Poor Families Striving to Save in Matched Children’s Savings Accounts: Findings from a Randomized Experimental Design in Uganda

Abstract

This study examines participants’ savings in children’s savings accounts (CSAs) set up for AIDS-orphaned children ages 10–15 in Uganda. Using a cluster randomized experimental design, we examine the extent to which families participating in a CSA program report more savings than their counterparts not participating in the program, explore the extent to which families who participate in the CSA program report using formal financial institutions compared with families who do not have a CSA, and consider whether families participating in the CSA program bring new money into the CSA or whether they reshuffle existing household assets. We find that participating in a CSA increased families’ likelihood to report having saved money. However, our results show no intervention effect either on the amount of self-reported savings or on the likelihood of using formal financial institutions. Further research is needed to understand whether use of a CSA helps families generate new wealth.

Introduction

Despite ongoing global efforts to reduce poverty, the struggle continues, particularly in sub-Saharan Africa, which is home to 29 percent of the world’s poor (Artadi and Sala-i-Martin 2003; Go et al. 2007; Fosu 2009). Although poverty is decreasing around the world, it remains mainly unchanged in sub-Saharan Africa. The region still has the highest poverty rate in the world, and almost every second person in the region lives on less than $1.25 a day (UN 2012; World Bank 2012). Additionally, in terms of non-economic dimensions of poverty, more than 29 percent of the 875 million people living in sub-Saharan Africa have no access to safe drinking water (UNICEF 2011), 23 percent of school-age children in the region are out of school (UNESCO 2010, 2011), and an average of eight children under the age of 5 in the sub-Saharan African region die from preventable diseases every minute (WHO 2011).

Until recently, most poverty-reduction initiatives, in both developed countries and developing countries, have focused on short-term income maintenance rather than on long-term asset building and economic empowerment for low-income individuals, which may partially explain the persistence of poverty (Sherraden 2001; Paxton 2003). In this context, assets are broadly defined as a stock of inputs, such as money or property, which can produce positive economic returns (Schreiner and Sherraden 2007; Nam, Huang, and Sherraden 2008). Recent innovations to target poverty include programs supporting the modest accumulation of financial assets by the poor. This has been done through informal institutions, such as rotating saving and credit associations (groups of individuals who contribute money to a common pool and use it to grant larger, one-time loans to each member) and self-help groups (groups of individuals who contribute money to a bank account in order to access bank loans). It has also been done through formal financial institutions, such as microfinance (the extension of credit and financial services to low-income clients), contractual savings (savings accounts that require participants to save a set amount for a set period of time), and matched savings accounts specifically targeted to the poor. One type of matched savings account is an Individual Development Account (IDA), in which participants’ savings are supplemented with matched funds from other public or private sources. Children’s savings accounts (CSAs) are similar to IDAs, except the savings account is opened in the name of a child in order to finance his or her education or future career. Poverty reduction through the promotion of asset ownership is grounded in the premise that access to assets, including savings, home ownership, education, and income-generating opportunities, enhances people’s capacities to make choices and pursue their life goals (Sherraden 1991, 2005; Sen 1999; Lerman and McKernan 2008; Ssewamala et al. 2010).

This article estimates the effect of a subsidized, matched, National Institute of Mental Health-funded CSA program called Suubi–Maka (or “Hope for Families” in the commonly spoken language, Luganda) on the saving performance of poor families in Uganda who are caring for orphaned and vulnerable children between the ages of 10 and 15. We use a randomized experimental design to explore whether families participating in the subsidized CSA program report more savings than their counterparts not participating in the program, examine the extent to which families who participate in the subsidized CSA program report using financial institutions, specifically banks, compared with families who did not have a CSA, and consider whether families in the subsidized CSA program bring new money into the CSAs or whether they reshuffle current household assets.

Background and Theoretical Framework

Institutional Theory of Saving

According to the institutional theory of saving, low monetary savings by the poor can be partially explained by the limited institutional opportunities available to them. People save not only because of their individual choices, but also because of institutional frameworks that encourage saving behavior, such as access (e.g., proximity of savings programs and products, use of electronic deposits), incentives (e.g., matched deposits, earnings on savings), information (e.g., educational programs to increase economic literacy), facilitation (e.g., use of automatic deposits, assistance from program staff), expectations (e.g., saving goals), restrictions (e.g., use of assets for specified/stipulated purposes only, such as home ownership, educational investment, or microenterprise development), and security of investments (Sherraden 1991; Schreiner and Sherraden 2007; McKernan and Sherraden 2008). Institutional theorists assert that low savings and low rates of asset accumulation among poor populations are partially explained by individuals’ lack of access to institutional constructs that encourage savings.

Children’s Savings Accounts (CSAs)

The institutional theory of saving guides subsidized CSAs. A CSA account is opened at a bank in a child’s name, which allows for separation between parental assets and children’s assets (Elliott, Kim, et al. 2010; Meyer, Masa, and Zimmerman 2010; Elliott, Choi, et al. 2011). Subsidized CSAs tend to incentivize savings through a match and/or account opening amount to the child and his or her parents, provide educational programming in the form of financial education and/or peer support, impose withdrawal restrictions, have mentorship programs and peer support intended to facilitate savings, and bring financial institutions closer to the participants, specifically when the account is opened. Accounts are usually (although not always) opened in schools and facilitated by a bank representative. Some programs may include in-school banking as part of the access.

In recent years, policymakers and social welfare programs have increasingly been calling for the promotion of CSAs as a means to empower the poor, including poor youth, both in western industrialized countries (the UK Child Trust Fund, for example) and in the poor developing countries of Africa, Latin America, and south Asia (for example, the YouthSave Consortium led by Save the Children, and funded by the MasterCard Foundation). Other examples include the Safe and Smart Savings program for vulnerable adolescent girls in Kenya and Uganda, supported by the Population Council; Early Start Savings Accounts, provided by PostBank in Uganda; Zawadi Accounts from the Akiba Commercial Bank in Tanzania; and Butterflies Children’s Development Bank, for the street and working children in South Asia. However, CSAs offering matching as a form of incentive or subsidy for opening and operating the account are not very common in developing countries. Specifically, as of the time of writing, we have only been able to identify three developing countries offering subsidized CSAs for the children and adolescents in our focus age bracket of 10–15 years: South Africa’s Fundisa accounts; Uganda’s SEED-Uganda, Suubi-Uganda, Suubi-Maka, and Bridges accounts; and Sri Lanka’s Singithi Kirikatiyo accounts. The extent to which subsidized CSAs can facilitate asset accumulation in developing countries may thus hold important implications for policymaking in a wide range of countries.

Participants in the treatment group of our study were provided with Suubi–Maka CSA accounts. These accounts provided them with a 2:1 match on money saved into the account, up to Uganda Shillings (UGX) 20,000 (approximately US$10) per month. The accounts were administered through recognized and regulated local banks. Participants were also provided with financial management training. The program encourages in-school banking as much as possible.

Literature Review

Numerous studies of subsidized matched savings accounts demonstrate that incentivizing savings helps the poor to save (Sherraden et al. 2003; Ssewamala and Sherraden 2004; Grinstein-Weiss et al. 2006; Han et al. 2009). Data from the American Dream Demonstration (ADD), the first and most extensive study of IDAs for 2,377 low-income individuals from across the United States, indicate that participants saved an average of $16.60 per month (Sherraden 2005), making an average of $2,586 in matched withdrawals (total deposits plus matching) between 1997 and 2002 (Han and Sherraden 2009). More recently, Michael Sherraden and colleagues examined the effect of subsidized CSAs on savings and asset accumulation for children enrolled in the SEED–Oklahoma study, a large-scale study of CSAs with 2,670 randomly selected newborn children, and find that more people in the treatment group (16 percent) held a participant-owned account, compared to the control group (1 percent). Results also show that participants in the treatment group saved more than participants in the control group ($47 vs. $13) (Nam et al. 2013). In southern Uganda, participants in the Suubi Program (2005–2008), which is an incentivized CSA program, saved an average of US$75.96 per year in net deposits. After the program matched their savings on a 2:1 rate, participants accumulated an average of US$228 per year, which is a substantial sum in a poor country like Uganda, and can cover the cost of anywhere between 1.5 to 2 years of post-primary education in the semi-urban areas where Suubi Program participants live (Ssewamala et al. 2009).

Kristin Richards and Bruce Thyer (2011) conducted a systematic review of peer-reviewed studies examining savings outcomes of IDA participants and conclude that most existing studies examining savings in these accounts are based on a single-group pretest-posttest research design (examples include Sherraden et al. 2003; Zhan 2003; Ssewamala, Lombe, and Curley 2006; Lombe and Ssewamala 2007; Christy-McMullin, Shobe, and Wills 2008; Curley et al. 2009; Grinstein-Weiss et al. 2010). Even studies with treatment and control groups, for example SEED–Uganda and Suubi–Uganda, draw conclusions on the basis of within-treatment group pretest-posttest differences in savings because they rely on administrative data to measure financial savings, which is only available for the treatment group. Without a between-group financial savings comparison (treatment vs. control group), it is difficult to draw firm conclusions on whether a subsidized IDA facilitates savings. Even when looking beyond research in developing countries, Yunju Nam and colleagues (Nam et al. 2013) conclude from a plethora of earlier research that there is little consensus on whether asset-building programs increase savings among the populations targeted by the programs.

