Financial Services to the Unbanked: The Case of the Mzansi Intervention in South Africa

The Mzansi intervention is a major initiative designed to provide banking services to the unbanked South African population. This study investigates the underlying variables that define the choice of a Mzansi account from a consumer perspective. Unlike previous studies, we do not assume that demand for financial services is a given but instead that it is underlain by perceptions and attitudes. Financial attitudes and perceptions are found to exert significant effects on financial choices. In particular, aspirations and forward-looking values are instrumental in facilitating access to finance.


Introduction
The lack of financial access by the majority of the population in developing countries is a significant concern. Because it may not be possible to directly observe the latent demand for financial products, supply-side constraints have thus far dominated the discussion. There is, however, another much more entrenched belief that "those who have access but choose not to use services pose less of a problem for policymakers" (Beck & Demirguc-Kunt, 2008). Over the years, through a coordinated effort, the barriers to access have been considerably lowered in many developing countries.
Relaxing the supply-side barriers brings the potential lack of demand for financial services to the forefront with important policy implications. Hence, studying the determinants of demand for financial services becomes an important research question.
The demand determinants define both the opportunities and the limitations of coordinated policy interventions designed to bring financial services to the needy. In this paper, we will examine a policy initiative, namely, the Mzansi intervention in South Africa, aimed at providing access to a basic preentry level of banking account for individuals who were previously excluded from the formal financial system. We investigate the drivers defining the uptake of Mzansi based on the genuine observed behavior of individuals and relate these results to their characteristics, attitudes and motivations. Section 2 discusses the features of the Mzansi account in detail, followed by a conceptual discussion in Section 3. Section 4 discusses data and methodology, followed by a discussion of the results in Section 5.

Mzansi Account
Basic or entry-level bank accounts designed to promote financial inclusion have been introduced in many countries, and the South African approach is unique in several aspects. First, it is based on a risk-sharing coordinated intervention that draws on a linkage between private commercial banks and a state-owned bank.
Second, the framework for this intervention is based on a voluntary financial agreement among the participating institutions. In South Africa, the exclusion from financial services was combined with racial discrimination, compounding the effects of poverty (Meagher and Wilkinson 2002). In this context, the issue of access to finance is no longer confined only to the issues of poverty and social disadvantage but becomes mixed with race and hence acquires a particular politically sensitive significance.

Conceptual Framework
Until recently, most analytical attention in the discourse on access to finance has focused on supply.
there is ample evidence of what supply-side conditions (product design, market conditions and regulation) are required for expanding access to financial services (Beck, Demirguc-Kunt & Honohan 2009;Claessens, 2006). The implicit assumption of existing actual demand for financial services has lately been questioned (Anand & Rosenberg, 2008 individual requires a service not because of their particular background (e.g., age or whether they are married) but because they think (wrongly or rightly) that they need it. Such beliefs will be measured by perceptions, opinions and attitudes. Conditioning economic action upon beliefs is nothing radically new: economic performance has long been known to depend on variables such as trust (Foss, 2012;Hunt, 2012)  while the latter tracks national financial access patterns and pathways in terms of products, service providers and household financial decisions and perceptions. We primarily use data from the second heading.
In particular, we use 102 variables measuring different finance-related attitudes and beliefs. These variables are grouped into the following categories: basic literacy, understanding financial terms, targets for financial advice, financial education desired and financial perception. As explained above, we expect that financial perceptions and attitudes affect the demand for Mzansi. There is, however, uncertainty about exactly which values and attitudes determine financial choices. It is therefore advisable to take a more exploratory stance and not simply investigate how such attitudes impact financial choices but also try to identify which variables have an impact. For this reason, we used all potential measures that we could identify in the dataset that could potentially fit into our conceptual framework.
The full list of variables and their descriptions is available upon request. Providing it here would, however, be inappropriate due to the length of the list. Another universally acceptable practice is to provide some summary statistics for the data. Such an approach would also be impractical. Furthermore, the variables used are essentially indicators, and as such, their summary statistics are not particularly illuminating. One needs, nevertheless, some overview of what these variables are. Table 1 provides an overview of the categories in which these variables are grouped. Here, we will only briefly describe these categories to facilitate the discussion of the results. The understanding of the financial terms category includes 17 variables that quantify whether the respondents comprehend the actual meaning of a number of financial terms, ranging from simple ones such as bad debt and loans to increasingly more complex ones such as pieces of financial legislation. In principle, these variables can be considered to have a natural ordering.
One could hypothesize that clients would need some basic financial understanding to select themselves into the Mzansi intervention, but having better understanding would mean that they would not be satisfied with the basic features offered by Mzansi and would require more sophisticated products.

