We follow the method proposed by McGillivray (2005) to develop a non-income composite index. The index differs from McGillivray’s (2005) in that it focuses on quality of life measured at a micro level rather than a macro level. The essence of this method is to extract the first principal component from selected social indicators, using principal component analysis (PCA), that explains the most variance in the data set. In line with the OECD’s handbook (2008), the first extracted component represents an objectively weighted composite index, in this instance of non-income quality of life. Secondly, we regress the composite index on the natural log of household income per capita. Lastly, we retain the residual, μ_i, from this regression and interpret it as an individual’s income-independent quality of life (IIQoL). This estimated function of the composite non-income quality of life index can be expressed as:

Qit=a+βlnyit+μit[Eqn 1]

where Q_it is the composite non-income quality of life index for individual i in period t (t = 2008 to 2015); and lny_it is the natural log of household income per capita for the same individual i in the same time period t, with μ_it being the residual term. This residual term μ_it is a purely statistical construct and is defined inter alia as IIQoL, which is central to our analysis and by definition orthogonal with respect to lny_it. To test the choice of PCA to weight the composite index, we also developed an index in which we applied equal weighting. Correlating the two indices, we find the Pearson’s correlation coefficient to be 0.94. We prefer and report the results using the PCA method of weighting, as it weights the independent variables according to the most variance explained in the data.

To prove that our newly constructed IIQoL is indeed independent of income, we correlate it with both objective and subjective monetary measures of well-being (other than income). Low correlations are an indication of this independence of the newly constructed index.

The generic model estimated (equation 2) applies to: (1) the whole sample, (2) a sub-sample of ‘poor female’ respondents, (3) a sub-sample of ‘poor females residing in rural areas’ respondents and to compare findings based on gender (4) a sub-sample of ‘poor male’ respondents:

Yit=β0+β1Xit+εit[Eqn 2]

where Y_it is the IIQoL, β₁X_it is a vector of demographic and economic variables including the variable of interest M&ILs, and the idiosyncratic individual error term is given by ε_it.

To estimate the specified model in equation (2), we use panel data estimation techniques and adjust these for the complex sampling design. Cluster corrections are necessary, as the assumption that our sample is extracted by means of simple random sampling is not adhered to. We deal with the cluster correction at the geographical level, assuming that people within a similar cluster might have similar levels of quality of life and therefore similar needs for microloans. Additionally, we make use of panel weights to correct for attrition.

To determine the most appropriate estimation technique between fixed effects (FE) and random effects (RE), we make use of the Hausmann test. Based on the results, we reject the null hypothesis which states that the difference in coefficients is not systematic (chi2 = 1754.67; p = 0.000) and therefore we estimate the model using FE. Although panel data estimation techniques have the benefit, over cross-sectional estimation methods, to address endogeneity, which arises from omitted variables and measurement errors, it does not address simultaneity. Therefore, to test for endogeneity arising from simultaneity, we make use of instrumental variable (IV) regressions using the FE estimator. In saying this, the data set did not offer a suitable option to instrument M&ILs, our variable of interest. For that reason, following Cameron and Trivedi (2010) we instrument M&ILs by its lagged variable. With the IV regression, we find the instrument to be strong, with the Kleibergen-Paap Wald F statistic = 222.998 and the Stock-Yogo weak ID test’s critical value being 10%. However, testing for endogeneity, the Durbin-Wu Hausman test shows that the variable of interest, M&ILs, is not endogenous (p = 0.4304). Therefore, we do not pursue the matter further and interpret the pooled ordinary least squares (POLS) (in the case of time invariant variables) and the FE models.

To test the robustness of our results, we also run regressions on individual non-income quality of life indicators, including education, health and housing (Table 1-A1 in Appendix 1). Similar results on the relationship between M&ILs and the individual social indicators are indicative of robust results. One should keep in mind that it is not ideal to use these individual social indicators to represent quality of life, as they are one dimensional and, secondly, they are not income-independent.

In all analyses, we run diagnostic tests as appropriate. With the exception of the age and age-squared variables, there is no evidence of multicollinearity. However, we find evidence of heteroscedasticity and this problem is addressed by making use of robust standard errors.