Additionally, it is not clear whether an IDA changes participants‧ savings behaviors in ways that increase family assets, or whether participants deposit money by reshuffling their assets. For example, Gregory Mills and colleagues (2008) conclude from their analysis of data from the experimental design component (with a control and treatment group) included in the ADD in Tulsa, Oklahoma that IDA-holders shift their financial assets from other accounts into the IDA. In another study, Chang-Keun Han and colleagues (2009) examine the effect of IDAs on five different types of assets: liquid assets (checking accounts, savings accounts, money market accounts, and certificates of deposit), other financial assets (sum of saving bonds, educational saving accounts, stocks, bonds, mutual funds, Christmas club accounts, and vacation accounts), total financial assets (adding value of retirement savings), real assets (values of real estate), and total assets (the sum of all the previous four). Their findings reveal that treatment group members experienced a decrease in liquid and other financial assets, which may suggest possible reshuffling of liquid and financial assets into IDA accounts. However, based on analyses of IDAs in the ADD that examine the relationship between IDAs and household assets by looking at total household liability (sum of debts and loans), household financial assets (combination of liquid assets and 13 other types of assets), and net worth (sum of financial assets and real assets net of total liability), Jin Huang (2010) finds no evidence that IDA participants reshuffled their assets.

One gap in the existing literature is that prior studies have focused primarily on low-income families in the United States. Little is known about the incentivized or subsidized matched savings programs that are also emerging in sub-Saharan Africa. Previous evaluations of CSAs in Uganda (Ssewamala, Han, and Neilands 2009; Ssewamala and Ismayilova 2009; Ssewamala et al. 2010) suggest that participants in the treatment group were able to save, but these studies did not compare savings in the treatment group with those in the control group and, therefore, did not include an experimental design to estimate the causal effect of the intervention on families’ saving behavior.

One particular way in which the successful use of IDAs and CSAs may differ between the United States and Africa is by providing access to formal financial institutions. Financial inclusion, or providing low-income individuals with opportunities to save in formal financial institutions, is one of the important policy claims behind IDAs. However, in sub-Saharan Africa, 76 percent of adults do not have an account in a formal financial institution whereas only 8.2 percent of adults in the United States are among the unbanked (FDIC 2011; Demirgüç-Kunt and Klapper 2012). In sub-Saharan Africa, lack of money to open an account is one of the most frequently cited reasons for not having a formal bank account. Specifically in Uganda, only 20 percent of adults have accounts in formal financial institutions, and 54 percent of non account-holders say that fees and costs of opening and maintaining an account are reasons for not having one (Demirgüç-Kunt and Klapper 2012, 2013). The extent to which participation in subsidized matched savings accounts programs increases the use of formal financial institutions for saving remains unclear.

Furthermore, most studies on the accumulation of financial assets focus on children and adolescents’ economic socialization, mainly by their parents and peers (Hibbert et al. 2004; Webley and Nyhus 2006, 2013; Kim, Lee, and Tomiuk 2009; Gudmunson and Danes 2011; Otto 2013). Very few studies specifically address asset accumulation by adolescents themselves.

Our article’s distinctive focus is on estimating whether saving behavior changes specifically as a result of families’ participation in subsidized CSAs for children and young adolescents ages 10–15. We want to emphasize the importance of distinguishing the matched CSAs from other forms of youth savings products that do not offer incentives such as seeding and matching deposits, and that do not restrict withdrawals based on the proposed use of the money. Furthermore, we estimate the effect of CSAs not only on the saving behavior of children, but also on the saving behavior of their caregivers.

Method

Data

The Suubi–Maka study (2008–2012) used a cluster-randomized experimental design in which 10 public primary schools in southwestern Uganda’s Rakai and Masaka Districts were randomly assigned to either a control condition or a treatment condition. Over the years, the Rakai and Masaka Districts have been heavily affected by wars and disease, including HIV/AIDS. As a result, there are high numbers of poor children and adolescents who are orphans (which we define as having at least one deceased biological parent). In Rakai alone, two in five children are orphaned.

To select the schools for the study, we screened 42 primary schools and selected 10 that matched our inclusion criteria. The 10 schools included in the study had similar socio-economic characteristics (detailed below). They were all public (government-supported) and located in the Rakai and Masaka districts. In addition, the schools had similarly sized populations of children (an average of 600 children in primary grades 1 to 7) and were at a comparable level of academic performance based on results of the Primary Leaving Examination (PLE), a nationally administered examination taken by students in the seventh grade (the last grade in Uganda’s primary schooling system). To identify the comparable level of PLE scores, we examined the distribution of PLE results for all 42 schools for the last three years and kept schools within one standard deviation from the mean. Furthermore, all schools had comparable infrastructure, resource availability, and demographic characteristics in terms of gender. All schools had to be mixed schools, enrolling both girls and boys. We excluded all boarding schools. After matching the schools on those characteristics, 10 schools remained for inclusion in the study. Following a baseline survey, each of the 10 schools was randomly assigned to either a control condition (n = 5 schools) or a treatment condition (n = 5 schools). The cluster randomization at the school level was intended to avoid contamination.

Following school selection, we selected children from each school for inclusion in the study. In order to be eligible for the study, a child needed to be orphaned (having lost one or both parents), enrolled in the last two years of primary school, and living with a family. Following informational meetings on the study conducted by project staff and the schools’ administrators during the beginning of the school academic term registration and parent teacher association meetings, parents or caregivers who expressed interest in being involved in the study were invited to participate. Each parent or caregiver had to give consent to participate, and each enrolled child had to give assent. Children were asked to assent separately from their caregivers or parents in order to avoid their being pressured to participate or not to participate. During the consent process, three families refused to participate, citing religious reasons.

Children in the control group (n = 167) received the usual form of care for orphaned students in the study region, including counseling and the provision of school lunch, school uniforms, and scholastic materials such as textbooks and notebooks. Children in the treatment group (n = 179) received the usual form of care outlined above, plus the opportunity to open a subsidized matched CSA at a nationally registered and regulated formal financial institution/bank. It is important to note that these accounts were in the children’s names, with a caregiver’s name listed on the account to address a law that prohibits children below age 18 from independently managing an account. The children received bank statements every month that informed them of their savings, and they received financial management training (described below). Each month, monetary savings up to Uganda shillings (UGX) 20,000 (approximately US$10) deposited in the CSA was matched by the Suubi–Maka Program at a 2:1 ratio. In other words, if a CSA participant deposited an equivalent of US$10 per month, that participant would receive an additional US$20 in their account every month (2:1 match rate), giving the participant a total of US$30 in his or her CSA at the end of the month, and US$360 at the end of a 12-month period. In addition to the participating child and family members, anyone could deposit into the child’s account. Note also that participants could withdraw their own money from the account at any time, but not the matched amount. The matched savings were restricted and could only be withdrawn to pay for school or education-related expenses including fees and tuition, or to start a small family business. In order to ensure that children used their matched savings per the stipulations of the study, the money was transferred directly to the school/educational institution. For example, in the case of school fees or other school-related expenses such as registration fees or examination fees, the check or voucher for the match was paid directly to the participant’s school. Children in the treatment group, together with their guardians, were also offered the opportunity to receive financial training, which consisted of three full-day sessions on financial management, covering basic financial literacy and management, and microenterprise development training. These sessions were conducted by financial institutions holding CSAs for the study (Kakuuto Microfinance, Centenary Rural Development Bank, and DFCU Bank).

Statistical Power Calculations

The study’s a priori (i.e., pre-data collection) power analyses were conducted for random coefficient models using the RMASS2 computer program in order to estimate the minimum detectable effect size for the group-by-time interaction in a two-group repeated measures design. Using a prior study in the same area and among a similar population as a base (the Suubi study 2005–2008; see Ssewamala and Ismayilova 2009), we conducted analyses for a range of continuous and categorical variables of interest (e.g. depression level, sexual risk-taking behavior, educational attainment). Power analysis inputs included the proposed number of time points (T = 3), proposed number of research participants (n = 300), anticipated attrition rate over a 24-month period (10 percent), and variance-covariance structure of the repeated measures. To address clustering not only within individuals but also within schools, the effective sample size was adjusted by the design effect. The DEFF (design effect) represents the factor by which the effective sample size (ESS) is reduced due to clustering for hypothesis testing purposes. A DEFF of 1.0 represents no clustering effect, whereas values of DEFF exceeding 1.0 indicate the presence of cluster variation that reduces the power of statistical tests. Effect sizes found through these analyses fell between 0.20 and 0.50, which constitute small and medium effect sizes, respectively (Cohen 1987). Consequently, these power analyses showed that with n = 300 we would have sufficient statistical power to obtain reliable point estimates, confidence intervals, and effect sizes for our hypotheses testing. The Suubi–Maka study pursued the goal of conducting a formative study to examine the feasibility of family-level economic empowerment intervention focused on savings. Therefore, given the formative nature of the study, no previous information on intraclass correlation (ICC) was available to assess autocorrelation of children within schools. As we indicate in the data section of our article, clusters (n = 10 schools) were selected based on the inclusion criteria detailed above. Additionally, the number of clusters was based on the study overall aim of ascertaining feasibility and acceptability of a family-level economic empowerment intervention focused on savings, and ascertaining short-term outcomes from such an intervention.

