	Child labor versus drinking	0.03	0.19	5.12	0.02
	Child labor versus school dropout	0.35	0.48	769.30	0.00
	Drinking versus school dropout	−0.03	0.74	4.13	0.04

Fig. 3.

Changes in M₀ for the different study groups

Fig. 4.

Robustness Check: Effect Sizes Across Various Random Weighting Structure (10–90). Note: The X-axis for the MPI poor outcome: differences in MPI poverty incidence from baseline to a given time point (range: −1 to 1). The X-axis for the C vector outcome: differences in deprivation scores from baseline to a given time point (range: −1 to 1). Bars represent the 95% confidence interval