The data used in this article comes from the first four waves of the National Income Dynamics Survey (NIDS), spanning the period 2008–2015. The NIDS is a face-to-face longitudinal survey, which is repeated with the same individual household members, every 2 years (NIDS 2016b). This data set focuses on the livelihoods of individuals and households over time. The reason for choosing the NIDS data set is because it provides rich data on non-income quality of life indicators at a micro level, which is not the case with other national surveys. The analysis is restricted to a balanced panel – adults 18+ years old (wave 1), who were successfully interviewed in all four waves and includes 9360 individuals in the four waves amounting to 37 440 observations. Table 1 provides the summary statistics for a select number of demographic variables in the balanced panel.

As seen in Table 1, there is a significant increase from 2008 to 2015 in the share of the population engaging in the M&IL market from 1.3% to 7.9%, although there was a minor decrease in wave 3. This finding of increased dependence on M&ILs holds for each of the cohorts of interest – ‘poor females’ and ‘poor females residing in rural areas’. The rate of increase in M&ILs among the poor cohorts compared to the whole sample is 3.5% higher on average. This higher level of activity by the poor could be interpreted as a higher number of individuals accessing the only financial service they are able to. This tendency to rely on loans is not uncommon in South Africa, with household debt as a percentage of nominal disposable income persistently remaining at high levels of more than 70% over the last decade (SARB 2019).

The percentage of the population that falls below the upper¹ poverty line declined significantly from 65.5% to 44.6% over the same time period. This is also true for females, as poverty decreased from 69% in wave 1% to 50% in wave 4 and among females in rural areas, where poverty decreased from 81% in wave 1% to 64% in wave 4. Notably, the percentage poor among the females and females in rural areas is higher than the percentage poor in the whole sample. However, although poverty rates decreased, we cannot conclude that higher levels of M&IL participation caused a decrease in poverty directly. Simanowitz (2010) argues that it would be very difficult to prove the positive effects of microloan initiatives on increasing the quality of life for South Africans, since many who participate are also recipients of welfare payments.²

With regard to the different geographical areas included in the sample: we refer to respondents residing in either rural or urban areas. In this article, people residing in rural areas are highlighted, as they are typically more marginalised and represent a substantial proportion of the whole sample: 44.71% in wave 1 to 50.5% in wave 4. Almost all inhabitants of rural areas are of African descent (on average 99.98%) as well as females (67%).

Firstly, to construct our IIQoL index, we apply PCA to our selected non-income quality of life variables: years of education, rooms in house per person (proxy for quality of housing) and objectively measured health. We extract the first component, which explains 58% (eigenvalue = 1.60) of the variance, deemed acceptable for the construction of a composite index (see Naudé, Krugell & Rossouw 2009; Greyling & Tregenna 2016). Secondly, we regress the non-income quality of life index on the natural log of household income per capita to derive the residual. The residual is what we identify as objectively measured IIQoL.

To prove the independence of our IIQoL measure from income, we correlate it to both objective and subjective measures of monetary well-being: expenditure per person and self-reported relative income. We find the correlation coefficient significant, though at very low levels of r = 0.10 and r = 0.07. At these low levels of correlation, we assume the independence of the IIQoL index and accept it as a good measure of IIQoL.

Figure 1 expresses the IIQoL of different groups as a percentage of the maximum IIQoL in South Africa. We find the mean IIQoL of the average citizen to be relatively low at 53.3%. This is not surprising, seeing that South Africa is a developing country with high levels of inequality and poverty. What should be noted is that within each of the samples represented – all South Africans, poor females, poor males and poor females in rural areas – those ‘with’ a M&IL have between 2% and 3% lower levels of IIQoL than those ‘without’. The sub-samples ‘poor females’ and ‘poor females residing in rural areas’, experience the lowest levels of IIQoL relative to the rest of the country.