We collected three waves of data on 346 participating children and their caregivers over the course of 24 months. On average, we selected about 35 dyads (that is, children and their caregivers) in each school. We collected baseline data prior to random assignment and collected wave 2 and wave 3 data at 12 and 24 months, respectively, following intervention initiation. Five schools were in the control group (n = 167 dyads) and 5 schools were in the treatment group (n = 179 dyads). The participating children were in their last two years of primary school prior to transitioning to secondary school. The average age of study participants was 13 years. No participant (child or caregiver) refused to participate in follow-up interviews. But, we lost some participants at the follow-up because they relocated from the study’s geographical area and did not tell the study’s implementing partners (Masaka Diocese) where they moved. In those instances, we were unable to locate the participants at follow-up. Within that context, the attrition rate (referring to children who we were not able to trace) at the 24-month follow-up was 7.3 percent (n = 13) for the treatment group and 9.6 percent (n = 16) for the control group.

Hypotheses

In accordance with our research questions on how the subsidized CSA program may increase savings, foster the use of financial institutions, and create new wealth, we propose three hypotheses. First, we hypothesize that, among the Suubi–Maka participants, families in the treatment group (those with CSAs) will report more savings than families in the control group (those without CSAs) over the course of the intervention. Second, we hypothesize that, over the course of the study, more families in the treatment group will report keeping their money in financial institutions such as banks, microfinance institutions, and credit unions, than those in the control group. Third, we hypothesize that, compared to families in the control group, families in the treatment group will not experience a reduction in the types of household assets they own over the course of the study, which may suggest that families in the treatment group did not have to sell existing household assets in order to contribute funds to their CSA.

Measures

We address all three of our hypotheses by measuring the differential change in the outcome variable over the course of intervention, at 12 months and at 24 months, for both the treatment and the control group. Our data come from three sources: a 90-minute individual instrument administered with children, a 90-minute individual instrument administered with parents or caregivers, and administrative data on savings obtained directly from the financial institutions where participants in the treatment group opened their CSAs.

To test whether families in the treatment group report more savings than families in the control group, we use four measures of savings. Self-reported saving by the child and self-reported saving by the guardian are both measured by dichotomous (yes/no) variables obtained by using the child and guardian’s answer to the question, “Do you currently have any money saved anywhere?” The self-reported amount saved by the child and self-reported amount saved by the guardian are both measured by continuous variables indicating the amount of money currently saved by a participant. These measures are obtained by using the child and the guardian’s response to the question, “How much money do you currently have saved?” It is important to note that this is a self-reported measure of money currently saved. Information on the self-reported amount saved is available only for participants who responded yes to the preceding question, “Do you currently have any money saved anywhere?” Therefore, these measures contain no zero values. Both measures are non-normally distributed (amount reported by child: Skewness = 5.41, Kurtosis = 35.6; amount reported by guardian: Skewness = 5.6, Kurtosis = 38), and, therefore, were transformed by a natural log.

To verify the self-reported data on savings, we used administrative data from financial institutions and available only for participants in the treatment group (for details, see the Analyses section), specifically looking into two measures: the amount saved in the CSA and whether the participant had an active CSA. The amount saved in the CSA is the sum of money that remains on the account after all the unmatched withdrawals are subtracted from all the deposits—in other words, this is the sum of money eventually matched by the program. Even if the unmatched withdrawals were used in the way intended by the program, we would not be able to verify this. Given our definition, the measure of amount saved in the CSA was obtained by subtracting total unmatched withdrawals from total deposits made in CSAs during the study period.

The measure of whether the participant had an active CSA is a binary variable coded as 1 = yes for all the participants who opened an account, activated it, and deposited at least once; and coded as 0 = no for all the participants who either never opened an account, opened an account but never activated it, or opened and activated an account but never deposited any money.

To test the extent to which families in the treatment group report using financial institutions, specifically banks, compared with families in the control group, we measure saving in financial institutions, dichotomized as yes/no, by recoding the guardian’s response to the following question: “Do you have money saved in any of the following places: bank (yes/no), credit union (yes/no), with your friend (yes/no), any other place (yes/no).” When the guardian reported saving money in a bank and/or credit union, the response category for our dichotomous variable was recoded as yes. For all other cases, the response category for our dichotomous variable was recoded as no. Although the accounts were open in the name of a child, the guardian’s name was also on the account in the capacity of a trusted adult (because children in the Suubi-Maka study were all below the age of 18). Moreover, family members were allowed to deposit money in these CSAs on behalf of the children. Consequently, as we learned during the booster session meetings with study participants and caregivers, adults considered these accounts as the family saving on behalf of the participating child. It is therefore more than likely that, when caregivers were answering the question as to whether they had any money saved in the bank, they considered whatever was saved in the CSA as part of the family savings to be reported.

To test whether families in the treatment group brought new money into the CSAs or whether they reshuffled existing household assets, we constructed a composite score of household assets. The score contains 16 dichotomous items indicating the household’s ownership of land, gardens and livestock, and means of transportation, such as, “family owns house,” “family owns land,” “family owns motorcycle,” “family owns banana garden,” “family owns cow(s),” and so on. The score ranges from 0 (household has no assets) to 16 (household has all types of assets). Only the guardians were asked the 16 questions on household asset ownership.

The main predictor variable in all three questions is the group-by-time interaction, which allows us to examine the marginal effect of the intervention on the outcome variable at a 12-month follow-up and a 24-month follow-up. We include the child and guardian’s ages, child and guardian’s gender, child’s orphanhood status (dichotomized as “single orphan” or “double orphan”), number of people in the household, and number of earners in the household (dichotomized as “single-earner family” or “double-earner family”) as control variables.

Analyses

Given the repeated measures nature of the data, measurements are likely to be correlated. To account for the clustered repeated-measures nature of data collected on the same children and their guardians over the three waves, we ran analyses using generalized estimating equation (GEE) models with schools treated as nested fixed effects and robust standard errors clustering on individual observations (xtgee command in Stata 12). Similar to multilevel models, GEE models account for within-subject correlations among responses over time and allow for time-varying covariates. However, unlike the subject-specific multilevel models, GEE models estimate population-averaged treatment effects (instead of a subject-specific treatment effect).

Out of the total sample (n = 346 dyads), 22 families had 2 adolescents enrolled in the study, and 1 family had 3 adolescents. The rest of the families (n = 323) each had only one adolescent in the study. Each family was represented by one guardian, identified as the primary caregiver for the child. This is the person who was interviewed throughout the study, who participated in all the financial management trainings, and who was the signatory on the subsidized CSA (in case of children enrolled in the treatment group). In order to decide whether estimates should be clustered by family, we conducted two types of analyses. First, we calculated intraclass correlations to assess homogeneity of model outcomes within a cluster (i.e., families). The results show small values of within-family variance relative to the total variance. Second, models nested in families were compared to models not nested in families using Likelihood Ratio tests. The results show no significant differences between models (Gelman and Hill 2007; West, Welch, and Galecki 2007). Based on these results, we concluded that analyses do not need to adjust standard errors to account for the fact that some families had multiple children in the study.

Analyses of baseline data also address clustering of individuals within schools. In table 1, which describes the sample at baseline, we report adjusted Wald F-statistics (design-based F) to examine individual-level variations while accounting for potential correlation between same-school observations. For outcome variables with values not equivalent between the treatment and control groups at a baseline, we conducted double-robust estimation analyses using the user-written Stata command -dr- (Emsley et al. 2008). These analyses allow for an understanding of whether the post-intervention values of outcome variables would have differed if baseline values were equivalent.

Table 1.

Description of the sample at baseline (n=346)

Predictor Measures	Percent or Mean [95% Confidence Interval]
	Total (n=346)	Treatment (n=179)	Control (n=167)	Design- based F
Child’s age (Mean; range: 10–17)	13 [13; 14]	13 [13; 14]	13 [13; 14]	0.00
Female child (%)	65 [57; 72]	65 [53; 76]	65 [55; 74]	0.001
Orphanhood status (%)
Single orphan	71 [64; 78]	77 [72; 81]	65 [56; 73]	8.77*
Double orphan	29 [23; 37]	23 [19; 28]	35 [27; 44]
Female guardian (%)	80 [70; 87]	78 [65; 87]	81 [65; 91]	0.2
Type of guardian (%)				3.27
Parents	35 [28; 43]	40 [31; 50]	30 [22; 40]
Grandparent	29 [20; 39]	21 [12; 36]	37 [29; 45]
Other relatives	36 [31; 42]	39 [32; 46]	34 [26; 42]
Guardian’s age (Mean; range: 18–87)	46 [43; 49]	44 [39; 49]	48 [46; 49]	2.59
Number of people in the household (Mean; range: 1–12)	6 [6; 7]	7 [6; 7]	6 [6; 7]	0.87
Single-earner family (%)	64 [59; 69]	64 [59; 70]	64 [56; 71]	0.02
Outcome Measures
Does child report having money saved anywhere? (%, YES)	19 [13; 28]	20 [14; 28]	18 [8; 35]	0.11
(a) How much money does the child report having saved? (in Uganda shillings) (Mean; range: 200 – 59,874)	6,002 [4,447; 8,350]	5,432 [3,641; 8,955]	6,634 [4,915; 9,897]	0.53
Does guardian report having money saved anywhere? (%, YES)	34 [24; 46]	39 [23; 58]	28 [23; 34]	1.93
(b) How much money does the guardian report having saved? (in Uganda shillings) (Mean; range: 6,000- 3,269,017)	89,322 [120,572; 6,003]	120,571 [98,716; 147,267]	54,176 [44,356; 73,130]	35***
(b) Does the guardian report saving in financial institution? (%, YES)	82 [72; 89]	86 [79; 91]	76 [58; 88]	2.26
Household assets (range: 0–16)	7 [6, 8]	6 [5,7]	8 [7, 8]	6.67*

