Measuring wealth inequality in South Africa: An agenda

ABSTRACT

Understanding wealth inequality has unique significance in South Africa. The co-existence of extreme poverty and extreme wealth is starkly visible. Apartheid-era inequality has persisted despite more than 20 years of democracy. Much of the recent research focus on inequality has been on inequality of income and of opportunities, especially quantitatively. With the recent project to release South African tax administrative data for research, this paper hopes to show how use of the tax administrative data can contribute to developing a refreshed quantitative analysis of wealth inequality, especially in estimating the top shares of the wealth distribution, and so contribute to the existing literature on wealth inequality in South Africa. The first section will explore why studying wealth inequality is of fundamental importance. The second section will review international data and methods used to research wealth inequality, before laying out suggested approach to doing such studies in South Africa.

KEYWORDS:

• Wealth inequality
• scoping
• administrative data

JEL CODES:

• C80
• D31
• D63
• O55

1. Introduction

Understanding wealth inequality has unique significance in South Africa where the co-existence of extreme poverty and extreme wealth is starkly visible. Orthofer (2016) estimates that the top 10 of the population own approximately 95% of all wealth while 80% of the population own no wealth at all. Apartheid-era inequality has persisted despite more than 20 years of democracy. Historically, there has been pioneering work both on measuring wealth distribution and analysing methods of wealth accumulation, especially during the apartheid era (e.g. see Wolpe (1972); McGrath (1982)). Much of the more recent research on inequality has shifted focus to inequality of income and of opportunities. The purpose and scope of this paper is modest: With the recent project to release and use South African tax administrative data for research, this paper hopes to show how tax data can be used to study levels of wealth inequality, how wealth is held, and the mechanisms through which wealth inequality is produced and reproduced. In doing so, this paper aims to firstly motivate for further release and use of administrative data, and second to show how studies using tax data can bring new quantitative perspective and uniquely contribute to the large body of work on wealth inequality in South Africa. The first section will explore why measuring wealth inequality is important to understand overall inequality. The second section will review international data and methods used to research wealth inequality in other countries, before laying out suggested approaches to doing such studies in South Africa.

The concept of asset-based inequality can be broadened. For example, there is stream of inequality literature in economics that that describes capabilities as assets (Sen, 1999). This approach is extremely important, but for the purposes of this paper, wealth inequality will be defined as the unequal distribution of household (money-metric) assets.

2. Why is studying wealth inequality important?

Wealth inequality is largely unexamined in the study of economic inequality, despite it being a stylised fact that it is more concentrated than income inequality. Income inequality only provides a partial understanding of overall inequality, and is unable to adequately explain persisting and deepening inequalities.

In contemporary economics, the theory of income inequality is organised around the labour market, where, as Stiglitz characterises, income inequality is the result of differing ‘returns’ to employing a worker, i.e. the productivity of a worker (Stiglitz, 2015) . This in turns reflects the capability of a worker, and so the labour market fairly distributes income to workers according to the contribution they make. The market failure in this case is that not everyone has access to same schooling, health care, and other components that allow fair competition in the labour market, and hence income inequality becomes a reflection of the distribution of these services. Indeed, Corak (2013) demonstrates that income inequality and inequality of opportunity are indelibly linked, severely hampering socio-economic (upward) mobility.

However, there are three shortcomings focusing solely on income inequality. The first is that Mincerian analyses that try to isolate the individuals’ determinants of earnings (i.e. human capital) cannot explain the high levels of income inequality, let alone overall inequality. Similar individuals receive quite different earnings, while seemingly irrelevant personal characteristics, including beauty and height, are often robust predictors of earnings (Bowles & Gintis, 2001). Further, contrary to anti-discrimination law, earnings are often driven by social determinants, such as race and gender (see Elson, 1999; Hinks, 2002; Ntuli, 2007; Kim, 2009). This implies that there is some other mechanism through which preferential access to the labour market is attained, such as wealth, but also social and cultural capital. Sociological research indicates that these elements play a key role in maintaining wealth concentration among elites (see Khan, 2010; Savage, 2015; Rivera, 2016). If it is differences in wages that causes wealth disparities, then variance in individual traits or meritorious ability fails to explain the massive disparities in private wealth. The second shortcoming is that the prevalence in non-labour income at the top end of the income distribution can’t be explained without analysing wealth. Non-earned income is almost exclusively at the top end of the income distribution (Lydall & Tipping, 1961), meaning the polarisation in income inequality is driven by ownership of assets, rather than labour market participation. Understanding wealth inequality is essential to understand the income distribution. The third is that inequality of opportunity, which drives income inequality, is driven both by income inequality (Macinko et al., 2003; Lynch et al., 2004; Subramanian & Kawachi, 2004), and wealth inequality (Ferreira, 2001; Zimmer, 2008; Nowatzki, 2012). This implies a significant role for wealth in explaining other forms of inequality.

There are important reasons to study wealth inequality in and of itself. Wealth is a stock variable, meaning that it is a quantity of money, stored in different ways, that is accumulated over time by inflows and/or depleted by outflows, and is transferable. In times of economic precarity, wealth allows consumption-soothing and self-insurance. ‘As households are exposed to increasing levels of risk, success in building personal assets is becoming increasingly important’(Davies, 2009). Hence understanding wealth inequality is important from a household’s economic welfare perspective.

