Estimating the Distribution of Household Wealth in South Africa Aroop Chatterjee Léo Czajka Amory Gethin April 2020 This working paper is the result of a collaboration between the Southern Centre for Inequality Studies and the World Inequality Lab. WID.world WORKING PAPER N° 2020/06 Estimating the Distribution of Household Wealth in South Africa∗ Aroop Chatterjee Léo Czajka Amory Gethin April 2020 Abstract This paper estimates the distribution of personal wealth in South Africa by com- bining tax microdata covering the universe of income tax returns, household surveys and macroeconomic balance sheets statistics. We systematically compare estimates of the wealth distribution obtained by direct measurement of net worth, rescaling of reported wealth to balance sheets totals, and capitalisation of income flows. We document major inconsistencies between available data sources, in particular re- garding the measurement of dividends, corporate assets and wealth held through trusts. Both household surveys and tax data remain insufficient to properly capture capital incomes. Notwithstanding a significant degree of uncertainty, our findings reveal unparalleled levels of wealth concentration. The top 10 per cent own 86 per cent of aggregate wealth and the top 0.1 per cent close to one third. The top 0.01 per cent of the distribution (3,500 individuals) concentrate 15 per cent of household net worth, more than the bottom 90 per cent as a whole. Such high levels of inequality can be accounted for in all forms of assets at the top end, including housing, pen- sion funds and other financial assets. Our series show no sign of decreasing wealth inequality since apartheid: if anything, we find that inequality has remained broadly stable and has even slightly increased within top wealth groups. ∗Aroop Chatterjee, Southern Centre for Inequality Studies - University of Witwatersrand; Léo Czajka, Université Catholique de Louvain; Amory Gethin, World Inequality Lab – Paris School of Economics. We thank the SA-TIED Data- lab team, as well as Facundo Alvaredo, Thomas Blanchet, Keith Breckenridge, Josh Budlender, Aalia Cassim, Lucas Chancel, Allan Davids, Andrew Kerr, Murray Leibbrandt, Thomas Piketty, Michael Sachs, Imraan Valodia and Eddie Webster for helpful insights. We also thank seminar participants from the Southern Centre for Inequality Studies, WiSER, School of Economics and Finance at the University of Witwatersrand, and SALDRU at the University of Cape Town. We acknowledge financial support from UNU-WIDER SA-TIED project, the Ford Foundation, the Sloan Foundation, the United Nations Development Programme and the European Research Council (ERC Grant 340831). This study is reproduced here with acknowledgement of the copyright holder UNU-WIDER, Helsinki. The study was originally commissioned under the UNU-WIDER project, Southern Africa - Towards Inclusive Economic Development (SA-TIED). 1 Introduction South Africa is by most contemporary measures the most unequal country in the world. This is a clear legacy of colonialism and apartheid, where minority rule was premised on racially motivated exclusion of the majority from ownership and participation in the economy. However, despite having a progressive constitution and policy mandate, post-apartheid democratic society seems to have reproduced inequality along the same lines. To study this evolution, most analyses of inequality have focused on inequality of incomes and opportunities, but relatively little attention has been given to wealth inequality. However the available evidence suggests that wealth is significantly more unequally distributed than income and thus may greatly contribute to maintain or exacerbate the discrepancies in standards of living and access to economic opportunities. Studying wealth inequalities is therefore crucial to accurately measure its concentration over time, identify the root causes of the current persistence of extremely high levels of inequality in South Africa, to eventually understand how to best overcome it. In this respect this study comes at timely moment given the current policy debate about the different reforms needed to address wealth inequality, such as wealth tax (Davis Tax Committee 2018) or expropriation without compensation.1 This paper estimates the wealth distribution in South Africa from 1993 to 2018, and advances the lit- erature in several ways. First, we systematically contrast all existing data sources in South Africa that can inform estimations on wealth, including macro-economic data, all relevant household surveys, and newly available tax microdata from 2010 to 2017. This inspection demonstrates the crucial lack of fully comprehensive and reliable data available to directly measure the distribution of wealth in South Africa. We further show that some key income components, which could be used to indirectly estimate wealth concentration, are also insufficiently captured, even by the most recent and accurate tax microdata avail- able. This is particularly salient for capital incomes such as rental income, interests and dividends - which almost exclusively benefit the very highest income earners. Secondly, we contribute to the methodological literature on wealth measurement (e.g. Saez and Zucman 2016; Roine and Waldenström 2010) by systematically comparing alternative methods of estimating the wealth distribution when only incomplete data is available, namely: partial direct measurement, rescaling and income capitalisation. Specifically, this paper is the first to systematically apply the income capitalisation method to estimate the distribution of wealth in South Africa. This method