^*p≤0.05,

^**p≤0.01,

^***p≤0.001

^(a)n=66

^(b)n=117

Boldface type indicates statistically significant F-tests

Information on the self-reported amount saved and on saving in financial institutions is available only for participants who reported having money saved. Eliminating all other cases (i.e., participants who reported not having any money saved anywhere) may jeopardize the benefits of random assignment and thus, undermine interpretation. Therefore, our estimation strategy for the self-reported amount saved and for the saving in financial institutions employed the Heckman method of selection model and inverse Mill ratio accounting for the limited dependent variable (Wooldridge 2012). The inverse Mills ratio was calculated and incorporated into regressions, as follows. First, we estimated the probability of participants reporting having money saved by running xtgee logit on the outcome measure of self-reported saving. Second, we calculated the inverse Mills ratio, lambda, to account for the fact that only part of the sample has values for the self-reported amount saved and for saving in financial institutions. Mills’s lambda is a ratio between the standard normal probability density function and standard normal cumulative distribution function. Third, we included lambda as an additional predictor into an xtgee regression run on the outcome measure of self-reported amount saved and saving in financial institutions. A statistically significant coefficient on lambda signified selection bias due to observing the self-reported amount saved and saving in financial institutions only for part of the sample. Including the inverse Mills ratio into the regression model thus helps to attenuate the influence of selection bias on parameter estimates.

To take advantage of the randomized control trial (RCT) design of the Suubi–Maka study, we use self-reported data on savings because they are available for both the treatment and control group at all three waves. Self-reported data on savings rely on children’s and guardian’s reports on whether they have money saved anywhere and how much money they have saved. Self-reporting bias may lead to inaccuracies in estimating the effect size. To address this risk and to verify self-reported data on savings, we use administrative data obtained directly from banks and available only for participants in the treatment group. Specifically, we ran bivariate analyses, for the treatment group only, to explore the correlation between the self-reported amount of money saved at wave 3 (self-reported data) and the amount saved in CSAs (bank administrative data). Additionally, to provide an estimate of the validity of the self-report data to capture savings behavior, we estimate the association between self-reported saving (yes/no) and having an active CSA yes/no, measured from the bank administrative data.

Results

Description of the Sample

The average age of the children in the study was 13 years. All of the children are orphans, having lost one parent (71 percent) or both parents (29 percent). Sixty-five percent of all the children in the study are girls. In 80 percent of cases, the children were cared for by a female guardian. The age of guardians participating in the study was 46 years on average, and ranged from 18 to 87. In 65 percent of cases, children were taken care of either by grandparents (29 percent) or by other relatives (36 percent), such aunts and uncles. On average, 64 percent of children lived in single-earner families where child’s guardian was the only source of financial support for the family. Participants lived in households with relatively large families; the average number of people in the household was six, which is two members above the reported average of four for the Rakai District of Uganda (Uganda Bureau of Statistics 2010). For details, see table 1.

As indicated in table 1, at baseline, 19 percent (n = 66) of the children in the study and 34 percent (n = 117) of the guardians reported having saved money. Children saving at baseline reported saving an average of UGX 6,002 (equivalent to US$3.30, given the exchange rate at the time of US$1 = UGX 1,800). At baseline, there was no significant difference in self-reported savings by children between the treatment and control groups.

Of the guardians who reported saving at baseline (n = 117), 82 percent reported keeping their money in financial institutions (banks or credit unions), and 18 percent reported keeping money elsewhere (e.g., with a relative, at home). They reported an average savings of UGX 89,322 (equivalent to US$50). The amounts were larger for guardians in the treatment group than for those in the control group (design-based F = 35, p <0.001).

Baseline results indicate that families had an average of 7 out of 16 assets included in the household asset composite score. Participants in the control group reported higher assets (8 out 16) compared with participants in the treatment group (6 out of 16 assets, design-based F = 6.67, p < 0.05).

Generalized Estimating Equation Models: Children’s Savings

To respond to our first research question of whether families participating in the subsidized CSA program report more savings than their counterparts, we examined savings outcomes by children as well as their caregivers. Table 2 presents results of generalized estimating equation regression for savings outcomes reported by children. Model 1 presents the self-reported saving (yes/no) and model 2 presents the self-reported logged amount of money saved.

Table 2.

Generalized estimating equation regressions for children’s savings across three waves

Predictor Variables	Model 1 Do you have any money saved anywhere?		Model 2 How much money do you have saved? (LOG)
	Odds Ratio [95% Confidence Interval]		Beta-coefficient [95% Confidence Interval]
Group by Time Interaction(a)	χ2(2) =9.19**		χ2(2) = 3.01
Marginal intervention effect: Wave 2	2.72**	[1.40; 5.29]	0.68	[−0.63; 1.99]
Marginal intervention effect: Wave 3	2.26*	[1.12; 4.57]	0.75	[−0.20; 1.69]
Child’s gender (reference: male)	0.45***	[0.31; 0.66]	−0.96*	[−1.92; −0.001]
Child’s age	1.03	[0.89; 1.19]	0.23***	[0.11; 0.36]
Child’s orphanhood status (reference: double orphan)	1.30	[0.86; 1.97]	−0.04	[−0.50; 0.42]
Guardian’s gender (reference: male)	0.76	[0.50; 1.17]	−0.31	[−0.75; 0.13]
Guardians-grandparents (reference: parents)	1.21	[0.76; 1.93]	0.28	[−0.19; 0.75]
Guardians-other relatives (reference: parents)	1.11	[0.75; 1.65]	−0.12	[−0.48; 0.23]
Guardian’s age	1.01	[0.99; 1.02]	0.00	[−0.01; 0.02]
Number of people in the household	0.92*	[0.85; 0.99]	−0.04	[−0.17; 0.09]
Number of earners (reference: double-earner)	0.80	[0.59; 1.11]	−0.04	[−0.48; 0.39]
School Fixed effects(b)	χ2(8) =22.84**		χ2(8) =15.56*
School 1	0.65	[0.28; 1.51]	−0.04	[−1.06; 0.98]
School 2	0.72	[0.40; 1.29]	0.15	[−0.42; 0.71]
School 3	0.94	[0.50; 1.77]	0.21	[−0.36; 0.78]
School 4	0.87	[0.39; 1.92]	0.25	[−0.48; 0.98]
School 6	0.88	[0.33; 2.35]	−0.10	[−0.76; 0.57]
School 7	3.34*	[1.25; 8.94]	1.21	[−0.42; 2.83]
School 8	1.49	[0.48; 4.62]	−0.40	[−1.18; 0.37]
School 9	0.59	[0.23; 1.51]	−0.45	[−1.57; 0.66]
Mills lambda			0.75	[−1.00; 2.51]
Constant (mean intercept)	0.23	[0.02; 2.72]	4.79*	[1.03; 8.56]
Number of observations	933		327
Wald Test for the overall model	χ2(22) = 165.45***		χ2(23) = 133.39***

^*p≤0.05,

^**p≤0.01,

^***p≤0.001

Boldface type indicates statistically significant results

^(a)Results of 2 DF Wald Test for joint interaction effect

^(b)Results of 8 DF Wald Test for school effect

As illustrated in model 1 of table 2, the odds of having any money saved anywhere are higher for the children in the treatment group than for the children in the control group at both 12 months and 24 months after baseline, which partially supports our hypotheses that families in schools randomly assigned to the treatment group saved more than those in schools assigned to the control group. In other words, compared to children in the control group, children in the treatment group were more likely to report having money saved both at wave 2 (odds ratio = 2.72; 95 percent Confidence Interval [CI] = 1.4, 5.29, p < 0.01) and at wave 3 (odds ratio = 2.26; 95 percent CI = 1.12, 4.57, p < 0.05). Girls were less likely to report having money saved, compared to boys (odds ratio = 0.45; 95 percent CI = 0.31, 0.66, p < 0.001). Also, children in larger families were less likely to report having money saved than children in smaller families (odds ratio = 0.92; 95 percent CI = 0.85, 0.99, p < 0.05).

Model 2 shows that, contrary to our hypotheses, there is no significant effect of intervention on the self-reported amount saved. In other words, we do not have statistical evidence demonstrating that children in the treatment group reported saving larger amounts than children in the control group. Furthermore, model 2 indicates that older children reported saving more money than younger children (Beta-coefficient [B] = 0.23, 95 percent CI = 0.11, 0.36, p < 0.001). Moreover, girls reported saving less than boys (B = −0.96, 95 percent CI = −1.92, −0.001, p < 0.05).