Methodology
In this particular study, we face a problem of an un- for the modeling framework of the decision to choose a financial service (Mzansi in this case). In particular, the logistic regression is a common choice for a modeling framework in such situations. Logistic regression belongs to the class of generalized linear models (Mc-Cullagh & Nelder, 1989). The next step is to find a method that can select which variables to include in the model. Unfortunately, estimating a 'grand' regression with all candidate variables and somehow restricting it is not a viable option. Similarly, Bayesian (or frequentist) model averaging procedures (as e.g., Białowolski, Kuszewski, & Witkowski, 2012) are computationally too expensive, and although they provide useful measures of model uncertainty, the latter can be difficult to interpret. The standard tools for statistical inference of regression results, such as t statistics and F tests, are based on the implicit assumption that the set of predictors is fixed in advance, which is, however, not the case here. We need to choose this set adaptively, using some formal (e.g., stepwise regression and all-subsets regression) or informal (e.g., researcher selecting variables that provide good fit) procedure.
Under the adaptive selection of regressors, the classical tests are biased and therefore unreliable tools, resulting in wrong models and erroneous inference.
Because the above variable and model selection procedures are based either directly on the F statistics or some related statistics, they will be biased. Two possible strategies can be used to circumvent the above problem. One is to apply bias-reducing adjustments to sequential F tests in an adaptive variable selection algorithm. The other option we will follow here is to implement a totally different approach to variable selection, namely, a regularization estimator.
In this paper, we apply what is most likely the best known and most popular regularization estimator, namely, the adaptive LASSO. Tibshiranni (1996)   Consequently, this method has been studied and extended (Fan & Li, 2001;Wang, Li, & Tsai, 2007).
Although consistent in terms of variable selection in that it retains the important variables, the original LASSO estimator applies a fixed amount of shrinkage to all coefficients, which can be a problem when the so-called oracle property is desired. In particular, the LASSO estimator can be an oracle only under certain special circumstances subject to non-trivial conditions (see Zou, 2006 for details). In simple terms, an estimator possessing the oracle property will have the same asymptotic distribution for the coefficient estimates as the 'oracle' estimator, or in other words, the estimator is implemented knowing which coefficients are zero. This step allows an oracle estimator to be used not only for variable selection but also for inference. Nevertheless, implementing an adaptive amount of shrinkage for each regression coefficient leads to estimators that possess the oracle property (Zou, 2006;Wang et al., 2007).
Financial Services to the Unbanked: the case of the Mzansi intervention in South Africa The adaptive LASSO can be formally defined as follows:

Results
We apply the adaptive LASSO to the logistic regression of the choice of Mzansi account on the set of 102 variables described in the data section. Table 2 presents the selected independent variables (i.e., the variables with non-zero coefficients) and their esti-     Yi and Xu. (2008). Popular alternative formulations are given in Griffin and Brown (2007), Park and Casella (2008) and Hans (2009;2010). The Bayesian adaptive LASSO and the Bayesian T-shrinkage were implemented following Sun, Ibrahim and Zou (2010).
A good overview of alternative Bayesian regularization priors is given in Fahrmeir, Kneib and Konrath (2010).
Because, unlike their frequentist counterparts, Bayesian LASSOs do produce coefficient estimates that are identically zero, some (hard) thresholding is necessary to achieve variable selection. In Table 4, we present two different versions of such hard thresholding, labeled 1 and 2 to encompass both a liberal and conservative choice for the latter purpose. One may notice that the main results in the paper (e.g., aLasso) typically fall in-between these two choices). For ease of interpretation, the labels of the variables with negative effects in Tables 2, 3 and 4 are given in underlined bold typeface.
Full results from these alternative estimators, as well as some other unreported estimators (e.g., several versions of Bayesian model averaging), are available from the authors upon request. The main purpose of these additional results is to demonstrate the robustness of our findings and interpretation, which we will not discuss in any detail. Without entering into too much detail, all methods chose similar variables, with certain methods replacing some of the variables selected by the lasso with one or more variables from the same group, which are amenable to similar interpretation. The reasons for these slightly different outcomes are technical and lie in the grouping properties of the different estimators and the way they handle correlated variables in small samples.
We now proceed to a discussion of the adaptive LASSO estimates (Table 2). First, the basic literacy variable does not discriminate between Mzansi account holders and non-holders. Because basic literacy can be expected to be a pre-requisite to any form of access to finance, it could be expected to be able to discriminate between individuals with and without such access. Note, however, that because the individuals without Mzansi comprise both people lacking access and people who use better accounts than the basic Mzansi access, such a variable cannot differentiate Mzansi holders from the rest. determinants (see e.g., Lusardi and Mitchell, 2013) explains why so many financial literacy measures are retained in our model. More specifically, the retained variables are as follows: Bad debt, Ombudsman, Interest rate capping, Debt counseling and Garnishee order or emolument order. Considering that this category has an almost natural hierarchy from simpler to more complex terms, the selected terms lie towards the lower end of this hierarchy. Considering the role of awareness (Bönte & Filipiak, 2012), in that not being aware of (or, here, not understanding) certain financial products can preclude access, one can view the above hierarchy as a proxy for the financial awareness of the respondents. It is thus logical that some basic awareness is necessary to step onto the access ladder. From this perspective, one can say that individuals choosing Mzansi only possess understanding of certain basic financial terms and thus have a minimal financial education/awareness. Such a basic understanding can be considered a pre-requisite for access to finance. We can go even further and state that the actual motivation to seek access to finance would require such a basic financial understanding, and then this motivation would lead to actively seeking financial services. If   Financial perceptions grid 1 -You find financial products that are provided by a local person to be cheaper than the products provided by a large organization -   191-206 2014 be overshadowed by the present climate of financial austerity. This result is nevertheless important and deserves further investigation. The last variable may appear slightly counterintuitive. It refers to the fact that having a will makes one less likely to opt into Mzansi.
What is counterintuitive is not the effect itself, which is rather logical, but the actual inclusion of this variable. This variable suggests a higher level of financial planning, which, logically viewed, would be characteristic of individuals with a better level of financial access.
Bearing in mind that its coefficient is very small, however, one should not attach to much significance to this particular result.

Conclusions
This study attempts to identify the underlying variables that distinguish users of Mzansi accounts from non-users. Drifting away from the body of literature that investigates supply-side determinants of access to finance, this research questions the assumption that willingness (a loose definition of demand) to access financial service is a given. The tools used to gauge willingness are perceptions and attitudes.
In terms of perceptions, transactional demand for money was observed as the most important driving factor for possessing a Mzansi account, followed by savings and debt. The positive contribution of savings and debt preferences suggests that sustained employment schemes, for instance, can lead to an up-scaling of the financial demands of Mzansi account holders. Hence, the sustainability of holding the account actively depends on the extent of policy influence. The observation that Mzansi account holders seek information from independent financial advisors indicates the level of importance attached to verifying information received from traditional banking personnel. That is, financial literacy programs should be informed by a diversity of sources, financial desires and revealed perceptions.
There are certain inherent limitations to this study.
First, we do not consider background variables, i.e., variables that could define the perceptions and attitudes. Such variables could, however, be added to our methodology to complement the characterization of the Mzansi choice. Furthermore, the cross-sectional nature of the data used only allows us to investigate static questions. Finally, the real choice is between Mzansi and other low-cost alternatives, which we did not consider in this particular application. Note, however, that all these limitations can be overcome in a subsequent study building upon the present results.
After all, the characterization of the Mzansi choice, i.e., what exactly defines it, is a prerequisite for any such alternative explorations. Furthermore, applying the analysis to dynamic data or incorporating alternatives would create endogeneity issues, which cannot be reliably addressed unless the basic underlying structure of a decision problem is fixed. The methods in this paper achieve this task and therefore provide a reliable platform for such extensions.