The Impact of Savings and Credit Cooperatives on Household Welfare: Evidence from Uganda

Savings and Credit Cooperatives (SACCOs) help in reducing the financial exclusion gap. This study examines whether SACCOs improve the welfare of households. Data used are from 2009/2010 and 2010/2011 World Bank’s Living Standards Measurement Surveys (LSMS) done in Uganda by the Bureau of Statistics. Treatment cases are households that saved in SACCOs only while control cases are those that did not use the services nor save in SACCOs, banks or microfinance institutions. Propensity Score Matching and a two-step Treatment Effects’ model are used. Findings show that SACCOs have a positive and significant impact on household dietary diversity score, food consumption score, household clothing/footwear expenditure, and school enrollment rates in Uganda. The results are robust to hidden selection bias. The results show that SACCOs play a key role in improving household food security, non-food expenditure, and human capital development for the poor facing financial exclusion from banks and traditional microfinance institutions.

(US$94.55 million), share capital of Shs 178 billion (US$80.91 million) and income of about Shs 60 billion (US$27.27 million). Through the Ministry of Trade, Industry and Cooperatives (MTIC), the government invested US$ 134 million for subsidized loans to individuals and small businesses through the government-owned Microfinance Support Center (MSC) to SACCOs (MTIC, 2016). A survey done in Uganda by EPRC (2013) reveals that SACCOs which were legally constituted, but not controlled by the central bank were an option of choice second to commercial banks in terms of adults holding an account at a financial institution. The share of the adult population that operated an account increased from 5% percent in 2009 to 21% in 2013. The study also reveals that about 61% of the total users of SACCOs were women and 87% of all adult users of SACCOs were in the rural areas of Uganda (EPRC, 2013). As indicated above, SACCOs are a potential source of financial services to a large fraction of Ugandans who are excluded from commercial banks and traditional microfinance institutions. By 2013 there were 1,900 operational SACCOs in Uganda. However, many SACCOs have organizational challenges that impede their service delivery. These include lack of proper financial oversight and capacity, poor bookkeeping, and inadequately skilled staff and boards (BoU & MoFPED, 2017). Some literature indicates that governance remains the major weakness SACCOs in developing countries (Labie & Pé rilleux, 2008;Cuevas & Fischer, 2006;Cornforth, 2004;Branch & Baker, 2000). Our contribution to the literature is that we examine the impact of SACCOs on household welfare in Uganda and to the best of our knowledge there has not been any rigorous study that has examined this impact in terms predefined household outcomes. We provide a rigorous econometric impact assessment of SACCOs using propensity score matching methods that are complemented by the two-step treatment effects model with bootstrap corrected standard errors.

Background
To evaluate the impact of access to SACCO services on household welfare, we first control for potential differences between the treatment and control cases. In this study we restrict our sample to households that use the services and actually save in SACCOs only (treatment cases) and compare them with those that do not use the services nor save in any formal or semiformal financial institution, such as banks, microfinance institutions, SACCOs, etc. (control cases). To control for possible hidden selection bias, this study adapts the propensity score matching (PSM) method following Rosenbaum and Rubin (1983), Dehejia and Wahba (2002), Jalan and Ravallion (2003), DiPrete and Gangl (2004), Smith and Todd (2005), Mendola (2007), and Caliendo and Kopeinig (2008). The advantage of using PSM is that it does not require exclusion restrictions or a given specification of the functional form of the selection model to construct the counterfactual as well as reduce self-selection bias. Denote the indicator variable for participation in a program or treatment as D = 1 for participants and D = 0 for non-participants. For a given treatment we have the observed mean outcome under the condition of  Rosenbaum & Rubin (1983) and Caliendo & Kopeinig (2008), the parameter of interest in this study is the average treatment effect on the treated group (ATT) where In practice we observe the mean outcome E[Y 0 |D = 0] but do not observe mean outcome, E[Y 0 |D = 1].
We thus use PSM to extract for comparison the observed mean outcome of the non-treatment cases, To fulfill the condition in (2) there is the conditional independence and the common support assumptions (Rosenbaum & Rubin, 1983). The predicted probability for each household is the propensity score, P(x) = Pr(D = 1| X) and the overlap condition implies that 0 < Pr(D = 1|X) < 1. From (2) we have

Testing the Quality of PSM Matching
We test the quality of matching to make sure that none of the observable characteristics are significantly different between treatment and control households after matching to reduce the effects of confounding observable characteristics. In addition, both the likelihood ratio chi-square statistic for the joint significance of all covariates and the pseudo-R 2 from the probit/logit of treatment status on covariates should decline after matching. The joint significance of all covariates should be rejected as given by the high p-value of the likelihood ratio chi-square statistic (Rosenbaum & Rubin, 1983).
small amounts of hidden selection bias. For instance, with a value of Γ=1.10 at 5% or 10% level of significance, there should be concern that hidden selection bias is a serious threat to the validity of the study findings. In this study we consider the value of Γ = 1.20 at 10% level of significance as the lowest cut-off safe point that is far enough away from Γ=1.0 to allay concerns about the influence of unobserved confounding on the ATT study findings. We supplement the Rosenbaum Γ tests with another test for hidden selection bias. Following Jalan & Ravallion (2003, we also test for potential remaining hidden selection bias of confounding factors using the Sargan-Wu-Hausman test.
We do this only on the sample of matched treatment and control households using Nearest Neighbor matching. Using only the matched sample of treatment and control households, we run an OLS regression of the outcome variable on the residuals from the probit selection equation, the propensity score, and a set of additional control variables that exclude the instruments used to identify exogenous variation in the outcome variable. If the coefficient on the residuals is significantly different from zero, then hidden selection bias is still a problem even after matching and it may compromise the estimate of the impact. If the coefficient on the residuals is not statistically significant, then we can assert that the impact estimate is a result of participation in the treatment.