Generalized Estimating Equation Models: Guardian’s Savings

Table 3 provides results of generalized estimating equation regression for savings outcomes reported by guardians, which are relevant to addressing our hypotheses that families in schools assigned to the treatment condition saved more and were more likely to report using a formal financial institution than families in schools assigned to the control group. Model 1 of table 3 examines self-reported saving (yes/no), model 2 examines the self-reported logged amount of money saved, and model 3 examines whether the guardian reported saving money in financial institutions (yes/no).

Table 3.

Generalized estimating equation regressions for guardians’ savings across three waves

Predictor Variables	Model 1 Do you have any money saved anywhere?		Model 2 How much money do you have saved? (LOG)		Model 3 Used formal Financial institution?
	Odds Ratio [95% Confidence Interval]		Beta-coefficient [95% Confidence Interval]		Odds Ratio [95% Confidence Interval]
Group by Time Interaction(a)	χ2(2) =6.47*		χ2(2) =9.8**		χ2(2)=0.34
Marginal intervention effect: Wave 2	1.84*	[1.06; 3.2]	−0.49	[−1.02; 0.04]	1.05	[0.27; 4.14]
Marginal intervention effect: Wave 3	1.95*	[1.09; 3.51]	0.07	[−0.42; 0.56]	1.36	[0.40; 4.61]
Child’s gender (reference: male)	0.91	[0.60; 1.38]	−0.02	[−0.35; 0.31]	1.01	[0.47; 2.14]
Child’s age	1.01	[0.85; 1.19]	0.08	[−0.06; 0.22]	1.21	[0.87; 1.68]
Child’s orphanhood status (reference: double orphan)	1.19	[0.80; 1.78]	0.10	[−0.23; 0.44]	0.99	[0.45; 2.16]
Guardian’s gender (reference: male)	0.31***	[0.2; 0.5]	−0.61*	[−1.13; −0.09]	0.98	[0.23; 4.08]
Guardians-grandparents (reference: parents)	0.94	[0.56; 1.59]	0.11	[−0.30; 0.52]	0.77	[0.27; 2.19]
Guardians-other relatives (reference: parents)	1.04	[0.68; 1.60]	0.38*	[0.03; 0.72]	1.52	[0.73; 3.18]
Guardian’s age	0.97***	[0.96; 0.99]	−0.01	[−0.02; 0.01]	1.07*	[1.01; 1.13]
Number of people in the household	1.09*	[1.02; 1.17]	−0.00	[−0.07; 0.06]	1.14	[0.96; 1.36]
Number of earners (reference: double-earner)	0.80	[0.59; 1.09]	−0.08	[−0.30; 0.14]	0.99	[0.54; 1.82]
School Fixed effects(b)	χ2(8) = 9.96		χ2(8) = 12.38		χ2(8) =5.19
School 1	0.74	[0.33; 1.68]	−0.51	[−1.33; 0.32]	0.79	[0.13; 4.66]
School 2	1.23	[0.62; 2.43]	0.35	[−0.19; 0.90]	0.93	[0.28; 3.02]
School 3	2.28*	[1.11; 4.69]	0.75**	[0.19; 1.31]	1.13	[0.28; 4.55]
School 4	1.70	[0.73; 3.95]	0.42	[−0.33; 1.16]	3.49	[0.27; 44.98]
School 6	1.42	[0.47; 4.33]	−0.12	[−1.20; 0.96]	0.73	[0.09; 5.98]
School 7	1.69	[0.54; 5.31]	−0.15	[−1.24; 0.94]	2.01	[0.17; 23.69]
School 8	1.78	[0.46; 6.91]	0.02	[−1.02; 1.06]	3.70	[0.26; 53.67]
School 9	1.74	[0.59; 5.15]	−0.40	[−1.39; 0.59]	0.77	[0.08; 7.77]
Mills lambda			0.15	[−0.57; 0.88]	0.51	[0.06; 4.62]
Constant (mean intercept)	1.38	[0.08; 23.29]	10.21***	[7.50; 12.91]	0.01	[0.00; 4.31]
Number of observations	933		402		374
Wald Test for the overall model	χ2(22) = 127.34***		χ2(23) = 120.46***		χ2(23) = 63.23***

^*p≤0.05,

^**p≤0.01,

^***p≤0.001

Boldface type indicates statistically significant results

^(a)Results of 2 DF Wald Test for joint interaction effect

^(b)Results of 8 DF Wald Test for school effect

Model 1 in table 3 indicates that, compared with guardians in the control group, guardians in the treatment group were more likely to report having money saved at both wave 2 (odds ratio = 1.84; 95 percent CI = 1.06, 3.2, p < 0.05) and wave 3 (odds ratio = 1.95; 95 percent CI = 1.09, 3.55, p < 0.05) providing partial support for our hypothesis that families in the treatment group would save more than families in the control group. In addition, model 1 indicates that a positive response to the question, “Do you currently have any money saved anywhere?” was more frequent among guardians living in larger households (odds ratio = 1.09; 95 percent CI = 1.02, 1.17, p < 0.05). Furthermore, female guardians (odds ratio = 0.31; 95 percent CI = 0.2, 0.5, p < 0.001) and older guardians (odds ratio = 0.97; 95 percent CI = 0.96, 0.99, p < 0.001) were less likely to report having saved money, compared to their male counterparts and younger counterparts, respectively.

Model 2 shows no significant intervention effect on the amount saved, which is counter to our hypothesis that families in the treatment group would save more than families in the control group. Additionally, model 2 demonstrates that, on average, female guardians reported saving 45.7 percent less than their male counterparts (B = −0.61, 95 percent CI = −1.13, −0.09, p < 0.05). Guardians who were the children’s relatives reported having more money saved (B = 0.38, 95 percent CI = 0.03, 0.72, p < 0.05) than guardians who were the children’s parents.

Like model 1, model 2 demonstrates significant results for school 3 (Kimaanya). Guardians whose children were enrolled at school 3 were more likely to report having money saved (odds ratio = 2.28; 95 percent CI = 1.11, 4.69, p < 0.05) and they reported saving more (B = 0.75, 95 percent CI = 0.19, 1.31, p < 0.01), compared with guardians whose children were enrolled at other schools.

Contrary to our hypotheses, model 3 indicates no significant effect of intervention on saving in financial institutions. In other words, we found no evidence showing that more families in the treatment group reported saving their money in financial institutions (such as banks, microfinance institutions, and credit unions) than those in the control group. Results of model 3 also show that the odds of guardians saving money in a financial institution were higher for older guardians (odds ratio = 1.07; 95 percent CI = 1.01, 1.13, p < 0.01) compared with their younger counterparts.

Robustness Check of Group Differences in Amount Saved Reported by Guardians

As illustrated earlier in table 1, baseline means of logged reported amounts saved by guardians were different between the treatment and the control groups and this difference was statistically significant. Therefore, it is important to understand whether the post-baseline means of this variable would have differed between the control and the treatment groups if values at baseline were equivalent. To address this question, we ran double robust estimation, the results of which are illustrated in table 4.

Table 4.

Double robust estimates of group differences in amount saved (LOG) reported by guardians

Self-reported amount (LOG) saved by guardian	Mean
	Treatment	Control	Difference (95% CI)	Z	P
Unadjusted
Wave 1 (N=335)	11.7	10.9	0.8 [0.5; 1.1]	5.56	0.000
Wave 2 (N=324)	11.06	11.18	−0.12 [−0.6; 0.4]	−0.44	0.658
Wave 3 (N=316)	11.8	11.07	0.8 [0.3; 1.3]	2.87	0.004
Adjusted
Wave 1 (N=335)	11.46	11.49	−0.03 [−0.9; 1.4]	−0.03	0.979
Wave 2 (N=324)	11.33	11.45	−0.12 [−1.03; 1.5]	−0.04	0.967
Wave 3 (N=316)	12.23	11.82	0.4 [−0.4; 2.5]	0.24	0.808

Estimates were generated using Stata command (dr). Clustering in schools was addressed by using the cluster bootstrap with 5,001 replications. Bias-corrected bootstrap-based confidence intervals are reported for the differences. Adjusted analyses control for effect of control variables listed in table 1 on self-reported amount saved as well as on group membership. Baseline score on self-reported amount saved by guardians was included as a predictor of self-reported amount saved by guardians at Waves 2 and 3. It was also included as a predictor of group membership at all three waves.

Table 4 contains results from unadjusted and adjusted analyses. Unadjusted analyses compare means of logged self-reported amounts saved by guardians between the control and treatment group at each wave. Adjusted analyses do the same, but under the assumption of a pseudo-balance of baseline values for logged self-reported amounts saved by guardians, along with other baseline covariates, as illustrated in table 1. We want to examine changes in the mean difference between the treatment and control groups at each wave when comparing unadjusted and adjusted analyses as well as statistical significance of the mean difference between the treatment and control groups at each wave. Results in table 4 show that mean differences obtained through adjusted analyses at both wave 2 and wave 3 are reduced, compared to mean differences obtained through unadjusted analyses. Furthermore, the mean difference between the treatment and the control groups at wave 3, although statistically significant when obtained through unadjusted analyses (mean difference = 0.8, 95 percent CI = 0.3, 1.3, p < 0.004), is no longer significant when obtained through adjusted analyses (mean difference = 0.4, 95 percent CI = −0.4, 2.5, p < 0.808). This suggests that if baseline savings reported by guardians were equivalent across the two study groups, post-intervention levels of savings reported by guardians would not be different between the groups.