allows to measure wealth inequality in spite of the absence of reliable microdata directly measuring wealth, by estimating wealth stocks from the income flows they generate. Thirdly, we improve on existing studies by using a combination of the above-mentioned methods, merging tax microdata with surveys to account for the fact that higher in- comes are better captured by fiscal data, and harmonising the resulting distribution with the National Accounts to ensure aggregates are consistent with macro totals. This paper thus contributes to the Dis- tributional National Accounts (Alvaredo et al. 2016) agenda by creating new ways to bridge the gap between macro and micro data to retrieve consistent distributional estimates. Finally, assuming that under-representation of top wealth groups in surveys has remained constant before 2011, we are able to use the income capitalisation method to reconstruct a time series since 1994. The rest of the paper unfolds as follow: section 2 reviews the literature on wealth in South Africa, section 3 discusses the available sources on aggregate wealth, section 4 discusses the available microdata sources on the distribution of wealth, section 5 discusses the different methods, section 6 presents the results, and concludes by comparing our favorite estimates with that of other countries. 1 For a broader discussion on the importance of research on wealth inequality in South Africa, see Chatterjee (2019). 1 2 Measuring the wealth distribution: South Africa in an international perspective Although the estimation of wealth inequality has a long history (e.g. Mallet 1908; Clay 1925; Langley 1950; Daniels and Campion 1936; Atkinson and Harrison 1974), recent improvements in available data and methodological approaches has led to a resurgence in studies on the distribution of wealth. Accord- ingly, a new body of literature has studied long terms trends in wealth inequality (Piketty 2011), the importance of tax havens (Zucman 2014), the use of different techniques to estimate wealth concentra- tion and the importance of combining available data sources in conducting this research (e.g. Garbinti, Goupille-Lebret, and Piketty 2018; Saez and Zucman 2016). To understand how we can estimate a wealth distribution in South Africa, we briefly review the main studies and techniques, and locate key papers on the South African case within this. We group the literature according to the essential components of such a studies: aggregate household wealth, survey sources of information about the distribution, and administrative data sources of information about the distribution. 2.1 Literature on aggregate household wealth Recent studies internationally have relied on official household sector balance sheet statistics (referred to as the household balance sheet). These form part of a system of national accounts that capture all economic activity, in the form of both stocks and flows, and which therefore provides internally con- sistent and internationally comparable estimates of aggregate household wealth. The development of national account statistics to include stocks and wealth concepts is relatively recent. For example, the US household balance sheet was systematically developed in the late 1980s (Wolff 1989), in Germany, the first official balance sheets were released in 2010 (Piketty and Zucman 2014). The System of Na- tional Accounts, an international standard for national accounting, first included guidelines for wealth only in 1993. The standards that inform present statistics come from 2008 (United Nations 2009). In this context, South Africa has firmly been part of this international trend, with the first household balance sheet estimated in 1999 (Muellbauer and Aron 1999). Since then it has been continuously consolidated and now forms part of the official quarterly statistical release of the National Accounts of the South African Reserve Bank (e.g. South African Reserve Bank 2015). The household balance sheet has been estimated backward until 1970 which interestingly allows for long-run analyses. The household balance sheets show that net household wealth, as a percentage of households disposable income, fell from an average of 315 percent in 1980-1998 to 283 percent for the 1999-2003 period, but rose again above 300 percent in 2005. The decomposition reveals a declining significance of liquid assets and the rise of share-holding, pension assets and debt, in line with South Africa’s liberalisation policies (Aron, Muellbauer, and Prinsloo 2006, 2007). Muellbauer and Aron (1999) estimate that from the early 1980s to 1997, the value of housing wealth declined, with pension wealth overtaking it as a proportion of personal disposable income in the early 1990s. Liquid assets, such as bank and building society deposits, declined from the early 1980s, while personal debt rose. After this period, pension wealth was a significant contributor to the recovery in household wealth, while housing wealth recovered due to valuation increases in the private property market and equity prices (Kuhn 2010). These trends continued until the global financial crisis affected property wealth and household debt in 2011 (Walters 2011).2 2 The aggregate balance sheet also provides useful sources of data for decomposition analysis, as per Piketty and Zucman (2014). See Orthofer (2015) for a study on the proportion of South African household wealth changes in South Africa from 1970 to 2014 that are a result of quantity (saving-induced) versus price (revaluation-induced) effects. 