The Treatment Effects Model (Switching Regression Model)
The decision to join and hold savings in a SACCO might not be exogenous to the households so we test whether or not assignment to treatment is endogenous using the Wu-Hausman test. We test the null hypothesis that assignment to treatment is exogenous. If the null is rejected, then we employ the two-step treatment effects model that also explicitly controls for hidden selection bias and compare the results to those obtained using PSM. We denote the outcome variable as Y i , and the treatment selection variable as D = 1 for the treatment households and D = 0 for the control households and x i as a vector of explanatory variables. We have the OLS outcome regression model given by: (4) Y i = β′x i + δD i + ε i where δ is the estimate of the impact of SACCOs on the outcome variable and ε i is the error term.
However, confounding factors will bias the estimate of δ. Therefore we use two-stage least squares approach while controlling for hidden selection bias. The first stage is the probit model that regresses the treatment selection variable, D i , on the vector x i of explanatory variables and a vector z i , of instruments. Denoting the D i * as the latent treatment selection variable, we have: The regression model observed only if D i =1 is given as: where the sample selectivity correction, λ(θ′w) is the inverse Mills ratio or the hazard function for the incidentally truncated distribution. Therefore the selection probit model is The predicted values of the treatment selection variable in (8) are used in the second stage OLS regression given by: where the treatment impact is now given by the parameter δ IV and the instrument variables in the vector z i , are assumed to be correlated with treatment D, but not with the error vector ε i .

Data Sources
The study uses household survey data from the World Bank's Living Standards Measurement Surveys

Indicator and Outcome Variables
In this study we restrict our sample to households that use the services of and actually saved in SACCOs only (treatment group) and compare them with those that do not use the services of nor save in banks, microfinance institutions and SACCOs or any formal or semi-formal financial institution (control group). The indicator variable is SACCO that takes a value of 1 for treatment cases and a value of zero for control cases. We define SACCO saving households as those hold savings in a SACCO only and not in banks or microfinance institutions. In most cases these households will have borrowed in the past from the SACCOs. We exclude households that only applied to borrow but did not save in the indicators for food security, which capture household food consumption and dietary diversity (Kennedy et al., 2011;Kennedy et al., 2010;Hoddinott & Yohannes, 2002 food and nutritional quality. Following the FAO, we computed the Food Consumption Score as a composite score based on 8 major food categories that any household member has consumed over the previous 7 days, multiplied by the number of days that the food category was consumed after being weighted by the nutritional importance of the food category. The total possible score ranges from 0 to 112. The major food categories were (a) main staples (cereals/tubers) with a weight of two; (b) pulses/legumes with a weight of three; (c) vegetables with a weight of one; (d) fruits/fruit juices with a weighted of one; meats, poultry, offals, blood, any fish type with a weight of 4; milk and milk products with a weight of 4; sugar/honey with a weight of 0.5, and oil and fats with a weight of 0.5 (WFP & FAO, 2012;Kennedy et al., 2011;WFP, 2009WFP, , 2008.

Descriptive Statistics
Descriptive statistics for the treatment group (households that use services of and hold saving with SACCOs, but not MFIs, banks) and the control group (households that do not use services of SACCOs, banks and MFIs) are presented in Table 1 below. The table compares a number of selected demographic and socio-economic variables for the treatment and control cohorts before matching the data. The independent two-sample t-test is used to test for significant differences between the means of the selected variables. The results show statistically significant differences in the means of various covariates between the two cohorts before matching, such as such electricity use, household income, and value of assets, HDDS, FCS, household clothing and footwear expenditure, and school enrollment rates. The results of quality of matching or covariate balancing are shown in the Appendix. As expected, matching achieves a reduction in the standardized bias, the pseudo-R 2 , the likelihood ratio chi-square and the statistical significance of the likelihood ratio chi-square. The reduced pseudo-R 2 indicates that covariates have very low explanatory power for selection into the treatment group. The reduced statistical significance shown by the p-values of the likelihood ratio chi-square indicate that there are no systematic differences in the distribution of covariates between the treatment and control cases after matching. That is, the hypothesis that both cohorts have the same distribution in the covariates after matching cannot be rejected. Tables 2 and 3    Note. ***,**,* indicate significance at 1%, 5%, and 10%, respectively. † † indicates the value of Γ at 5% level of significance. † indicates the value of Γ at 10% level of significance. a denotes the t-value for Lambda in the hazard function of the treatment effect model. Exchange Rate; US$1.00 = Shs 2,200