Comparing the poor female and male sub-samples, males ‘with’ and ‘without’ M&ILs enjoy higher levels of IIQoL than their female counterparts. In general, the IIQoL for males is higher than for females and, the IIQoL for poor males ‘without’ these kinds of loans is even higher than the IIQoL as determined for the whole sample.

These initial results suggest possible negative effects of participating in the M&IL market on those it was supposedly created to help, and it also alludes to the existence of a gender gap. These possible relationships as well as the associated implications will be further investigated in Section ‘Results (dependent variable = IIQoL)’ as other factors might also be at play.

We find the estimated models for the sample as a whole and the sub-samples statistically significant (p = 0.00). To interpret the results, we use the FE estimations, although in the event of time invariant variables we also refer to the POLS results (see Table 4).

Table 4 reveals an interesting result, namely that M&ILs are statistically significant and negatively related to IIQoL for all sub-samples, except for ‘poor males’ in which the relationship is negative, but not significant. Therefore, we assume that South Africans in general ‘with’ M&IL have a lower IIQoL than those ‘without’. The not significant relationship to the ‘poor male’ sub-sample implies that M&ILs, controlling for all other factors, are not related to their IIQoL. Based on previous research (Karlan et al. 2016), it seems that poor men involved in the M&IL market use the funds to provide for current consumption expenditure needs, which is not related to non-income quality of life, explaining the not significant result. In contrast, as females are mostly the primary caretakers of children, they often spend their M&ILs in sectors related to the non-income quality of life of their children, such as education, housing and health.

To test the robustness of this negative relationship between M&ILs and IIQoL, we test it using the individual indicators of quality of life. However, these single indicators of education, health and housing, suffer from the critique of being income dependent and unidimensional (as discussed in Sections ‘Introduction’ and ‘Methodology’), contrary to our IIQoL index (OECD 2008; Saltelli et al. 2007). Notwithstanding this, the results in Appendix 1 (Table 1-A1), are similar to those reported in Table 4, in that these individual indicators of quality of life are statistically significant and negatively related to M&ILs.

The estimation results pertaining to the whole sample indicate both a gender and geographical inequality when it comes to IIQoL. Male respondents are deemed better off than females, which confirms the initial findings in Section ‘Principal component analysis’. Turning to the geographical area, as expected, the IIQoL of respondents residing in rural areas is lower than that of their urban counterparts. This is in line with the literature which suggests that females, and more specifically females residing in rural areas, are more marginalised than other groups (Bhorat et al. 2006; Kirsten 2011). This finding emphasises the importance of addressing the second and third research questions (see Table 4, columns 3 – 5).

The time variable for waves 2 to 4 is significant and negative for the whole sample indicating that the IIQoL decreased from wave 2 to wave 4. The reasons for the decrease in IIQoL, from wave 2 to wave 4, are associated with health (as time goes by it is more likely that a person was exposed to more health-related problems) and factors related to housing, such as service delivery. However, the wave variable is not significant for the poor sub-samples, except for wave 2 for the ‘poor female’ respondents. The poor might be less affected by the lapse in time as their health, type of housing and service delivery could have been less affected or not affected at all. One should remember that respondents in rural areas report to be healthier than those in urban areas. Furthermore, the housing of the poor is basic with limited services; thus the deterioration thereof is much less likely than for their richer counterparts.

In terms of race, the POLS results show that Coloureds have lower levels of IIQoL, whereas White South Africans experience higher levels, relative to Africans. This is indicative of the persistent inequality in access to health, housing and educational services, attributed to the racial discrimination of the old apartheid regime (Van der Berg 2011).

Age performed as expected in explaining IIQoL for the whole sample as well as the sub-samples ‘poor females’ and ‘poor females residing in rural areas’. This statistically significant and positive quadratic (U-shaped) relationship is not surprising, since the young generally have higher levels of health while the older respondents have better housing and more assets and access to service delivery than those in their middle ages. Therefore, it mirrors the findings from the subjective well-being literature (Frijters & Beatton 2012). In saying this, the opposite holds true if only the sub-sample ‘poor males’ is considered, with both the young and old experiencing lower levels of IIQoL than those in their middle ages. This might be since many of the middle-aged men are employed and residing in urban areas, which is positively correlated to IIQoL relative to the young and the old, who might not be employed and likely residing in rural areas, which is commonly accepted to be related to lowers levels of IIQoL.