Wealth inequality also impacts economic performance. Piketty (1997) and Ghatak et al. (2002) demonstrate how wealth inequality affects credit rationing, influencing capital allocation and overall investment. Higher wealth concentration results in investment in financial products, meaning capital is tied up in financial products rather than the real productive economy (Stiglitz, 2015), depressing economic growth and hampering the creation of decent jobs. Bagchi & Svejnar (2015) use the Forbes listing of billionaires to uncover a negative relationship between wealth inequality and growth, especially where wealth is acquired through political connections. Wealth inequality also has an impact on productivity, public good provision and occupational choice (Banerjee & Newman, 1993; Bardhan et al., 2000; Bardhan et al., 2007)

Wealth inequality also affects society more generally. ‘Wealth … brings empowerment … to enforce your rights, intimidate others, influence politics. Limits to power of the wealthy are less severe than those on the power of the poor or middle class’(Davies, 2009), demonstrating how increasing wealth concentration undermines the democratic representation of all parts of societies in favour of narrow interests. Indeed, among the narratives of the Great Financial Recession were that the influence of the rich allowed financial excess to spiral (Stiglitz, 2012), leading not only to the crash but also government responses to the crash, generally cutting spending on social policies to manage the fiscus and pay for quantitative easing programmes, exacerbating hardships of the non-rich. High-wealth concentration describes the presence of a group who have disproportionate control over, or access to, resources – otherwise defined as economic elites (Khan, 2010). Elites

secure political and administrative connections in order to maximize their profits, … develop exclusionary practices in higher education in order to preserve their privileged access to top educational credentials, … reproduce their privileges through elite lifestyles or, among other possible examples, … convert their economic capital into other forms of capital. (Jodhka & Naudet, 2017)

Wealth inequality then provides a lens into inequality in other spheres of inequalities, as wealth becomes an effective predictor of health, educational and labour market outcomes (Killewald et al., 2017). Understanding (and addressing) wealth inequality is crucial to maintaining an inclusive and stable society.

Wealth inequality also represents social injustice. Wealth is built by inflows, which depends on the effectiveness of creating savings, investment income and capital receipts. Those with a higher starting level of wealth are able to build wealth quicker than those with lower or no wealth to begin with, leading to increasing wealth inequality. In South Africa and elsewhere, the starting levels of wealth have been configured by dispossession and discrimination (Conley, 1999; Terreblanche, 2002). This, together with gendered and racial labour market discrimination, has supported the wealth accumulation for certain members of society. Piketty (2014) demonstrates that in periods when the return to wealth (what he defines as r) exceeds the overall growth rate, the gap between those that earn predominantly through wealth, which in many countries is inherited, rather than newly created, and those that earn through participation in the labour market will widen, and will continue to widen substantially. This role of inherited wealth is demonstrated in the UK (Atkinson, 2018), where transmitted wealth (expressed as a percentage of national income) rose from under 5 per cent in the 1970s to around 8 per cent in 2006, equivalent to the proportion of pensions and annuities in total gross household income. Various studies of the USA show that intergenerational transfers are significant, if not more important in wealth accumulation than life-cycle savings (Kotlikoff & Summers, 1981; Gale & Scholz, 1994). Several studies on the Scandinavian experience also show the importance of hereditary wealth (Black et al., 2015; Boserup et al., 2016; Adermon et al., 2018). The complexity and diversity of inheritance rules in the global South also varies the impact of intergenerational asset ownership, especially in relation to gender distribution (Deere & Doss, 2008).

Given the historical role of dispossession and discrimination in capital and labour markets, and intergenerational transfer of wealth (through inheritance), wealth inequality therefore ‘captures the historical legacy of low wages, personal and organizational discrimination, and institutionalized racism’ (Oliver & Shapiro, 2013). In the US financial inheritances may account for between 10% and 20% of the average difference in black–white household wealth (Menchik & Jianakoplos, 1997). While most of the focus lies in the role of inheritance in reproducing wealth at the top end, Oliver & Shapiro (2013:5) argue that ‘The effect of this inherited poverty and economic scarcity [for African Americans] for the accumulation of wealth has been to “sediment” inequality into the social structure’.

Applied to South Africa, wealth inequality is a strong indicator of Apartheid-era injustice and inequalities perpetuated into the present. Capital accumulation strategies were supported by the apartheid’s state support of, on the one hand, concentration and centralisation of white-owned capital, and on the other hand, a state regulated and coercive labour market (i.e. through the creation of an industrial reserve army in the Bantustans) (Legassick & Wolpe, 1976; O’Meara, 1996), and so wealth inequality was deliberately structured along race (and gender) lines. These wealth accumulation strategies are at the root of current inequalities in health (Coovadia et al., 2009), and education (van der & Berg, 2007; Chisholm, 2012), just to name two other spheres, and have, in turn, perpetuated the wealth inequalities, as it has been contended, policies and economic activities benefited those who had gained from apartheid (through ownership of assets and access to education).¹ It is on the basis of these historical processes of exploitative wealth accumulation that Terreblanche suggests a wealth tax to provide a ‘satisfactory degree of systemic justice’ (Terreblanche, 2018:19). Developing a wealth distribution for South Africa, and linking other data sources to it, can then help quantify the historical impact on present inequalities, and provide a different lens with which to analyse education, health and other spheres of inequalities.

More studies, especially with new sources of data, are crucial for current policy-making on reducing wealth inequality. First, equity is one of the core principles in tax policy. However, historically, the

structure of taxation … discriminated against income and in favour of wealth, wealth acquisitions, and capital gains. This benefited those people who could switch back and forth between income, wealth and capital gains to reduce their tax liabilities and penalised others, largely wage earners, who could not. (Harbury & Hitchins, 1979)

Hence evaluating tax policy in the presence of better understanding of wealth is important. Second, a policy response to wealth inequality in South Africa has been to propose a ‘wealth tax’. The viability of such a tax was investigated by the Davis Tax Commission (The Davis Tax Committee, 2018), which concluded that the quality of existing data on wealth holding needed to be improved to decide if a wealth tax is the most appropriate policy in light of how wealth is held. To do this, the following is required:

• Develop a more accurate wealth distribution that includes information on its components of wealth

• Analyse how wealth concentration is influenced by intergenerational transfers

• Study how wealth inequality affects income inequality and labour market outcomes

3. Wealth distribution – How has it been created globally?

This section will summarise the approaches taken to studying wealth inequality globally and will be structured as follows: (1) definition of wealth; (2) wealth distribution data and methods (top shares, non-top shares, combined); and (3) Analysis of the distribution.

3.1. Definition of wealth

Wealth is broadly defined as non-financial and financial assets over which ownership rights can be enforced and that provide economic benefits to their owners. As per the international standards set in the System of National Accounts (United Nations, 2009), these include tangible assets (real estate and consumer durables), fixed claim assets (cash, deposits, etc.), corporate equities, equity in unincorporated businesses (farms, small businesses), and other various miscellaneous assets. Valuations of these need to be considered carefully (see Davies (2008) for an overview of definitions and considerations)

Researchers can consider an expanded definition of wealth, particularly in the Global South, where rural livelihoods are governed by non-market institutions. For example, cattle is a store of wealth in sub-Saharan Africa (Jarvis, 1980; Turner, 2004; Stroebel et al., 2008), while in parts of India, trees are grown separately as a form of insurance to guard against the risk of market participation in cash crop agriculture (Ravindran & Thomas, 2000). There are issues to consider when bringing in a wider definition of wealth, relating to availability of data, valuation and conceptual relevance. This is discussed in Appendix 1.

3.2. Creating a wealth distribution

A wealth distribution simply describes how the total wealth (defined above) is divided among the population, or what proportion of total wealth is held by what proportion of the population, ordered from least (e.g. bottom 1%) to the most (e.g. the top 0.1%). Broadly speaking, the availability of data sources influences the method used to estimate a wealth distribution. Administrative data is used to estimate top shares (e.g. the top 2%), while household surveys used to estimate bottom shares (e.g. the bottom 98%). These varied sources are combined to create a functional distribution, which is then applied to the National Accounts record of aggregate wealth.

3.2.1. Top shares – Estate duty method

Estate duty is a tax paid on the estate (money and property) of a deceased person. The term ‘estate duty’ is interchangeably used with the term ‘inheritance tax’, though there is an important difference – estate duty is determined and applied on the assets of the deceased, whereas inheritance tax is assessed on beneficiaries’ share of the assets. The estate duty method was first used in 1908 (Mallet, 1908), as estate duty was the first and for a long time the only source of tax revenue and administrative data collected that revealed a person’s total assets and liabilities.² The method essentially uses the dead as a sample of the living. This is done by taking the value of assets recorded in the estate duty records (the dead), and by multiplying it by (the inverse of) the mortality rate, deriving an estimate of what the value of assets of those living would be. This is then compared to the total personal sector wealth (from an external source, such as the national accounts) and population figures to assign these observations to a position in the upper end of the distribution. The estimates are sensitive to missing wealth and the selection of the mortality rates. Missing data come from three areas: (1) Under reporting; (2) Tax evasion and avoidance (illegal and legal ways of minimising tax); (3) under valuation of assets. Using alternative sources of data can help fix the estimates. For example, the National Accounts record national household wealth by assets. The estimate from the estate duty method can be ‘grossed up’ to match the National Accounts totals, with the missing data assigned to different parts of the distribution using various statistical methods. Tax evasion and avoidance however are harder to track, given the transnational nature and expertise in moving wealth to offshore locations or keeping then other types of tax entities (Zucman, 2014). However, building a wealth distribution without accounting for these is still useful as the bias almost certainly understates the concentration of wealth at the top end, so would not change the structure of the wealth distribution. It would also provide practical information about a tax base that is accessible. See Appendix 2 for further technical explanation.

3.2.2. Top shares – Income capitalisation method

The investment capitalisation method also has a long history, with early references to its methodology appearing in 1913. Simply put, it applies a yield multiplier to the distribution of investment income recorded for each taxpayer to calculate the asset base on which the income was earned.

The choice to perform this method is highly dependent on how strong the tax system is on enforcing taxpayers to submit information on investment income, if financial assets make up a significant portion of total wealth, and the level of disaggregated information available. The methods to fit it to a distribution, and issues around missing wealth and valuations are the same as discussed in the previous section. See Appendix 3 for further technical explanation.

3.2.3 Top shares – Rich lists

Rich lists are lists of large wealth-holders compiled globally by Forbes magazine (though nationally other sources exist, for example in the UK, Sunday Times also compile a rich list for the UK). They provide information to examine the top end of the wealth distribution. However, there are some concerns in using this information. First, it is compiled using interviews from a range of people linked with the billionaires, but the numbers aren’t ‘easy to validate’ (Alvaredo et al., 2016). Additionally, it reports the wealth to hundred millions USD. Therefore, many individuals share the same rank causing a discontinuous ranking. Nevertheless, this source of information does provide useful information to inform missing and understated wealth at the top end of the distribution.

3.2.4. Non-top shares – Sample and household surveys

Household surveys offer a different perspective to the wealth distribution. These surveys furnish information about pension holdings and savings for other methods, but also provide information about those not in the tax system. Surveys to record wealth holdings have been less frequent, and subject to statistical concerns, making it less useful to develop an accurate and continuous distribution. There are four main concerns. The first is the relatively low response rate, leading to underrepresentation from upper wealth groups. As the majority of wealth studies have shown, the top end of the distribution is where wealth is concentrated and nature of wealth can be studied. The second is that of incomplete information, and high potential for understatement in terms of both completeness and valuation. Third, incomplete coverage in survey design means that some types of assets are excluded, and so the definition of net worth is not comprehensive. (More recent surveys may be more comprehensive, though this creates continuity issues). The fourth is sampling error, which becomes more amplified at the top end of the distribution given the fewer numbers of the wealthy. More recent surveys have tried to rectify some of these problems, with limited success.³ Despite the unsatisfactory information about the upper tail of the wealth distribution, surveys play a very important role, either to reconcile other estimates, or to combine with other methods. This helps provide a more complete picture of the wealth distribution (Atkinson & Harrison, 1974; Alvaredo et al., 2016).

3.2.5. Combined methodologies

Combined methods use the three sources described above to piece together a wealth distribution over time, and use sophisticated techniques to provide continuity. One of the most comprehensive studies to do this is that of Garbinti et al. (2017), who combine income tax data, inheritance registers, national accounts and wealth surveys to create a consistent, unified wealth distribution series by percentiles for France over the 1800–2014 period. They use estate-multiplier method from 1800 to 1970 period and link up this with a new series for the 1970–2014 constructed using a mixture of income capitalisation and survey-based method. This study provides valuable techniques to apply: how to use wealth surveys and Pareto adjustments using billionaire rankings to supplement other methods; where fiscal sources don’t exist, how to develop flexible, non-parametric generalised Pareto interpolation methods (details are in Appendix 4).

3.3. Analysing wealth inequality

3.3.1. Intergenerational mobility

In studying the wealth distribution, we also aim to understand how such inequality is produced and reproduced. A fairly simplistic, but important, answer is that ‘wealth inequality may be driven by differences in saving behaviour, or in the intergenerational transfers received’ (Cagetti & De Nardi, 2008). This can be broadened to find out which is more prevalent: newly created wealth (which can include savings) or hereditary wealth. In relation to the latter, we can investigate the various outcomes for those with inherited wealth, and test if they are significantly different from those without inherited wealth.

Household surveys that have information about sources of income and assets are one source for these studies. Linking various tax datasets can also help to give better information –for example matching income tax data to gift and estate tax data could shed light on the fraction of wealth coming from inheritances (as opposed to self-made) (Saez & Zucman, 2014). The link allowed Joulfaian (1994) to make some powerful findings: the average inheritance is approximately three times that of the heir’s income (child, as opposed to spouse) and wealthy parents are more likely to have higher-income children.

The Scandinavian analyses (Black et al., 2015; Adermon et al., 2018; Boserup et al., 2016) benefits significantly from registration information available that link children to parents (more information on the data and research methods is in Appendix 5).

3.3.2. Calibrated models of wealth distributions

Following the creation of wealth distributions, quantitative models of wealth inequality can help to understand through what channels such wealth distributions are created. In this way, the models help understand what policy makers should target to change the wealth distribution. For example, Benhabib et al. (2011) demonstrate that ‘when idiosyncratic rates of return across generations are a significant source of wealth inequality, reducing estate taxes, or … capital income taxes, can significantly increase wealth inequality in the top tail of the distribution of wealth’.

3.3.3. Other

There have also been other connected analyses related to wealth inequality and distributions. Studies that look at the link between wealth inequality and policies include analysing the potential for a wealth tax (Londoño-Vélez & Ávila, 2018), investigating the link between monetary policy and household wealth (Domanski et al., 2016), and looking at the impact of land reform policies (Assunção, 2006). Housing, a particularly important component of wealth is especially interesting given its contested role in property rights, wealth, spatial and overall inequality especially in transition and developing countries.(There are many studies that investigate this, including: Yemtsov (2007); Power (2012); Christophers (2018); Fuller et al. (2019)).

4. Wealth inequality in South Africa

Wealth inequality research has not been comprehensively approached, and it is the hope of this paper to motivate for such. However, there has been some important research in this area, which provide important context.

One of the earliest studies on South African wealth inequality uses primarily the estate duty method (McGrath, 1982). Mcgrath uses estate records drawn the Natal Supreme Court in Pietermaritzburg for every estate lodged in 1975. He notes that 75% of the estates accounts were White, Asians making up 21% with only 2.6% from the Coloured category and 1.5% from the Black African category. The Black African category is too small to be representative. The mortality multiplier applied is specific to the age, gender and racial group of the deceased. Mcgrath uses the South African Life Tables 1967–71, where black/Africans were excluded as they were not included in the vital registration system, and were also deemed to be citizens of the independent homelands (Dorrington et al., 2004). For the estimates at this time, this does not cause a problem, given that black/African sample was too small, and largely legally prevented from owning wealth. However, actuarial analyses show that these life tables demonstrate an underestimation of mortality rates for the white population and an irregular pattern for the coloured population (Bah, 1998). Given the sensitivity using this method to the mortality multipliers, these estimates of wealth require testing with more refined mortality multipliers. Following the estimation of wealth, information in the estate duty records allowed the wealth holdings to be decomposed by occupation as well as race, gender and age groups.

Orthofer (2016) uses the National Income Dynamics Study (NIDS), a household survey, and a sample of personal income tax records, to estimate a combined wealth distribution, while adjusting the distribution to take into account the totals in the national accounts. The NIDS data capture information about the households non-financial assets and mortgages, while also ascertaining these details for each household member. The personal income tax data are sample from the self-assessed income tax records, which captures non-labour income, and so the assets that generate the taxable incomes. Orthofer follows the capitalisation technique using average investment returns for each asset class. The data provided in the sample is at a high level of aggregation, specifically, local interest, other investment, foreign interest and foreign investment. Orthofer uses averaged returns in these very broad categories, and so the estimate may be sensitive to the capitalisation multiplier. Given the data limitations, this study is crucial in starting the exploration of how to use the various datasets available to estimate a wealth distribution.

Mbewe & Woolard (2016) explore two waves of the NIDS survey to examine the cross-sectional distribution of wealth in South Africa. Having created a net wealth variable for each households, the NIDS survey allows for negative wealth, which is often missed in other data sources. However, there doesn’t seem to be any analysis on how representative the top tail is, given the likely bias stated above, and whether corrective actions were taken in sampling. This makes understanding the top shares challenging. However, this provides a crucial source of information for the rest of the wealth distribution.

Once the functional distribution is created using these methods, this must be applied to consistent categories of wealth in the National Accounts’ household balance sheet. The development of this data source is promising for this exercise (Aron et al., 2008; de Beer et al., 2016).

5. Research agenda for South Africa

The recent project to use administrative tax data from South Africa (Ebrahim & Axelson, 2019) can ignite new research into measuring wealth inequality:

5.1. Data organisation

5.1.1. Estate duty

Estimating top shares using the estate duty methodology requires accessing estate duty information from the South African Revenue Service. To start building a historical series, archived estate duty information needs to be investigated.

5.1.2. Investment income capitalisation

This data sits within SARS income tax dataset. To start building a historical homogenous series, archived income tax information needs to be investigated.

5.1.3. Linking estate duty to income tax data

This section is focussed on creating a database to investigate hereditary wealth and its impact on labour market incomes.

It should be noted that inherited wealth is not taxed, as these are assets (so not subject to income tax), and capital gains is settled by the estate, rather than the inheritor. But transfers of property from a deceased estate to an heir or legatee entitle the estate to capital gains rollover, which could provide a linking mechanism.

A second link to develop between the datasets comes through donations tax data. Information from here can be used to link taxpayers in the IRP5 dataset. Information of incomes from trusts, donations, and gifts may offer some understanding of hereditary wealth impact on labour market participation. The likely irregularity of donations and gifts may not make it suitable for analyses, though this route can still be investigated

5.1.4. Household survey and sample data

The primary wealth data in household survey comes from NIDS. This data source also provides information on hereditary links. Other useful survey data to bring in comes from Stats SA: Living Conditions Survey (LCS), General Household Survey (GHS), and the Income and Expenditure Survey (IES). Other surveys from private sector financial services companies – such as the Momentum/Unisa Household Financial Wellness Index surveys – and other organisations that focus on this area, (e.g. The Association for Savings and Investment South Africa and Eighty20), can provide supplementary data to check on financial assets at the top end of the distribution.

5.2. Wealth distribution

To accurately estimate the wealth distribution, both methods (estate duty and investment income capitalisation) are required to estimate the top shares In addition, developing accurate mortality and capitalisation multipliers would be extremely important. Combining household data for the bottom shares with estimates from the top shares would then complete the creation of the distribution. The functional distribution then can be applied to the national wealth estimates from household balance sheet data from the National Accounts. Information on the top shares can be supplemented with information from Forbes’ Rich List.

5.3. Analysing wealth inequality

There are considerations specific to estimating a wealth distribution in South Africa, and quantitative analyses of wealth inequality more generally. First, more detailed understanding of the distribution is required. This includes: overall trends, including measures of central tendency and composition, decomposing debt and assets (crucial to understanding how we distinguish cases of low net wealth, where asset ownership is debt funded, versus low asset ownership), stratification of asset distribution (i.e. by race, gender, location), and the application of different measures of inequality, as discussed in Cowell (2000).

Second, ‘missing’ or unaccounted-for wealth, though a consistent issue in developing a wealth distribution in all countries, must be given special attention in the South African context as a specific phenomenon (e.g. see Ashman et al. (2011)). There is prevalence for both legal and illegal strategies to reduce the values of assets recognised domestically, including offshoring, and tax-base erosion to reclassify wealth through multinational arrangements (Wier & Reynolds, 2018). This requires a separate focus that should include journalistic data sources such as the Paradise Papers. In light of South Africa’s unique history, analysing ownership of natural capital is important (Lange et al., 2018).

Third, following completion of the wealth distribution, Calibrated Models of Wealth Inequality can investigate the channels that produce outcomes that match the extreme levels of wealth inequality in South Africa. Given South Africa’s unique history, this would require models to go beyond standard savings-based models and incorporate hereditary modes of wealth transmission. The models should then assist develop an understanding of policy options and potential benefits and costs.

Finally, the relation to wealth to other economic and social phenomenon should be analysed. Studies on intergenerational mobility should explore the impact of inherited wealth on intergenerational wealth and wealth mobility, income and labour market participation. Labour market participation investigations can include, but not limited to, the impact of (hereditary) wealth on employability, income, job duration, labour market progression, business income and other relevant variables. The variation of wealth along the distribution can be investigated in relation to macro and micro economic trends (e.g. exchange and interest rates, growth, regulatory changes, etc). The link between wealth concentration and other spheres of inequality, such as health and health systems, education, and security, could open up new perspectives on these issues.

5.4. Policy analysis

Each part of the work project provides important information in developing policy to address the high levels of wealth inequality. Research can focus on:

• Which policy instruments are available to influence each component of wealth (not limited to tax policy)?

• Through what mechanisms and channels the policy would work

• What are the impacts, both positive and negative, of the policy interventions?

• What are the institutional and administrative requirements needed for successful policy implementation?

• What are the risks and how can the risks be mitigated?

6. Conclusion

Key to understanding and addressing the high wealth disparities in South Africa is to build up data sources on the distribution of wealth. This will help us locate how much wealth is held, how it is held, and how it is passed on to maintain the patterns of inequality. I have reviewed methods used worldwide, as well as in South Africa, to suggest that both the data and the methods are available and applicable. After building a distribution, further analysis can be done to understand the type of interventions that work. However, the measurement of wealth is only the first important step. A multidisciplinary approach will can use the quantitative information to isolate further areas of investigation, and provide a richer understanding of the how wealth inequality is produced and reproduced. This multidisciplinary approach can provide the evidence to develop policy aimed at redressing wealth inequality.

Acknowledgements

The study was originally commissioned under the UNU-WIDER project, Southern Africa – Towards Inclusive Economic Development (SA-TIED). I gratefully acknowledge guidance, comments and suggestions provided by Imraan Valodia, David Frances, Kim Jurgensen, Murray Leibbrandt and Amina Ebrahim.

Disclosure statement

No potential conflict of interest was reported by the author.

Additional information

Funding

This work was supported by United Nations University World Institute for Development Economics Research [contract number 605UU-2402].

Notes

1 The purpose of this paper is to locate the relevance of quantitative analyses using tax administrative data both to these historical processes and their current manifestations, and so it is beyond the scope to provide a review of this large literature. For historical accumulation processes, see for example Wolpe (1972), O’Meara (1996), Kaplan (1977). For debates around how democratic-era policies exacerbated historical inequalities, see for example Streak (2004); Leibbrandt et al. (2011), Seekings & Nattrass (2008).

2 This is described in Atkinson and Harrison, 1978. The modern progressive income tax was not created until around 1913 in most countries (Piketty et al., 2006)

3 The Wealth and Asset Survey in the UK was launched in 2006, and used tax data to identify wealthy addresses. These addresses were oversampled. In the case of the ONS survey, the response rate didn’t improve significantly, and incomplete responses, especially about business assets, also contributed to concern about the upper wealth ranges specifically. In the French household finance and consumption survey, the improvements have been meaningful, though ‘its sample size is still too small to go beyond the 99th percentile’(Garbinti et al., 2017).

Appendices

Appendix 1. Issues to consider if using broader definition of wealth

1. Availability of data: While some household surveys do include cattle as part of household assets (e.g. National Income Dynamics Survey in South Africa), other forms of household assets that are specific to a localised culture are unlikely to be captured (e.g. the NSS All-India Debt and Investment Survey Data include livestock, but not trees)

2. Valuation: These alternative stores of wealth play a conceptually broad socio-economic role, such as social currency, source of subsistence, provider of power, security (Ravindran & Thomas, 2000; Turner, 2004). A market valuation of these items would inaccurately reduce them to a single-role commodity in market exchange. i.e. a value de-linked to the value placed on them by the society.

3. Relevance and comparability: Although inclusion of these items provide a more complete picture, it is unlikely that ownership of these assets are driving the concentration of wealth. Indeed including these items would also make inter-country comparisons more challenging.

Appendix 2. Estate duty estimation method

The principle of the method is as follows:

Assuming that those of a particular age and sex dying in a given year are representative of the living population, the overall distribution may be obtained by blowing up the estate data by a mortality multiplier equal to the reciprocal of the mortality rate. (Atkinson & Harrison, 1978)

Thus the key bits of information for this method are: (1) Data for Estate values (2) mortality rates.

Data from revenue authorities have historically been summaries of estate values by assets type by age group. Asset type information helps with missing data (dealt with below). The most recent studies adapt these earlier methods to take advantage of the release of micro-data from tax administration records. Micro-data in this context are the actual estate duty tax returns of the entire taxable population, or of a sample. Studies that build a longitudinal wealth distribution series use mixed data types: Piketty et al. (2006) use the estate duty method to estimate wealth concentration in Paris and France from 1807 to 1994, using a series constructed from samples of archived individual estate tax returns for the years 1807–1902, and tabulations compiled by the French tax administration in the years after that. Kopczuk & Saez (2004) use the estate duty method to estimate the top wealth shares in the US from 1916, also using mixed formats of estate tax information. Kopczuk and Saez only analyse the top 2% of the wealth distribution, as, due to large exemption levels, only a small fraction of decedents were required to file estate tax returns.

Where there are only high level tabulations, estate information provided are cross-tabulations by size of gross estate and age groups. For each age group and gender cell, the estate multiplier is the product of the average mortality for the cell and the social differential mortality factor. The authors multiply the number of decedents and the amount of gross estate reported by the estate multiplier, obtaining the distribution by gross estate brackets for the living population. There is a separate adjustment for the multiplier in the top bracket, given the small number of observations. The Pareto distribution is then applied to estimate the thresholds and amounts corresponding to each fractile (e.g. top 2%, top 1%, … , top 0.01% thresholds). Shares of wealth are calculated by dividing the wealth amounts accruing to each group by total net-worth of the household sector in the United States. For years where the top % information is not covered (in the earlier years 1916–1945), those shares are calculated using a Pareto extrapolation.

Where there are samples of returns, coverage of the largest estates (in the top 0.01%) is 100%. Any estates below the filing thresholds were ignored, given that not all estates below the threshold file their estates, and so there is no way to tell what proportion of estates are filed. The Inland Revenue Service take the samples during the processing of returns itself, which are stratified through three variables: year of death, age and size of gross estate. While all returns above a certain high level of wealth are included in the data, those below that level are sampled using complex methods by the IRS, including Poisson probability sampling method (during the processing of returns) and Neyman allocation scheme to use an estimate of the living population to determine the strata allocation for sampling (Woodburn & Johnson, 1994). Thus estimation of the distribution using this method will require involvement of the South African Revenue Service from the start.

Shares of wealth are calculated by dividing the wealth amounts accruing to each group by total net-worth of the household sector in the United States. The trend in the overall wealth shares is then analysed in terms of its composition. Wealth is divided into six categories: (1) real estate, (2) bonds (federal and local, corporate and foreign) (3) corporate stock, (4) deposits and saving accounts, cash, and notes, (5) other assets (including mainly equity in non-corporate businesses), (6) all debts and liabilities.

There are two areas of this approach that are source of bias: missing wealth, and the multiplier. Missing wealth can be split into two – smaller estates and incomplete wealth from the larger estates. Estate data only captures large estates that are registered to the revenue services, and so smaller wealth holdings, including for example savings, are omitted (for example, Atkinson states that in 1970 Inland Revenue statistics, there were 2 90 000 estates comparted to a 6 39 000 deaths). However, this part of the distribution is better represented through household surveys, which is discussed below. For the larger estates, missing data, missing items, tax evasion and avoidance, and undervaluation of assets may provide another source of bias. To deal with missing data, Atkinson & Harrison (1978) use the National Accounts information on asset type to build more reliable totals, and then allocate the difference to the distribution (this method is developed for more recent studies using investment income method, discussed later). Missing items relate to stores of wealth that are not required for estate duty collection. Life insurance, of particular importance at the top end of the distribution, need to be extrapolated from other data sources. With regards to tax evasion and avoidance, Kopczuk and Saez evaluate studies on tax evasion to determine this does not provide a major risk to their study. This is unlikely to be the case in countries that do not have such an authoritative tax collection system as the IRS. This is a challenging area, as much personal wealth is stored in different tax entities outside the scope of these studies, namely in companies and trusts and in tax havens, rather than in a personal or household capacity. Other valuation concerns relate to non-transferable stores of wealth, for which there are no market prices, such as pensions. Valuation can occur either on a cash surrender value (i.e. before maturity of the policy) or the full payout, though here they find the results are robust according to either valuation method. Pension information is provided according to the cash surrender value, with realisation value potentially only impactful for deaths under the ‘pension age’.

The second area of potential bias, the mortality multiplier questions how good death is in sampling the living (of the wealthy). The two main concerns in with multipliers are: (a) mortality multipliers tend to understate the number of wealthy persons and their total wealth, producing errors in the size of the wealth distribution; and (b) the deceased may not be a representative sample of the population, leading to bias in the estimate (Lyons, 1975). Initially, a general multiplier was used on estates below a certain amount, and an occupational and/or social class mortality rate on estates above that amount. Lampman (1962) used mortality rates from life assurance companies. A sensitivity analysis of the two types of rates showed that there was small effect on the cumulative shares of total wealth held by a given percentage of the population, but the effect on the absolute numbers in each range was significant (Anthony Barnes Atkinson & Harrison, 1974). Even when looking at shares, where there is a difference of approximately 1% of total wealth in the top 1%, this could be vital in understanding the breakdown of that share. Kopczuk & Saez (2004) use baseline mortality rates (from the Human Mortality Database), which decompose death rates by year, age and gender. They then adjust this using the mortality differential for white college graduates (by gender), relative to the average population and are assumed constant over the whole period. Another source of mortality rates are the actuarial mortality tables. However, they are constructed differently, and due to changes in the customer base of insurance companies with the rapid expansion of financial services, comparing the differential mortality rates and tracking its progression over time make it difficult to use. Given that wealth estimates’ sensitivity to mortality rates, as described above, means that this is an area that requires constant refinement. The assumption that mortality differential does not change over time may lead to systematic bias in the mortality rates. The assumption that within the year, gender, age cell, mortality rates are constant, may also lead to bias. If higher mortality rates lead to lower wealth (through higher health expenditure, tax planning, etc), then the multiplier and wealth will be positively correlated, biasing wealth shares downwards.

Appendix 3. Investment income estimation method

A person could have an investment, such as equity in company X. Every year, the company pays out R10 dividend or investment income. If we know from another source that the yield is 10%, we can estimate that the equity value is R100. The investment income method uses the investment income from tax data, and an assumption about what the yield is, to reverse out an estimate of the value of the estimate. There are two different methods of calculating the yield multiplier, which also inform the data requirements. The first is by ascertaining the average yields on different categories of asset, and multiplying it by the proportion of income from that asset of the total income (i.e. composition of investment income). This is simple where investment income data is classified by type. However, where this is not the case, asset composition data need to be estimated from other sources. Atkinson & Harrison (1978) take the asset compositions from estate duty method. They then combine this with yield data calculated on 28 categories of assets. Assets are grouped according to different methods used to estimate the yields, mainly to be transparent on the resulting reliability. For example, yield on cash deposit accounts is from commercial banks data, and so is more reliable than yields on unquoted ordinary shares, which used quoted shares as a proxy. Categories of assets excluded are those that generate non-taxable income (e.g. tax-exempt savings products), capital gains (as that income does not correspond to a continued wealth holding), or rent from owner-occupied houses. Investment income data only covers those with assessable income beyond a threshold, and so then only represents the uppermost ranges in the estate estimates. The data for this were tables that summarised net investment income by ranges in ‘surtax’ returns, meaning only income above a very high income threshold is included. However, the 28 categories of assets enables more detailed yield multipliers to be used. The wealth estimates using this investment income method are highly sensitive to the yield multipliers, namely the choice of yields, and the estimates of the asset composition.

In more recent work, Saez & Zucman (2014) use a different method of capitalisation. They calculate a capitalisation factor that is a ratio of the equivalent category’s aggregate Flow of Funds (or national accounts) wealth to the tax return income. In doing so, it reduces risk of estimating yields, and by design that the tax income-based wealth estimates are consistent with the national accounts’ wealth estimate. Pre 1962, no micro data were available, and so a series of top incomes constructed from tabulations of income and its composition by size of income is used. After 1962, the authors use a large sample of taxypayers’ tax returns. The authors use nine categories of capital income: taxable interest (generated by fixed income claims), tax-exempt interest (generated by state and local bonds), dividends and capital gains (generated by corporate equities), and business and rental income (generated by closely held businesses and non-home real estate). This is sufficient as the yield multiplier is calculated on the basis of the equivalent categories in the Flow of Funds. The first step is to report the shares of taxable capital income by fractile relative to the total population. The second step is to capitalise the investment income over the asset classes. Within each asset class, the authors assume that everybody has the same capitalisation factor, which is a strong assumption. However, the authors study foundational wealth to show that the return is not different among asset classes as wealth increases, thus the assumption holds. Equities can result in capital gains income and dividend income. Realised gains also provide useful information about stock ownership, but the selling of stocks is lumpy (say stock gets sold all at once at retirement age, rather than gradually or cyclically). A mixed capitalisation method is used here, so that gains are ignored when ranking individuals into wealth groups, but taken into account when computing the top shares. This decision does not really affect the top shares, given that those who receive high dividends also received high capital gains, and so how these are distributed across groups does not change dependent on whether gains are included or excluded. Dealing with assets that do not generate taxable income, namely pensions and owner-occupied housing, is the third step. This was excluded by Atkinson and Harrison. These categories are not that important for the top wealth shares, but nevertheless are included. The value of owner-occupied housing is inferred from property taxes paid, assuming all property owners pay the same property tax. In reality, this varies across and within States, so using tax addresses would improve this calculation. However, this is not seen as a big problem given that only 5% of the wealth of the top 0.1% is from housing. Pension fund, which in the US account for a third of total household wealth, is more evenly distributed than overall wealth, and so is distributed in line with the Survey of Consumer Finance (SCF), a household survey, and a similar process is followed for life insurance and non-taxable fixed income claims (i.e. government bonds). Trust wealth is estimated by using the trust income in the individual’s tax return. Offshore wealth is accounted for by distributing a separately estimated series and distributing it similarly to trust income (i.e. highly concentrated). The robustness of these estimates is checked by reconciling them with estimates from estates.

The choice to use this method is well informed by the following studies. Alvaredo et al. (2016) reject using the income capitalisation method, citing the insufficient breakdown of investment income categories in recent UK tax data. In Atkinson & Harrison (1974), estate data categorisations are used to estimate the components in investment income. This hybrid technique is still not deemed sufficient, with six categories of assets. Saez & Zucman (2014) specifically use the investment income capitalisation technique because there is more disaggregated data (11 categories). Interestingly, when looking at US wealth from the Flow of Funds (equivalent to the national accounts), they noted financial assets were a more significant component compared to France or UK, and hence this method seemed more appropriate.

Appendix 4. Combining methods and data sources: Garbinti et al. (2017)

Table

Methodology	Data	Years
Estate duty	Large individual-level micro-samples of estates	1800–1902
Estate duty	Detailed tabulations by age and asset category	1902–1970
Estate duty	National micro-samples of inheritance tax returns	1977, 1984, 1987, 1994, 2000, 2006 and 2010
Income capitalisation	Income tax tabulations	All years
Income capitalisation	Sample of income tax returns	1970, 1975, 1979, 1984. 1988–2010
Income capitalisation	All income tax returns	2010–2012

The tax micro files provide individual level information about the component assets that generate income, and, as explained in the previous section, average rates of return for each component is used to calculate the stock of the asset. Assets that do not generate taxable income, such as owner-occupied housing and life insurance, are imputed using housing and wealth surveys. The surveys are divided by age, then in each age category by financial income, and then in each age/financial income category by labour and replacement income. The proportion of individuals holding the asset in the group (extensive margin) and the share of the asset owned by the group (intensive margin) is calculated. For imputed housing, a debt ratio is calculated for the group that takes into account a mortgage/bond. In the income tax micro files, groups are defined according to the same dimensions (age, financial and labour incomes). Within each group, the authors randomly select tax units who own the asset according to the extensive margin computed in the survey. Those tax units are assigned the proportion of total asset, adjusted for the debt ratio in the case of imputed housing. Where this information from the survey is at household level, the values would have to be allocated to the tax unit. Finally, the different components of capital income are calculated by simply multiplying each asset by the corresponding economic rate of return. Interesting to note is that interest and dividend income are defined differently across the years, for example with income from mutual funds first classified as interest before 2005, and then dividends. This led the authors to jointly capitalise taxable interest and dividends and then reclassify them into equities or bonds proportionally to the respective importance of interests and dividends in the individual income.

There are some years where micro-samples were not available. These missing years were interpolated by using the asset categories from national accounts and applying linear trends in within-asset-class distribution.

Appendix 5. A study of intergenerational wealth in Denmark

Boserup et al. (2016) use meticulously collected data on wealth in Denmark from both the statistical agency and the tax administration to generate a baseline sample of child cohorts who are 45–50 years old in 2010 and their (biological) parents observed at the same point in the life-cycle. The authors take the average wealth of children over the three-year period 2009–2011 and measure (average) parental wealth 25 years before, corresponding to the median age of the parents when getting the children, i.e. 1984–1986. Thus, parents and children are approximately the same age when wealth is measured. They first provide non-parametric evidence of the relationship between child-cohort and parent-cohort wealth, in the middle of life-cycle, showing a strengthening relationship towards the top of the distribution, with a child average rank going from percentile 68 to percentile 73, when going from percentile 99 to percentile 100 in the parental wealth distribution. They also investigate the role of bequests in intergenerational wealth, but interesting to note is that they do it without direct information on bequests, but create an experiment that exploits inheritance laws whereby a spouse can retain undivided possession of an estate, and inheritance of that estate only occurs after death of both parents. They use this to create a treatment group where the parent dies in 2010, and a control group where the parent does not die in 2010, and compute the percentile ranks for each individual separately in each group, and look in each group at the mobility pre- and post- parental death.