2 2.2 Distribution of wealth using complete micro administrative data Some researchers have been able to take advantage of national wealth databases, which consist of com- plete records of assets and liabilities. For example, Boserup, Kopczuk, and Kreiner (2016) use admin- istrative tax records from the Danish Tax Agency (SKAT), which collects, in addition to information of various income sources, information about the values of asset holdings and liabilities measured at the last day of the year for all Danes. Wealth and debt components, such as all deposits, stocks, bonds, value of property, and deposited mortgages, as well as all types of debt, are third-party reported and linked to the individual ID numbers. There are also records that allow matching the identification numbers of parents and children to study the intergenerational transmission of wealth. The data is unfortunately not currently organised like this in South Africa. However, it is important to note that all the components to do so are in place. Third party reporting is already done by all financial institutions to the South African Revenue Service (SARS), but the data has not been made available. Organising this data to link with individual income tax records would not only benefit researchers, but also SARS, as such a dataset can be used to cross check the consistency of the reported income level with the change in net wealth during the year under the assumption of a given estimated consumption level. Incidentally these types of records have proved extremely successful in helping tax enforcement (Kleven, Knudsen, Kreiner, Pedersen, and Saez 2011). The identification numbers linking individuals to their parents are administratively required by South African Home Affairs. Combining this information with administrative data about wealth would allow researchers to study intergenerational dynamics in wealth transmission. 2.3 Distribution of wealth according to survey data Data on aggregate wealth informs trends in levels and composition of wealth at the national level, but does not allow us to specify how wealth is distributed over the population. In the absence of disaggre- gated administrative data, studies on wealth distribution typically rely on survey instruments such as the Survey of Consumer Finances (SCF) for the US - which provides regular information on assets and debts since the 1960s - or the Household Finance and Consumption Survey (HFCS) which collects data on the assets and liabilities of households in 18 European countries since 2010. The National Income Dynamics Survey (NIDS), a household panel survey conducted in South Africa, has collected information on the assets and liabilities of South African residents at the household and individual levels in three waves since 2008. Comparing wealth aggregates from the SARB with the one they estimated from the NIDS, Daniels and Augustine (2016) observe that the NIDS survey only captures approximately 3 per cent of the financial assets recorded in the national accounts. Lower levels of financial assets in the NIDS suggest that the survey failed to capture the top end of the wealth distribution. Non-financial assets from the NIDS are around four times higher than the national accounts due to the inclusion of durable assets, not usually included in measures of aggregate net wealth. Mbewe and Woolard (2016) explore two waves of this data, 2010-2011 (wave 2) and 2014-2015 (wave 4) to build measures of household net worth. They estimate that the share of the top 10 per cent accounts for 87 per cent of total net assets in wave 2, and 85 per cent in wave 4 (excluding durable assets). Furthermore, the bottom decile has negative wealth, while the next seven deciles together hold 4 and 7.6 per cent of net wealth in waves 2 and 4 respectively. Exploiting NIDS’ demographic dimensions, they also reveal that approximately 60 per cent of the top decile are White individuals.3 Within-race inequality is very pronounced as well. The Gini index for wealth among the White group is equal to 0.74, compared to 0.98 among the Black group (wave 2). 3 Notice however that the authors exclude the Asian group from their estimates due to undersampling. 3 There is potentially another source of survey data measuring wealth, but this has not been considered for this paper due to the proprietary nature of the data. Van Tonder, Van Aardt, and De Clercq build a set of distributional balance sheets using private sector survey data, the Momentum/Unisa Household Financial Wellness Index surveys conducted between 2011-2015, which cover 12,500 households. They merge these surveys with data from the Bureau of Market Research (BMR) Household Income and Expenditure Database to derive 2016 distributional balance sheet statistics. A slightly different point of reference, they find that the top income decile holds about 51 per cent of household net wealth. This viewpoint however only limits the wealth held to those in the labour market, and so underestimates wealth and its distribution at the top end. The survey records aggregate household net wealth at R 7,344 billion in the fourth quarter of 2017, compared to the National Accounts which estimate it at R 10,835 billion in 2017. However, the methodology used to derive these figures remains very opaque, which makes it hard to understand sources of differences with NIDS and SARB data. Making this survey and data publicly available in an appropriate form to researchers would be key to contrasting it to other data sources and to improving our understanding of the wealth distribution in South Africa. In any case, studies relying on survey data are limited when estimating the top shares due to under- sampling, non-response and under-reporting issues which are particularly pronounced at the top end of the distribution. Given the extreme levels of wealth concentration in South Africa, this implies that estimates based on surveys can only depict a truncated picture of the reality. In this regard, administra- tive sources are generally more exhaustive and usefully complement estimates relying on survey data alone. 