Trust, which is an important determinant of IIQoL, is only positive and significant for the sub-sample ‘poor females in rural areas’. This is not surprising since women staying in rural areas are seen as having better social capital in the form of stronger bonds, that is, links to people based on a sense of common identity (‘people like us’) (Keeley 2007). This trust also forms the basis of their IIQoL as they depend on one another for support (child caring, toiling of lands) and survival (sharing of basic necessities) (Mutopo 2014; Vercillo 2016).

Employment is statistically significant and, against expectations, negative for the whole sample as well as for the ‘poor female’ sub-sample. South Africa suffers from very high levels of unemployment and very limited employment opportunities. In the whole sample we find 68% of the respondents are unemployed and among ‘poor females’ this percentage is higher at 84%. The fact that being employed is negatively related to IIQoL contradicts theory, although in the South African context, where the social welfare system is very well developed, it might be plausible.

More than 67% of the whole sample and 83% of the ‘poor female’ sub-sample reported to be recipients of some kind of social welfare grant. These grants are mostly dependent on a means test. Therefore, if a person is employed, they might earn more than the means test and lose their right to the grant. Those that are employed earn an average wage of R3560 per month and ‘poor females’ earn R1070. If household size is considered, the mean wage income per household for the whole sample is R1170 and for ‘poor female households’ it is R445 per month. Thus, their income is very low and travel costs to their place of work present a real cost associated with working, which could sometimes be more than half of their earnings.

Now, if individuals make use of the welfare system, we find the average household income per month per person to be R283 and for ‘poor females’ lower at R220. As the mean income earned is relatively low, the opportunity cost to work is high, and therefore respondents might prefer to be unemployed and receive a grant. They then have more time to spend on improving their IIQoL. This situation reflects the unintended disincentive of a social welfare system to discourage productive employment (Blau & Robins 1986; Camissa 1998; Keane 1995; Weaver, Shapiro & Jacobs 1995). This dependency on government grants is not sustainable or conducive to development, especially if the unemployment rate is already high.

Contrary to the above results, employment is not statistically significant in either the ‘poor male’ or the ‘poor female residing in rural areas’ sub-samples. This implies that being employed is not related to the IIQoL of ‘poor males’ and ‘poor females residing in rural areas’. As already mentioned, females are the primary caretakers of relatively large households, thereby shouldering most of the responsibility regarding children’s education and health care. This means that a large portion of their time is spent on household and childcare duties. At the same time, these unskilled females, a common phenomenon in rural areas, have very few real employment opportunities. This means that it is unlikely that they will be employed and therefore employment is not related to their IIQoL. As for the ‘poor male’ sub-sample, there are relatively few job opportunities for unskilled labour and they do not have the main responsibility of caring for children. Therefore, they most likely also earn some type of government grant, which makes them indifferent to being employed or unemployed.

Safety, measured as the likelihood of being a victim of a crime, rated on a scale from likely to very unlikely, is significant and positive for all samples. This indicates the importance of safety to the IIQoL of all South Africans, who are being impacted by crime on a daily basis both in the form of property (housing) and bodily harm (health).

Lastly, the results pertaining to the ‘poor female’ and ‘poor male’ sub-samples clearly contradict general populace⁵ in the subjective well-being literature, which shows that females often have higher levels of well-being than males. Since 2008, the South African Social Attitude Survey (SASAS) shows that females indeed do report lower levels of quality of life than males in South Africa (HSRC 2010). Kirsten (2011) points out that ‘African females living in rural areas’ is the most deprived demographic group in South Africa and that, even after 20 years of redistribution policies, their level of quality of life is still unacceptable. May and Norton (2012) argue that African women, in general, do not have control over their own income which excludes them from the decision-making process and the threat of violence remains a major form of control by men over women. This indeed, corroborates our findings and highlights the plight of this gender in terms of:

Work done by Teixeira and Chambers (1995) summarises it nicely: