This study uses Wave 5 of the National Income Dynamics Study (NIDS) (Southern Africa Labour and Development Research Unit 2018), which is a nationally representative survey that is part of a longitudinal survey of individuals and their households on income, expenditure, wealth and other socioeconomic measures in South Africa, to answer two research questions: (1) Are household heads’ characteristics associated with the distribution points of household wealth differently across districts in South Africa? and (2) Is there a significant difference in the variability of household wealth distribution within districts compared to that of between districts in the aforementioned associations?

The investigation of wealth and other related phenomena continues to attract particular attention in economics and other social sciences, both in developed and developing countries (Balestra & Tonkin 2018; Branson et al. 2024; Chatterjee, Czajka & Gethin 2020, 2022; Chelwa, Maboshe & Hamilton 2024; Jin & Xie 2017; Joubert & Van der Merwe 2021; Mojsejová & Marcinová 2023; Muyanga, Jayne & Burke 2013; Schmidt & Sevak 2006; Shifa et al. 2023; Sierminska, Frick & Grabka 2010; Soseco 2021; Vo et al. 2023; Vo & Ho 2023). One reason for such interest is that, beyond providing a financial cushion for households during economic hardships, wealth, particularly in the form of physical assets such as land or real estate or investments such as shares, can also serve as a channel for intergenerational mobility. Chelwa et al. (2024) note that additional income streams, such as rental income and dividends, can be generated through wealth. The reasoning outlined previously aligns with the Asset-Based Welfare Theory, which emphasises that asset accumulation is essential for long-term well-being and economic security (Bryant, Spies-Butcher & Stebbing 2024; Johnson & Sherraden 1992; Sherraden 1991; Steele 2020; Walks 2016). This is particularly relevant when examining the relationship between the characteristics of household heads and the wealth they accumulate. In addition, this study draws on stratification theory, especially in the context of systemic and structural inequalities in South Africa. Factors such as educational attainment, race, gender and ethnicity of household heads are considered either barriers or catalysts to wealth accumulation among South African households (Pfeffer & Waitkus 2021; Spilerman 2000).

Therefore, understanding this phenomenon is crucial for policymakers and academics alike. Besides, some studies aim to understand the dynamics of wealth inequality, while others focus on detecting its determinants, among other things. This study aligns with the latter perspective as it seeks to shed light on the relationships between household heads’ characteristics and household wealth.

The importance of scholarly attention to this issue is clear, as public policies designed to improve people’s lives should be grounded in sound scientific evidence. Before proceeding, it is worth pointing out that wealth in general remains a central issue in South Africa, both in public discourse and academic research. This is largely because of the persistently high levels of inequality in the country, despite the implementation of progressive public policies by the democratic government since 1994. The empirical analysis in this study aims to contribute to this ongoing debate by exploring whether the characteristics of household heads serve as determinants of household wealth in South Africa.

Another noticeable feature in the literature is the deficiency of studies that focus on developing countries, in particular when it comes to the employ of micro-level data (e.g. individuals and households) as opposed to studies that focus on the developed world. This is because of the scarcity of reliable data and other related challenges, such as the complexities involved in measuring wealth itself. For example, Chatterjee et al. (2022) discuss some limitations of the wealth data in the NIDS.

Regarding the NIDS, it must be noted that some studies (Chelwa et al. 2024; Daniels & Khan 2019) offer a counter-argument to the view expressed by Chatterjee et al. (2022), specifically highlighting steps taken in Wave 5 of the NIDS to improve the data. These works have concluded that the wealth variable(s) in Wave 5 of the NIDS is suitable for cross-sectional analysis. For example, Branson et al. (2024) and Chelwa et al. (2024) have used Wave 5 of the NIDS to assess wealth in South Africa at the micro-level.

Despite this positive development regarding micro-level data and empirical application in South Africa, two important methodological aspects still merit mentioning. Initially, the issue of NIDS data clustering has been overlooked in previous studies. It should be noted that individuals’ or households’ characteristics in the NIDS are nested within districts, which plausibly makes these observations exhibit dependency within districts. Therefore, traditional single-level econometric techniques used in previous empirical studies may not be suitable for the NIDS data. On the contrary, multilevel or fixed-effects modelling can be a suitable alternative.

While there is an ongoing debate in the social sciences literature regarding the suitability of multilevel versus fixed-effects and other modelling techniques for clustered data, as summarised in Huang (2016) and Oshchepkov and Shirokanova (2022), this study sets aside that important discussion for another time and instead opts for multilevel modelling guided by the second research question. In other words, by adopting a multilevel modelling approach, this study does not treat the issue of clustering in the NIDS data as just a nuisance, as is the case in studies that employ fixed effects and other prominent modelling techniques in economics. Furthermore, as Mulamba (2022) argues, multilevel modelling involves accounting for district-level clustering by dividing the variance of the disturbance term into two random components: a group-specific (i.e. district) random effects component and an error term applicable to all observations. This approach is crucial for addressing the second research question posed in this study, as further discussed in the methodology section.

Moreover, consistent with the argument presented by Bond-Smith et al. (2024), multilevel modelling in spatial economics implicitly recognises the influence of location – such as districts – on the relationships under study. This suggests that location- or district-specific factors can influence the economic phenomena under investigation. For instance, a cultural norm that promotes or hinders wealth accumulation might be prevalent in certain districts or regions, while it may be absent in others. From a multilevel modelling point of view, one does not necessarily need to observe these district-specific factors. Instead, the group random effects component of the model takes care of that, as is discussed further in the methodology section of this study.

The preceding discussion highlights the importance of understanding or gaining insight into location-specific factors for public policy-formulation purposes. While the approach taken is a step in the right direction, it should be pointed out, however, that the exploration of all dimensions of location-specific factors is beyond the scope of the present article. Further investigation therefore is warranted.

Moreover, the application of conventional multilevel modelling (or any other single-level regression, for that matter) assumes that independent variables are related to the dependent variable solely through its location parameter, typically the mean. This also means that the outcome (dependent) variable is affected by the predictors in a similar manner, both in terms of magnitude and direction, across the entire distribution. In addition, this assumption seems inappropriate, particularly when studying a phenomenon such as household wealth in South Africa, where wealth or income inequality is well established (Chatterjee et al. 2022; Fortuin, Grebe & Makoni 2022; Leibbrandt & Diaz Pabon 2021; Wittenberg 2017). This reflects the fact that the influence of household heads’ characteristics may vary across different levels of wealth distribution. For example, the association at a given wealth rank, i, may differ in sign, size and significance from that at wealth rank j, because of heterogeneity in the household wealth distribution and many other factors. Thus, rather than merely applying a traditional multilevel modelling approach, it is important to incorporate both multilevel and quantile regression aspects as a more comprehensive, quantitative approach.

Against this background, this study employs a linear quantile multilevel modelling approach (LQMM) to analyse the NIDS data in order to answer the two questions posed. Besides the fact of NIDS data clustering and the suitability of the research methods discussed precedingly, the quest for understanding the relationship between household headship and wealth is not limited to one group of countries. For instance, Vo and Ho (2023), Vo et al. (2023) and Wittenberg (2017) are but a few of the researchers investigating this relationship in developing as well as in developed country contexts. One reason for focusing on household headship is that some of its selected characteristics may hide important societal norms that can explain the phenomenon of household wealth accumulation in South Africa.

Thus, it is not surprising that this study focuses on household headship in South Africa, given the country’s historical context. The exclusion of the Black majority and women from access to economic opportunities can be attributed to the establishment and enforcement of the apartheid regime in South Africa, which may still have lingering effects in the present era. As a result, it is plausible that female-headed households may face greater challenges in wealth accumulation than their male-headed counterparts.

Based on the above-stated, this study hypothesises that women (or any other previously or currently disadvantaged groups) may have begun accumulating wealth under less favourable conditions because of historical exclusion. Similarly, ethnicity or race plays a role, as non-black African-headed households may benefit from historical privileges to which black African-headed households did not have access. Thus, the focus of this study is on these household heads’ characteristics: gender, ethnicity, education, age and marital status.

The structure of this study is as follows. The article begins with an introduction to the study, outlining the problem, main questions, objectives and significance. This is followed by a discussion of the previous studies on the topic and how they are related to this study. Next, the article describes the study’s methodology and data. The discussion of the main results and findings follows, concluding with recommendations..

This section discusses previous studies on the subject matter. Some aspects of these studies were already introduced earlier (e.g. the issue of methodology), so the focus here is on exploring additional aspects not yet covered. In this respect, research in South Africa focuses predominantly on income rather than wealth, particularly when it comes to micro-level analysis. However, there is a strand of literature that focuses on some aspects of wealth in South Africa. For instance, studies focusing on understanding wealth inequality and distribution have concluded that, even in post-apartheid South Africa, the country continues to experience high levels of wealth inequality (Chatterjee 2019; Chatterjee et al. 2020, 2022).

Other studies have focused on gender and race (ethnicity) to examine individual wealth gaps in South Africa (Casale & Oyenubi 2024; Chelwa et al. 2024). The study conducted by Casale and Oyenubi (2024) argues that understanding wealth at the individual level is crucial, as household-level aggregation can obscure important differences given that household members are not necessarily homogeneous. These findings confirm the existence of gender-based wealth disparities. Chelwa et al. (2024) compared racial wealth gaps in the United States of America (USA) and South Africa, given the similarities in their histories of discriminatory policies that prevented certain segments of the population in both countries from owning property and fully participating in the economy or led to the dispossession of their assets. These authors conclude that the Black-White wealth gap pattern in South Africa is similar to that in the USA and explain that these gaps exist at different levels of education and age of individuals.

As discussed in the introduction, wealth data discrepancies and measurement challenges at the micro-level have been prominent in earlier studies focusing on South Africa. Notably, the wealth data from NIDS, particularly in Wave 5, is reliable (Chelwa et al. 2024; Daniels & Khan 2019). However, despite the availability of this rich micro-level data, no study has yet investigated the relationships between household wealth and the characteristics of household heads in South Africa. The suggestion made by Casale and Oyenubi (2024) to assess individual wealth rather than household wealth is valid. Nevertheless, this idea presents challenges when considering the economic well-being of individuals. This notion of economic well-being should be understood within the context of the household, which is a functional unit, where members share the benefits of wealth, irrespective of who owns specific assets such as income or housing, as well as the effects of debts within the household.

Based on the preceding, one should note that focusing solely on individuals presents a risk of overlooking the critical aspect of household economic well-being and may lead to the assumption that resources are not pooled among household members living together. In addition to using households as the unit of analysis, this study incorporates individual wealth, as outlined in Casale and Oyenubi (2024), by including an independent variable called ‘Working Household Members’. This variable represents the number of household members who reported being employed or self-employed, and it is assumed that these members have the economic ability to possess wealth (discussed further in Section 3.3).

While there is a scarcity of studies examining the relationship between household wealth and headship in South Africa, a few related studies focus on this topic in other developing countries (Jin & Xie 2017; Vo & Ho 2023; Vo, Vo & Ho 2023). This study builds on their findings to address the gap in research specific to South Africa. In addition, as mentioned in the introduction, the methodology used here offers a unique contribution, particularly in modelling hierarchical (clustered) data such as that from the NIDS.

As already discussed in the introductory section, an LQMM approach is adopted in this study to answer the two research questions. This section outlines the reasoning behind the chosen econometric modelling approach, the estimation process employed and the data utilised for analysis.

Following the stepwise estimation approach outlined in Section 3.2, two regressions were estimated at each of the four selected quantiles: a null LQMM and a corresponding random-intercept LQMM, totalling eight regressions. However, because of space constraints, only the LQMM estimates at Q0.25, Q0.50, Q0.75 and Q0.85, along with the ICCs for the null LQMM models, are presented in Table 4. Including ICCs is essential, as they are applied to assess whether the LQMM approach is more suitable than single-level quantile regression at each predetermined quantile.

It can be seen that the reported ICCs in Table 4 exceed the conventional threshold of 0.05, indicating that 93%, 97%, 93% and 49% of the variation in household wealth at Q0.25, Q0.50, Q0.75 and Q0.85, respectively, are explained by district-specific latent factors. This confirms that the LQMM approach is more appropriate for the data.

Given the various LQMM estimates presented in Table 4, this study adopts a straightforward approach to assess each coefficient in terms of the three Ss: significance, sign and size, to facilitate interpretation. In this context, a few main observations arise. Before delving into the discussion of the results, it is important to again highlight the rationale for applying the LQMM in this study. The primary motivation was to account for within-district clustering while also capturing the distributional variation in ‘Household wealth’. A key advantage of this approach is its ability to provide a nuanced understanding of how predictors influence different points of the conditional distribution, beyond the mean.

However, it must be acknowledged that the LQMM, particularly as implemented using the R package proposed by Geraci (2014), does not inherently address potential endogeneity arising from omitted variables, measurement error or simultaneity. Although within-district centring of predictors helps mitigate group-level confounding and enhances interpretability, it does not fully resolve endogeneity concerns. Therefore, while the estimates presented in Table 4 offer valuable insights into heterogeneous effects across quantiles, they should be interpreted with caution in light of possible endogeneity.

Firstly, the random-intercept LQMM models at Q0.25 and Q0.5 show similar coefficient patterns based on the three Ss, indicating that the effects of independent variables on household wealth are comparable for both less wealthy households and those at median wealth levels in South Africa.

Secondly, overall, the estimates at Q0.75 and Q0.85 quantiles are not comparable to each other, nor to those at Q0.25 and Q0.5, based on the three Ss. For example, the coefficient for ‘Age’ is significantly larger at Q0.85, while ‘Gender’ is only significant at Q0.75. This suggests that, above the median value, household wealth in South Africa is influenced differently, offering valuable insights for a more differentiated approach to public policy design or intervention. It is also worth noting that, for statistically significant LQMM estimates, Table 4 reports larger coefficient magnitudes. This implies that changes in household heads’ characteristics have a greater impact on wealth for wealthier households in South Africa.

Thirdly, turning now to the independent variable of interest, the household head’s age and education level consistently explain household wealth, regardless of the household’s wealth status. Notably, the relationships between these two characteristics and household wealth are both positive, as expected. While LQMM modelling was not applied in some studies, it is worth noting that these findings are consistent with previous research (Vo & Ho 2023; Vo et al. 2023). Specifically, the results for age show that a 1% increase in the household head’s age is associated with a 0.37%, 0.37%, 0.37% and 0.42% increase in household wealth, on average, at Q0.25, Q0.5, Q0.75 and Q0.85, respectively, all other factors being held constant within a cluster. Regarding education, the results imply that the effects of a household head having completed at least secondary education on the logarithm of household wealth are 0.095, 0.095, 0.093 and 0.123 at Q0.25, Q0.5, Q0.75 and Q0.85, respectively.

The marital status of the household head positively explains household wealth only at quantiles lower than Q0.85. As discussed in Section 3, one possible explanation for this positive relationship is that married couples benefit from combining their sources of income, thereby accumulating more wealth than households with unmarried heads. However, this does not apply to wealthier households, as the coefficient for ‘Marital status’ at Q0.85 is not statistically significant.

In addition, the relationships between ‘Gender’ and ‘Gender*Education’ with ‘Household Wealth’ remain unconfirmed, except at Q0.75, where a positive association is observed in both cases. These results indicate that a male-headed household is linked to an increase of 0.011 logarithms in household wealth, while a male-headed household whose head has completed at least secondary education corresponds to an increase of 0.033 logarithms in household wealth within South African districts at Q0.75, holding all else constant. Although this trend is only significant at Q0.75, it highlights existing disparities in South Africa, particularly based on gender. As discussed in Section 3.3, it is evident that female-headed households encounter challenges in wealth accumulation, especially if the head lacks secondary education.

Fourthly, all control variables except for ‘Residence’ explain household wealth at all quantiles, as shown in Table 4. As anticipated, there is a negative impact on household wealth levels within districts when a household is predominantly black African, as indicated by all four LQMM models. This suggests that ethnicity is an additional dimension contributing to wealth disparities in South Africa. One possible explanation, as argued in this study, is the public policies of exclusion implemented by the apartheid regime, the lingering effects of which continue to disadvantage the black African majority in the country to accumulate wealth, as opposed to other races.

Moreover, ‘Household wealth’ is positively associated with ‘Working Household Members’, indicating that households with a relatively high number of working household members can accumulate more wealth than those with fewer working household members. This suggests that having more working household members provides the household with multiple sources of income, positioning it to accumulate wealth, as discussed in Section 3.3. In contrast, the number of household members represented by ‘Size’, regardless of their employment status, is negatively related to the level of household wealth.

After discussing the LQMM estimates, it is now time to assess whether the analysis conducted addresses the research questions posed. Regarding the first question, except for less wealthy households and those at the median level (Q0.25 and Q0.50), household wealth is generally affected by some household heads’ characteristics in terms of magnitude. For the second question, the reported ICCs reveal significant differences in the variability of household wealth distribution within districts compared to between districts. Therefore, through the application of the LQMM modelling approach, this study has attempted to answer the research questions.

This study used the LQMM modelling approach to examine whether household heads’ characteristics are associated with household wealth across South African districts and to explore differences in the variability of household wealth within and between districts. Rich data from Wave 5 of NIDS was utilised for this purpose. After implementing a rigorous estimation procedure, the analysis confirmed that certain household characteristics are indeed related to household wealth, as expected. In addition, it was found that these relationships vary in magnitude between wealthier and less wealthy households, and significant differences in wealth variability were observed within and between districts in South Africa.

The findings of this study highlight the need for targeted policies to promote household wealth accumulation in South Africa. For example, the Reconstruction and Development Programme’s (RDP) housing initiative should be strengthened and localised at the district level. In the long run, such a programme can enable households to accumulate wealth, as homeownership represents a significant asset that can also serve as collateral for acquiring other types of assets. Similarly, education policies and instruments, such as the National Student Financial Aid Scheme (NSFAS),¹ require reinforcement, given this study’s evidence that a household head’s educational attainment is closely linked to household wealth accumulation. Land redistribution also remains crucial, not only for agricultural development but also as a key pathway for asset accumulation among previously disadvantaged households and communities.²