Analyses to Verify the Self-Reported Data on Savings

To verify the self-reported data on savings, bivariate analyses examined associations between the self-reported data and administrative data obtained directly from banks where CSAs were opened for participants. It is important to note that administrative data reflect whether, by the end of the intervention, families opened up CSAs and deposited in those accounts, and the total amount saved per participant remaining in the CSA at the end of the intervention, after accounting for unmatched withdrawals. Therefore, associations are tested between administrative data on savings and the wave 3 values of self-reported data, controlling for school fixed effects (because data is clustered at school level) and accounting for within-subject correlation of self-reported savings information due to the repeated measures nature of data.

Results of bivariate analyses (table 5) show a statistically significant association between having an active CSA (administrative measure) and self-reported saving by the child (χ² = 4.94, p < 0.05). This administrative measure was also significantly associated with self-reported saving by the guardian (χ² = 13.06, p < 0.001) and with the guardian’s likeliness to have savings in a financial institution (χ² = 10.58, p < 0.01). Furthermore, results show a statistically significant association between the amount saved in the CSAs (administrative data) and the amount the guardian reported having saved (χ² = 26.8, p < 0.001). No significant association was found between the amount saved in the CSAs (administrative data) and the amount the child reported having saved. Additionally, bivariate correlation analyses showed that the guardian’s self-reported measure of amount saved shares 48.5 percent of its variability with administrative savings data, while the child’s self-reported measure of amount saved shares 19.6 percent of its variability with administrative savings data. These findings suggest that, despite potential bias due to using self-reported data (for details see Limitations section below), self-reports by guardians are significantly associated (correlated) with administrative data and, therefore, represent reasonably accurate information on savings.

Table 5.

Association between the self-reported and administrative data on savings

SELF-REPORTED DATA	ADMINISTRATIVE DATA
	Having active CSA		Amount saved in CSA
	Coefficient	95% CI	Coefficient	95% CI
Self-reported saving by children
Self-reported saving (Yes/No)	0.78*	[0.09; 1.47]
Wald test (a)	χ2=4.94*
Self-reported amount saved			0.14	[−0.13; −0.41]
Wald test			χ2=0.97
Self-reported saving by guardians
Self-reported saving (Yes/No)	1.19***	[0.54; 1.83]
Wald test	χ2=13.06***
Self-reported amount saved			0.81***	[0.5; 1.11]
Wald test			χ2=26.8***
Guardian reports saving in financial institution (Yes/No)	1.78**	[0.71; 2.85]
Wald test	χ2=10.58**

^***p<0.001,

^**p<0.01,

^*p<0.05.

Boldface type indicates statistically significant results.

^(a)Wald test is for a parameter of interest being equal to zero

Generalized Estimating Equation Models: Household Assets

Table 6 provides results of a generalized estimating equation regression run on household assets, which addresses the hypothesis that the savings program would encourage savings behavior among families assigned to the treatment group without leading them to sell off or reduce their accumulation of other assets, when compared to families in the control group. At wave 3, guardians in the treatment group reported having more types of household assets than guardians in the control group (B = 0.64, 95 percent CI = 0.14, 1.14, p < 0.05), providing some support for this hypothesis.

Table 6.

Generalized estimating equation regressions for household assets

Predictor Variables	Beta-coefficient [95% Confidence Interval]
Group by Time Interaction(a)	χ2(2)=10.98**
Marginal intervention effect: Wave 2	−0.19	[−0.62; 0.25]
Marginal intervention effect: Wave 3	0.64*	[0.14; 1.14]
Child’s gender (reference: male)	−0.32	[−0.82; 0.18]
Child’s age	0.24*	[0.04; 0.43]
Child’s orphanhood status (reference: double orphan)	0.01	[−0.41; 0.43]
Guardian’s gender (reference: male)	−0.73*	[−1.35; −0.11]
Guardians-grandparents (reference: parents)	−0.21	[−0.87; 0.45]
Guardians-other relatives (reference: parents)	0.28	[−0.22; 0.77]
Guardian’s age	0.03**	[0.01; 0.05]
Number of people in the household	0.13**	[0.05; 0.2]
Number of earners (reference: double-earner)	−0.29*	[−0.57; −0.02]
School Fixed effects(b)	χ2(8)=30.79***
School 1	0.65	[−0.47; 1.78]
School 2	0.59	[−0.36; 1.54]
School 3	−1.69**	[−2.76; −0.61]
School 4	−0.28	[−1.67; 1.12]
School 6	0.87	[−0.02; 1.76]
School 7	1.22**	[0.41; 2.04]
School 8	0.76	[−0.39; 1.90]
School 9	0.70	[−0.18; 1.57]
Constant (mean intercept)	2.48	[−0.56; 5.52]
Number of observations	932
Wald Test for the overall model	χ2(22)=150.56***

^*p≤0.05,

^**p≤0.01,

^***p≤0.001

Boldface type indicates statistically significant results

^(a)Results of 2 DF Wald Test for joint interaction effect

^(b)Results of 8 DF Wald Test for school effect

Female guardians reported having fewer types of household assets compared with their male counterparts (B = −0.73, 95 percent CI = −1.35, −0.11, p < 0.05). Also, guardians who were the only earners in the family reported having fewer types of household assets than guardians in double-earner families (B = −0.29, 95 percent CI = −0.57, −0.02, p < 0.05). Furthermore, older guardians reported having more types of household assets than younger guardians (B = 0.03, 95 percent CI = 0.01, 0.05, p < 0.01). Similarly, guardians living in larger households reported having more types of household assets than guardians living in smaller households (B = 0.13, 95 percent CI = 0.05, 0.2, p < 0.01).

Robustness Check of Group Differences in Household Assets

As illustrated in table 1, guardians assigned to the control group reported significantly more types of household assets than guardians in the treatment group. This raises the question of whether the post-baseline differences in household assets between the control and treatment groups would be even greater if these means at baseline were equivalent. To answer this question, we conducted double robust estimation, illustrated in table 7, which produces pseudo-balance on baseline values of household assets.

Table 7.

Double robust estimates of group differences in household assets

Household Assets	Mean
	Treatment	Control	Difference (95% CI)	Z	P
Unadjusted
Wave 1 (N=335)	6.2	7.7	−1.5 [−2.75; −0.7]	−2.9	0.004
Wave 2 (N=324)	6.4	8.1	−1.73 [−3.05; −0.8]	−2.9	0.003
Wave 3 (N=316)	6.5	7.55	−1.04 [−1.9; −0.35]	−2.6	0.008
Adjusted
Wave 1 (N=335)	6.99	6.96	−0.03 [−0.3; 1.3]	0.12	0.908
Wave 2 (N=324)	7.09	7.8	−0.71 [−1.5; −0.3]	−2.37	0.018
Wave 3 (N=316)	7.19	7.07	0.12 [−0.3; 0.8]	0.49	0.623

Estimates were generated using Stata command (dr). Clustering in schools was addressed by using the cluster bootstrap with 5,001 replications. Bias-corrected bootstrap-based confidence intervals are reported for the differences. Adjusted analyses control for effect of control variables listed in table 1 on household assets as well as on group membership. Baseline score on household assets was included as a predictor of household asset scores at Wave 2 and 3. It was also included as a predictor of group membership at all three waves.

Table 7 demonstrates that mean differences between the treatment and the control groups at wave 2 are statistically significant for both unadjusted (mean difference = −1.73, 95 percent CI = −3.05, −0.8, p < 0.003) and adjusted analyses (mean difference = −0.71, 95 percent CI = −1.5, −0.3, p < 0.018). However, mean differences between the treatment and the control groups at wave 3, while statistically significant for unadjusted analyses (mean difference = −1.04, 95 percent CI = −1.9, −0.35, p < 0.008), are no longer statistically significant for adjusted analyses (mean difference = 0.12, 95 percent CI = −0.3, 0.8, p < 0.623). This suggests that if baseline household assets were equivalent between the treatment and control groups, post-intervention levels of household assets at the 24-month follow-up would not be different between the groups. This partially supports our hypotheses that, compared to families in the control group, families in the treatment group will not experience a reduction in measurable household assets over the course of the study.

Discussion

On one hand, CSAs increased the likelihood that poor families would report having saved money. This finding holds for both children and their guardians. A larger proportion of families in schools randomly assigned to the treatment group reported saving throughout the course of the intervention than families in schools randomly assigned to the control group. This finding is in line with previously documented evidence that poor people, even in sub-Saharan African countries with a significant proportion of the population living on less than US$2 a day, can and do save when offered incentives and institutional mechanisms in the form of matched savings and financial education (Ssewamala et al. 2010).

Additional analyses of savings in Suubi–Maka accounts (see appendix 1) show that, out of 179 families enrolled in the treatment group and offered an opportunity to save, 66 percent (n = 118) opened CSAs and deposited money as a result of their participation in the study. In 18.4 percent of cases (n = 33), accounts were never opened. In other cases, accounts were opened but never activated (7.8 percent) or accounts were opened but had no deposits made (7.8 percent). Through our formal booster sessions with study participants and bank officials, we learned that one of the reasons for not opening a bank account was that guardians were not able to submit all the required documentation to the bank (e.g. birth certificates or passport photos). Not opening or activating a CSA did not disqualify a child from being part of the study.

Moreover, we found no statistically significant differences between the characteristics of participants who opened or activated their CSAs and those who did not. In the 118 CSAs opened, families saved an average of US$3.04 per month. Extrapolating from the average number of months accounts were open (18 months), children are estimated to have saved an average of US$54.72, which, after being matched by the program on a 2:1 match rate, comes to US$164.16. Given an average cost of approximately UGX 65,000 per academic term in a typical public secondary school in the study area (rural Rakai), the saved amount is enough to cover approximately five academic terms. Note, though, that analyses comparing the treatment and the control groups in our study show no significant effect of the intervention on the self-reported amount saved either by children or by their caregivers. Furthermore, we find no significant effect of the intervention on participants’ likelihood to save in formal financial institutions (e.g., banks, microfinance institutions, or credit unions). However, as explained earlier, given our research questions, these comparisons were made only among families in the control and treatment groups who reported having saved money, which reduced the sample size and, thus, the power to reliably identify differences between groups.

In sum, we find partial support for our hypothesis that, among the Suubi–Maka participants, families in the treatment group (those with CSAs) would report more savings than families in the control group (those without CSAs) over the course of the intervention. On the one hand, the intervention increased the likelihood of having money saved. On the other hand, the intervention had no significant effect on the reported amounts saved. Furthermore, we found no evidence to support our hypothesis that, over the course of the study, more families in the treatment group would report keeping their money in financial institutions such as banks, microfinance institutions, and credit unions, than those in the control group.

We do find gender differences in savings reported among both children and their guardians. Females (both children and their guardians) were less likely to report having money saved. In addition, female guardians reported having smaller amounts saved than did male guardians. The significance of gender in explaining savings in matched accounts has not been consistently supported by empirical evidence. While most studies point to the importance of gender in affecting savings performance (Mills et al. 2008; Han et al. 2009; Huang 2010; Friedline and Elliott 2011), research on savings performance among poor families in sub-Saharan Africa (Ssewamala, Ismayilova, et al. 2010; Ssewamala, Karimli, et al. 2010) has found no significant variation in savings by gender. It is important to note that all of these studies use administrative data on savings (available through financial institutions) as opposed to self-reported data on saving performance. Against that backdrop, we conducted additional analyses of administrative savings in Suubi–Maka accounts (shown in appendix 2, which is available online) to examine the relationship between main demographic characteristics employed in this study (including gender) and two measures of administrative savings: having an active CSA (yes/no) and amount saved in CSAs. Unlike self-reported measures of savings, administrative measures of savings did not significantly vary by either the child’s or guardian’s gender. However, similar to self-reported savings, administrative measures of savings significantly varied by the guardian’s age and the type of guardian. More specifically, families with older guardians had greater savings in CSAs than families with younger guardians (B = 0.05, 95 percent CI = 0.02, 0.08, p < 0.01). Furthermore, families in which the participating child was reportedly being cared for by grandparents saved less than families in which the participating child was reportedly being cared for by a surviving parent (B = −1.31, 95 percent CI = −2.33, −0.29, p < 0.05).

The findings shed some initial light on the issue of whether match-savings interventions may provide an incentive for families to reshuffle their assets, rather than growing them. Our findings suggest that, over the course of the intervention, families assigned to schools in the treatment condition did not reduce other types of assets as they increased their savings behavior when compared to families in schools assigned to the control condition, which is consistent with our hypothesis that families would put new money into the CSA rather than reshuffle existing household assets At all three waves, including the baseline (Designed-based F = 6.67, p < 0.05), the 12-month follow-up (F = 7.31, p < 0.05), and the 24-month follow-up (F= 6.51, p < 0.05), there are statistically significant differences between the treatment and control groups in the proportion of families owning certain items (i.e. house, rental property, land, bicycle, car, coffee garden, potato garden, cassava garden, goats, and pigs; see appendix 3, which is available online). At wave 3, control group participants did report having more types of assets than treatment group participants, but this was also true at baseline and at other waves. Our study does not provide evidence that families in the treatment group relinquished the types of assets they held in order to take advantage of the 2:1 match on savings. Of course, it could be that liquidation did occur in ways that are not captured by our measure. Specifically, families may have liquidated some but not all of a particular type of asset, e.g., they sold a goat and kept one.

Limitations

Using self-reported data on savings is, perhaps, the biggest limitation of this study but, at least among participants in the treatment group who opened CSAs, we were able to estimate the validity of self -reports by comparing them to data on savings from the financial institution. To take advantage of the randomized experimental design of the study, we used the self-reported items administered to both the treatment and control condition at all three waves, including baseline. This is not unlike many studies that rely on self-reports of household income, expenditures, and consumption, especially when gathering data on savings specifically by children and adolescents (Elliott 2009; Elliott et al. 2010, 2011). Nevertheless, the sensitivity of self-reports to specific forms of bias may lead to inaccuracies in estimated effect sizes (Gibson 2006; Beegle et al. 2012). Our analyses comparing self-reported data on savings to administrative records suggest that adults may provide a more accurate assessment of savings than children, but even among guardians, less than half of the variance in savings from administrative records was explained by guardians’ self-reports of savings. Furthermore, it is possible that families save money and then withdraw it to pay for a child’s schooling or to invest in a small business. Any money that was withdrawn would be excluded from savings as measured by this study because participants were asked to report the amount of money they had in savings at the time of the interview. Future studies that use self-reports to measure savings should include a question that addresses withdrawals preceding the measurement period (for example, withdrawals from savings a few weeks or days before the interview period).

As noted above, another limitation is that the measure of household assets used in this study is a count of different types of assets and does not capture variation in assets within different categories (e.g. families having 2 goats and families having 3 goats are treated as equal). Without in-depth information on families’ consumption and expenditure patterns and families’ financial management strategies, we do not know exactly how families in either the treatment or control groups managed to save money. Families in the study are taking care of AIDS-orphaned and vulnerable children in one of the poorest regions on earth. One can speculate that, in order to save and invest in their children’s future education, these families must be making some sacrifices. Unfortunately, we do not know what kind of economic sacrifices these families are making. Future studies would benefit from detailed financial diaries, indicating daily and weekly consumption and savings patterns, and showing how these patterns may be affected by a family’s participation in a subsidized matched CSA similar to the one examined in this article.

Finally, our results point to significant school effects for some schools (additional analyses not reported in this article show persistent significant school effects regardless of which school we choose as the base/reference school). We do not know for sure the reason for these school differences. We can only speculate that they might be due to closer physical proximity of some schools to financial institutions. Perhaps participants in schools located closer to the center of the political district (for example, school #3) may have had readier access to the physical branch location of a given financial institution to make a deposit, since most financial institutions tend to locate their branches in the center of a political district. This argument assumes that children were making deposits after school, enabling them to minimize transport charges for traveling to the bank’s physical location. Further research is needed to advance our understanding of what exactly accounts for significant school effects on savings. In addition, larger studies with a greater number of clusters (schools) would help address the concern that a small number of clusters may not provide enough power to reliably estimate random slope variances at the school level.

Conclusion

Compared to other development approaches, including direct investments in education, health, and physical infrastructure development, CSAs specifically aim to improve the financial well-being of children by changing the financial behavior of their families in ways that increase savings and asset accumulation. CSAs are based on the proposition that changing parents’ and children’s saving attitudes and behaviors contributes to the long-term goal of inter-generational mobility out of poverty. Further research and additional evidence are required, however, to fully support this claim. It is important to emphasize that prior research has assessed the effect of participation in CSAs on the well-being of orphans. Specifically, Fred Ssewamala and colleagues examine the effect of the intervention on orphans’ educational outcomes (performance, educational plans, and aspirations) and psycho-social outcomes (orphans’ future orientation, mental health, and sexual risk-taking behavior) in Uganda (Ismayilova, Ssewamala, and Karimli 2012; Ssewamala et al. 2012; Han, Ssewamala, and Wang 2013). In this article, our focus is exclusively on estimating the effect of participation in CSAs on savings by children and their caregivers.

Our findings are not conclusive. Although participation in CSAs increased the likelihood that families reported saving, we find no intervention effect on the self-reported amount saved or on the likelihood of saving in formal financial institutions. Given these results, we suggest that further research is necessary prior to making policy recommendations. More specifically, a richer assessment of intra-household financial management strategies is needed to understand who makes the decision on saving, how this decision is communicated and implemented within the family, and how this decision varies depending on the relationship to the child (an orphaned child hosted by an extended family member vs. a biological child; a girl child vs. boy child; a HIV-positive child vs. HIV-negative child). In addition, administrative data on savings among both treatment and control group participants would strengthen the rigor of evaluations of savings programs. However, this step is unlikely to be sufficient to gauge the potential effect of CSAs on savings in the context of poor, unbanked households in developing countries. For example, in Uganda, where only 20 percent of the population have bank accounts (Demirgüç-Kunt, Beck, and Honohan 2008), verifiable data on savings in the control group may not be available merely because participants in the control group are unlikely to have bank accounts. Saving may most likely occur through means that may be best captured through reports from family members themselves.

Appendix

Appendix 1.

Saving performance in Child Savings Accounts (n=179)

Variable	n	Percent or Mean	95% Confidence Interval
Opening Child Savings Accounts (%)
Opened, activated, and deposited	118	66	[41.5; 84.1]
Opened, activated, but never deposited	14	7.8	[5.1; 11.9]
Opened, but never activated	14	7.8	[1; 43.6]
Did not open	33	18.4	[5.6; 46.1]
Average number of month account was open (Mean, range: 9–20)	146	17.8	[16; 19.6]
Average amount saved, UGX (Mean, range: 100- 38,160)	118	5,477	[2,437; 8,516]
Lower quartile (25% percentile)		1,053	[883; 1655]
Upper quartile (75% percentile)		7,383	[5,263; 11,510]

Appendix 2.

Association between demographic characteristics and administrative savings in CSA (n=179)

Predictor Variables	Model 1 Has an active CSA?		Model 2 Total savings in CSA (LOG)
	Odds Ratio [95% Confidence Interval]		Beta-coefficient [95% Confidence Interval]
Child’s gender (reference: male)	1.23	[0.87; 1.74]	0.23	[−0.31; 0.77]
Child’s age	1.16	[0.99; 1.35]	0.13	[−0.11; 0.37]
Child’s orphanhood status (reference: double orphan)	1.12	[0.69; 1.84]	0.55	[−0.29; 1.38]
Guardian’s gender (reference: male)	0.92	[0.61; 1.37]	−0.27	[−1.02; 0.48]
Guardians-grandparents (reference: parents)	0.76	[0.40; 1.44]	−1.31*	[−2.33; −0.29]
Guardians-other relatives (reference: parents)	0.84	[0.53; 1.32]	−0.18	[−1.03; 0.67]
Guardian’s age	1.01	[0.99; 1.03]	0.05**	[0.02; 0.08]
Number of people in the household	0.99	[0.92; 1.07]	−0.03	[−0.18; 0.12]
Number of earners (reference: double-earner)	0.76	[0.51; 1.13]	−0.49	[−1.34; 0.36]
School Fixed effects(a)	χ2(4) =41.52***		χ2(4) =213.78***
Constant (mean intercept)	0.02***	[0.002; 0.17]	−2.52	[−6.09; 1.06]
Wald Test for the overall model	χ2(13) = 52.78***		χ2(13) = 333.21***

^*p≤0.05,

^**p≤0.01,

^***p≤0.001

Boldface type indicates statistically significant results

^(a)Results of 4 DF Wald Test for school effect

Appendix 3.

Household Assets

	Percent or Mean [95% Confidence Interval]
	Total (n=346)	Treatment (n=179)	Control (n=167)	Design-based F
Household Assets (Mean, Range: 0–16)
Baseline	6.9 [6; 7.8]	6.2 [4.98; 7.4]	7.7 [7.2; 8.1]	6.67*
12-month follow-up	7.25 [6.2; 8.3]	6.4 [5; 7.8]	8.1 [7.8; 8.5]	7.31*
24-month follow-up	7 [6.4; 7.6]	6.5 [5.7; 7.3]	7.5 [7.2; 7.9]	6.51*
Individual items in the Household Assets scale
1. Family owns house (Yes, %)
Baseline	80.6 [68; 88.8]	74 [56; 86]	87.4 [77; 93]	3.63
12-month follow-up	81 [65; 91]	72 [52; 86]	91 [79.5; 96.6]	6.3*
24-month follow-up	81 [66; 90]	71.7 [53; 85]	91 [88; 93.7]	14.8**
2. Family owns rental property (Yes, %)
Baseline	10.7 [6.9; 16]	14 [9.4; 20.5]	7.2 [4.2; 12]	5.5*
12-month follow-up	11.7 [6; 20.9]	17 [10; 28]	5.6 [3; 10]	11.5**
24-month follow-up	13.5 [8; 21.9]	19.6 [15; 25]	6.8 [3; 14]	10.9**
3. Family owns land (Yes, %)
Baseline	82 [70.6; 89.7]	73.6 [59; 84]	81 [85; 94.7]	12.6**
12-month follow-up	82 [71.7; 89]	75 [64; 83.6]	89 [80.5; 94.5]	7.2*
24-month follow-up	82 [74; 88]	76.9 [67; 84]	87.6 [78.6; 93]	4.5
4. Family owns bicycle (Yes, %)
Baseline	46.4 [36; 56.8]	39 [27.6; 52]	53.9 [42.8; 64.6]	3.74
12-month follow-up	53 [42; 63.6]	43.9 [32; 56]	63 [56; 69.8]	9.4*
24-month follow-up	45 [36; 55]	39.5 [29; 51]	51.6 [41; 61.9]	3.03
5. Family owns motorcycle (Yes, %)
Baseline	9 [5; 14.9]	7.9 [3; 18.9]	10 [5.7; 17.6]	0.29
12-month follow-up	9 [4.5; 18]	11 [4; 25]	7.5 [2.5; 20]	0.4
24-month follow-up	10.8 [6.5; 17]	12.8 [6; 24.5]	8.7 [4.7; 15.5]	0.9
6. Family owns car (Yes, %)
Baseline	3 [1; 8]	5 [2; 11]	1 [0.2; 5.8]	3.9
12-month follow-up	3 [1; 6.5]	5 [3.8; 7]	0.6 [0.06; 6]	5.97*
24-month follow-up	5 [3; 8.5]	5.8 [2.9; 11.3]	4 [2; 9]	0.4
7. Family owns banana garden (Yes, %)
Baseline	73 [61.8; 82]	64 [50.9; 75]	83 [76; 88]	11.5**
12-month follow-up	74.5 [62; 83.7]	69 [51; 83]	80 [68; 88]	1.7
24-month follow-up	73.6 [62; 82.5]	66.5 [50; 79.5]	81 [71.7; 88]	4.3
8. Family owns coffee garden (Yes, %)
Baseline	46.4 [34.5; 58.7]	37.6 [29; 46.8]	55.7 [39.7; 70.6]	4.98
12-month follow-up	50.5 [33.7; 67]	39 [25.8; 55]	62.5 [38; 81.8]	3.4
24-month follow-up	39.8 [30.6; 49.8]	31.8 [23.9; 40.9]	48 [41.9; 55]	11.2**
9. Family owns sweet potato garden (Yes, %)
Baseline	71 [57; 82]	63.5 [43.7; 79.6]	79.6 [69; 87]	3.57
12-month follow-up	77 [60; 88]	64.7 [46; 79.7]	90.6 [83.6; 94.5]	16.1**
24-month follow-up	63.8 [54.9; 71.8]	60.7 [52.7; 68]	67 [51.5; 79.7]	0.7
10. Family owns cassava garden (Yes, %)
Baseline	70 [56.9; 80]	61.8 [46.6; 75]	78 [64; 88]	3.9
12-month follow-up	76.9 [62.6; 86.9]	65 [49.7; 78]	89 [79.7; 94.7]	12.7**
24-month follow-up	69.8 [64; 74.9]	68.8 [59; 76.8]	70.8 [64.5; 76]	0.2
11. Family owns other gardens (e.g. with beans, maize, greens, etc.) (Yes, %)
Baseline	67.5 [55; 77.7]	59.6 [41.8; 75]	76 [66.6; 83.5]	4.2
12-month follow-up	74.5 [57.8; 86]	63 [44; 78.6]	86.9 [77.7; 92.6]	10.2*
24-month follow-up	83.5 [74; 90]	79 [62; 89.9]	88 [80; 93]	2.2
12. Family owns cow(s) (Yes, %)
Baseline	21 [14; 30.8]	25.8 [16; 38]	16 [8; 29]	1.9
12-month follow-up	24 [17.4; 32]	26 [20.2; 32.8]	21.9 [11.3; 38]	0.36
24-month follow-up	18 [12; 26.6]	18.5 [9.9; 31.9]	18 [10; 29]	0.01
13. Family owns goat(s) (Yes, %)
Baseline	30.7 [25.4; 36.6]	25.8 [19; 33.5]	35.9 [31; 41]	6.3*
12-month follow-up	34 [25.9; 43.7]	24.9 [19.8; 30.7]	44 [38; 50.7]	27***
24-month follow-up	29.9 [23; 37.9]	25 [21; 30]	34.8 [24; 47]	3.22
14. Family owns pig(s) (Yes, %)
Baseline	53 [40.6; 65.7]	43 [30.8; 56.6]	64 [46; 78.7]	4.56
12-month follow-up	51 [39.8; 62]	39.9 [33; 46.7]	63 [50.8; 74]	13.9**
24-month follow-up	62.6 [49.6; 74]	49.7 [39; 60]	76 [72.7; 79.7]	33.4***
15. Family owns poultry for sale (Yes, %)
Baseline	19 [13; 26.8]	16.3 [9.6; 26]	22 [14.9; 31.7]	1.23
12-month follow-up	14 [10; 20]	15 [8.9; 24]	13.8 [9; 20.6]	0.09
24-month follow-up	13.8 [9; 20]	17.9 [12; 25]	9 [4.8; 17.5]	4.2
16. Family owns other animals (Yes, %)
Baseline	6 [4; 9]	7.9 [5.8; 10.6]	4.8 [2; 9.5]	2.24
12-month follow-up	7 [3; 15.8]	10 [4; 24]	3.8 [1.5; 8.9]	3.6
24-month follow-up	7.8 [4; 13]	6.9 [3; 15]	8.7 [4; 17]	0.22

^***p<0.001,

^**p<0.01,

^*p<0.05.

Boldface type indicates statistically significant results.