Impact of SACCOs on HDDS and FCS
The HDDS results in Tables 2 and 3 above are robust to whichever method of estimation that is used. households who held savings with SACCOs had higher mean Food Consumption Score (FCS) than that of the control group. The difference in the mean FCS between the two groups is statistically significant.
That is, when a household chooses to engage in SACCO savings services, on average, their mean FCS increases by 19% due to inter-temporal flexibility in consumption smoothing opportunities provided by SACCOs. The results are consistent and robust to whichever method of estimation that is used.
The findings above are in tandem with those from randomized controlled trials done by Ksoll et al.  (2013) suggest that increased household food expenditures can be linked to increased food quality. Household savings improve food security through increased food access. This is usually linked to the household's ability to purchase food or produce food. Zeller and Sharma (2000) argue that savings provide a pathway by which households accumulate capital to smooth consumption in difficult times. Inter-temporal consumption smoothing through savings helps households deal with income shocks or unexpected increases in expenditures. In Uganda vulnerable HHs self-insure against idiosyncratic risks across periods by holding precautionary savings in the form of relatively liquid assets (Kiiza & Pederson, 2006). Thus households that hold precautionary savings are able to adjust their income and consumption and in turn stabilize their food security through diet diversity, quantity and quality of food.

Impact of SACCOs on Household Clothing and Footwear Expenditure
The estimated impact of access to SACCO savings services on annual household clothing and footwear expenditure for the period 2010/2011 and 2009/2010 are presented in Tables 2, and 3 above; and Table   4 below. The difference between the mean annual household expenditure on clothing and footwear for the treatment and control group is about Shs 90,000 (US$41). This difference is statistically significant.
The results are consistent and robust to whichever method of estimation that is used. In Table 4 below we indicate that the mean household size of the treatment group is not statistically different from that of the control group after matching the data of the two cohorts. For instance, after Kernel matching, the mean household size for the control group in the 2010/2011 sample is 7.880. That of the treatment group is 8.351, the difference between the two means is not statistically significant, p-value = 0.272.
For the 2009/2010 sample, the mean household size for the control group after matching is 7.533 and for the treatment group its 7.585 and p-value is 0.907. Therefore it cannot be argued that, on average, the treatment group household size was far greater than that of the control group which would explain the large difference in clothing and footwear expenditure between the two groups. This finding is consistent with the study by Dupas and Robinson (2013) who conducted a randomized control trial www.scholink.org/ojs/index.php/jepf Journal of Economics and Public Finance Vol. 7, No. 3, 2021 47 Published by SCHOLINK INC.
(RCT) in Kenya for savings and find a positive impact on private expenditures, especially for market women. However, some studies that have used randomized evaluation methods find no statistically significant impacts of savings on non-food expenditures (Karlan et al., 2017;Beaman et al., 2014). For example, Karlan et al. (2017) find no impact of savings on non-food expenditures (such as transport, clothing, electricity, and petrol) for a clustered randomized evaluation spanning three African countries which include Ghana, Malawi, and Uganda. Our findings suggest that after controlling for household size and annual household income, treatment households spend more on clothing and footwear than the control households due to the inter-temporal flexibility in consumption smoothing opportunities provided by SACCOs.

Impact of SACCOs on Household School Enrollment Rates
The results in Tables 2 and 3

Tests for Robustness of the Results
We test for robustness of the results using several methods. First, we check whether the results are robust to the method of estimation. We examine the impact of SACCOs on HDDS, FCS, household expenditure on clothing and footwear, and school enrollment rates. We run the propensity score matching method with different algorithms and also employ the two-step treatment effects method while controlling for hidden selection bias. In all cases the results are very similar. Second, following Rosenbaum (2002Rosenbaum ( , 1987, we generate estimates of the magnitude of hidden selection bias that are necessary to invalidate the ATT study findings, that is, the parameter gamma Γ and check its value at 5% and 10% level of significance. The closer Γ is to the value of 1.0, the more sensitive the findings are to small amounts of hidden selection bias. In our study the lowest value of Γ is 1.20 at 10% level of significance (see Tables 2 and 3). We consider this a safe point that is far enough away from Γ = 1.0 to allay concerns about the influence of unobserved confounding on the ATT estimates. All the other estimates of Γ are far enough from 1.0 and indicate that our ATT estimates are not sensitive to hidden selection bias at 5% and 10% level of significance.
Third, we follow Ravallion (2003, 1999) and test for potential remaining hidden selection bias due to confounding factors using the Sargan-Wu-Hausman test after matching the treatment and control groups using Nearest Neighbor matching. For all the results in Tables 2 and 3 above, we run an OLS regression of each outcome variable on the residuals from the probit selection equation, the propensity score, and a set of additional control variables that exclude the instruments used to identify exogenous variation in each of the outcome variable mentioned above. If the coefficient on the residuals is significantly different from zero, then hidden selection bias is still a problem even after matching the data. In all cases, we find that the coefficient on the residuals is not statistically significant