2.4 Distribution of wealth using estate duty and personal income tax data Estate duty data has been commonly used to estimate wealth distributions. Mallet (1908) used estate duty data as early as 1908, and this method has since been developed to estimate a historical series even with minimal but useful summary data, as in, for example, Atkinson and Harrison (1974) and Piketty, Postel- Vinay, and Rosenthal (2006). Indeed, one of the earliest studies to look at the distribution of wealth in South Africa used estate duty returns in the Natal province in 1974/75 (McGrath 1982). The data was obtained from Master of Supreme Court (rather than from the tax authority, which had no demographic information attached to the estate duty records). Adjusting the provincial data to make it nationally representative, and using a mortality multiplier, the assets of the deceased were used to estimate the assets of the living. McGrath estimated that the top 5 per cent of the population owned 88 per cent of total household wealth. Assuming that 94 per cent of all wealth was held by the white population at that time (as per the information in the estate duty records), the demographic breakdown also provides some interesting statistics. Among the White population, the top 10 per cent held 65 per cent of wealth, while among the Coloured and Asian groups, it held 96 and 94 per cent of wealth respectively. The advantage of this approach is that estates of the deceased are directly measured. Unfortunately, even though there is an estate duty collected regularly in South Africa today, we have not been able to access data on estate duty records. Access to such data would be crucial to improve our understanding of the wealth distribution and its intergenerational dynamics. Another set of data useful to estimate the top shares of the wealth distribution is personal income tax data. This provides indirect information on wealth through declarations of income derived from capital ownership. These incomes can then be capitalised to estimate their asset bases (see Saez and Zucman (2016) and Garbinti, Goupille-Lebret, and Piketty (2017) for recent studies in the case of the US and France respectively). In South Africa, Orthofer (2016) has used reported incomes from administrative personal income tax data (PIT) to approximate the wealth distribution, comparing this to the distribution of wealth measured in the NIDS. Given that this study takes an approach which is closest to ours, let us briefly discuss it in more detail. For her estimates using PIT data, which does not cover the lower end of the distribution because of filing rules, Orthofer fits a lognormal distribution below the filing threshold to simulate a bottom end. She then uses the sum of investment incomes (interest, dividends and rental 4 income) and pension contributions as a proxy for wealth. For her estimates using NIDS data, she takes the sum of assets and liabilities reported in the survey; she also re-samples the top of the distribution from a Pareto distribution to account for the underrepresentation of top income and wealth groups in the survey. In both cases, she estimates that the top 10 per cent share lies between 90 per cent and 95 per cent, and the top 1 per cent share between 50 per cent and 60 per cent. This study has made an important contribution in being the first to use both surveys and income tax data to measure wealth inequality in South Africa. However, it suffers from at least three major limitations that we seek to address in this paper. Firstly, her PIT estimates only cover specific components of wealth, those who correspond to investment income (i.e. financial assets) and pension contributions (i.e. pension assets). They exclude owner-occupied housing wealth altogether, which we find to amount to as much as 28 per cent of household wealth in 2018 (see section 3). Secondly, her results based on PIT data do not account for the fact that the composition of income is not the same as the composition of wealth. Because assets have different rates of return (for example, bonds tend to have lower rates or return than corporate shares), the income capitalisation method requires applying differential multipliers by asset class. As a result, Orthofer’s estimates using PIT data better correspond to the distribution of financial incomes than to the distribution of wealth. The difference between these two distributions is now well- known in the literature: in the case of the United States, for instance, Saez and Zucman (2016) show that wealth is typically less concentrated than capital income. When it comes to estimates using the NIDS, we believe that the extraordinarily high levels of wealth concentration found by Orthofer are in large part due to the mismeasurement of pension assets. Accord- ing to her results, the top 1% owns as much as 99% of pension assets in the economy (see table 5, p. 18). This seems unrealistically high, given that more than 10% of the South African adult population contribute to pension funds, and at least 6% of the South African adult population received private pen- sion income from a pension fund in 2017.4 Looking closer at the NIDS, we find this inconsistency to be due to massive under-reporting in the survey data: indeed, a large share of pensioners and wage earners with positive pension contributions declare having no pension asset, which is by definition impossible. We correct for this discrepancy by imputing pension assets to individuals contributing to pension funds or receiving private pension income, using predictive mean matching. This increases the share of in- dividuals with positive pension assets in the survey data from 6% to 16% of the adult population. It also improves considerably the coverage of aggregate pension assets in the survey, which increases from about a third of the macro total reported by the SARB to close to 100%. This paper builds upon this existing literature by combining surveys and tax data, but with significant differences. Firstly, we systematically contrast all data sources that can inform estimations on wealth inequality (including all household surveys useful for this purpose, as well as tax data). Secondly, we directly combine surveys and tax data at the individual level, rather than resampling individuals from these two types of datasets. This requires us to thoroughly harmonise income concepts, but it allows us to study the entire distribution with a greater level of precision. Thirdly, we also capitalise incomes from surveys alone, and compare our results to those obtained when combining surveys and tax data. To the best of our knowledge, our study is the first to apply the income capitalisation directly to survey data, with no correction for the under-representation of top income groups, and to assess the consequences of this under-representation on the measurement of top wealth inequality. Quite surprisingly, we find in our case that both approaches yield very similar results. We interpret this as evidence that while surveys do understate the concentration of incomes at the top end, they still allow us to capture the core structure of wealth concentration as long as assets and liabilities are properly rescaled to match macroeconomic totals. 4 Authors’ computations using data from income surveys (pension contributions) and data from the matched IRP5-ITR12 income tax panel (pension income). 5 Fourthly, we systematically match specific wealth components with the corresponding balance sheets totals. Our wealth distribution is therefore fully consistent with official macroeconomic figures published by the South African Reserve Bank. Correcting for such micro-macro discrepancies is in our view crucial to both better measuring the distribution of wealth and improving the international comparability of existing studies. To be sure, the estimates of aggregate wealth published by statistical institutes are far from being perfect, and they are also likely to suffer from measurement error. Yet, the framework outlined by the United Nations’ System of National Accounts does represent the best attempt to provide internally consistent, comparable and measurable definitions of household wealth. Ignoring the fact that surveys massively understate major components of wealth seems in our opinion a much stronger assumption than attempting, even in an imperfect way, to address existing measurement errors. We know with a relatively high degree of confidence, for instance, that the NIDS survey does not cover more than 4% of bonds and stock held in the economy. Estimates which do not attempt to account for this problem effectively leave out more than 30% of household wealth held in South Africa. To our knowledge, our study is the first one in South Africa to correct for these micro-macro gaps. Fifthly, the tax microdata we use covers the entire universe of taxpayers. This allows us to go beyond the top 10 per cent or the top 1 per cent and to derive estimates of average wealth covering the very top end of the distribution. As we show in section 5, understanding wealth inequality within top wealth groups is absolutely crucial for the study of wealth inequality in South Africa, both in terms of measurement and policy, as the top 0.1 per cent alone owns a substantial share of household wealth. Finally, our methodology allows to estimate a time series, rather than a point estimate, that gives us a sense of longer term trends since 1993. 3 Data sources on aggregate wealth This section discusses the data sources available to measure total household wealth and its composition in South Africa. We then turn to a systematic comparison of micro and macro data sources in the next section. 3.1 The household balance sheets In South Africa, the first comprehensive attempt to estimate the value of total household wealth in the economy goes back to Muellbauer and Aron (1999), who collect and combine a number of data sources to provide figures on the market value of the assets and the liabilities of the household sector since 1975. The South African Reserve Bank (SARB) has since then updated and revised these figures on a yearly basis. Non-financial assets are divided into residential buildings and other non-financial assets. Residential buildings correspond to the market value of residential properties owned by households. Other non- financial assets include both land underlying dwellings and business assets. Financial assets are divided into interest in pension funds and long-term insurers, assets with monetary institutions, and other financial assets. Interest in pension funds and long-term insurers corresponds to all pension assets holdings of the household sector. It is the sum of the total assets of official pension and provident funds (series KBP2215 in Capital Markets Statistics), the total liabilities of private self- administered pension and provident funds (KBP2339), and the liabilities of long-term insurers under unmatured policies from the pension business (KBP2215).5 Assets with monetary institutions include 5 The original estimates of the South African household balance sheets done by Muellbauer and Aron (1999) excluded life insurance assets and all other assets associated with the non-pension business of long-term insurers. However, these items are now included by the SARB in line with the SNA guidelines. 6 Figure 1: The evolution of household wealth in South Africa, 1975-2018 -50 0 50 100 150 200 250 300 350 % o f n at io na l i nc om e 19 74 19 76 19 78 19 80 19 82 19 84 19 86 19 88 19 90 19 92 19 94 19 96 19 98 20 00 20 02 20 04 20 06 20 08 20 10 20 12 20 14 20 16 20 18 Residential buildings Other non-financial assets Other financial assets Assets with monetary institutions Interest in pension funds and long-term insurers Mortgage debt Other debt Notes: This figure shows the level and composition of household wealth in South Africa between 1975 and 2018, expressed as a share of the net national income. Source: authors’ compilation based on data from the South African Reserve Bank. all forms of currency and deposits with banks, mutual banks, the Land and Agricultural Bank, and the Post Bank, as well as notes and coins held by households. This category therefore includes both deposits generating interest income (savings accounts) and other liquid assets which have no corresponding mea- surable income flow (such as cheque accounts, notes or coins). Other financial assets include investment in government and public entities stock, collective investment schemes, corporate bonds and equities, other long-term deposits and households’ investment in foreign assets. Finally, the SARB decomposes household debt into mortgage advances, corresponding to loans provided by the commercial bank sector, and other debt, which includes trade credit, personal bank loans, credit card debt, instalment sales and lease agreements, non-bank loans granted by micro-lenders, and other loans. Figure 1 plots the evolution and composition of total household wealth between 1975 and 2018, ex- pressed as a share of national income. Aggregate net wealth has followed a U-shaped curve, declining from about 300 per cent of national income in 1975 to 220 per cent at the end of the 1990s, and rising back to more than 260 per cent at the beginning of the twenty-first century. In 2018, financial assets amounted to two years of national income. Within financial assets, pension assets have risen signifi- cantly and are now the biggest component of household wealth. Non-financial assets amounted to one year of national income in 2018, declining in importance over the years from just under two years of national income in 1974. Household debt rose significantly between 1975 and 2008, in large part due to a boom in mortgage advances in the early 2000s (see Figure A4), and has slightly declined since then, amounting to about 55 per cent of the national income today. 7 3.2 The limitations of available balance sheets: from institutions to asset classes As discussed in section 2, we aim to estimate the distribution of household wealth by combining infor- mation on capital income flows with directly measured stock data, not previously done for South Africa. There are at least five limitations to available balance sheets statistics which we discuss below: the decomposition of non-financial assets, the decomposition of housing wealth into tenant-occupied and owner-occupied, the decomposition of financial assets, the decomposition of pension and life insurance assets, and the inclusion of wealth held offshore in tax havens. Land underlying dwellings The first issue is that the other non-financial assets category provided by the SARB includes both land underlying dwellings and business assets. These two components are arguably distributed very differ- ently. For our purpose, in particular, it is reasonable to assume that land underlying dwellings is dis- tributed similarly to residential buildings – therefore defining total housing assets are the sum of land and residential buildings. We assume that 70% of other non-financial assets corresponds to land underlying dwellings, the remaining 30% amounting to the assets held by unincorporated businesses. Tenant- versus owner-occupied housing The two important components of "Residential buildings" are tenant-occupied housing, and owner- occupied housing. Available studies combining surveys with tax microdata typically assume that the distribution of tenant-occupied housing can be well approximated by the distribution of rental income, while owner-occupied housing assets are better captured using direct measurement available from sur- veys or administrative data (Saez and Zucman 2016; Garbinti et al. 2017). Unfortunately, the "Resi- dential buildings" category published by the SARB does not provide this decomposition, so we choose to estimate the proportions from survey data. The General Household Survey is the only survey that systematically asks both tenants and owners to provide a value for the dwelling in which they live. Our results show that between 22 per cent and 26 per cent of household housing assets are owned by households renting to private individuals over the 2013-2018 period (see appendix for method). Non-pension financial wealth "Assets with monetary institutions" and "other financial assets" gather together very different forms of financial assets. "Assets with monetary institutions" include both non-interest bearing deposits such as cheque accounts, which do not generate any income flow, and interest bearing deposits which generate interest income. "Other financial assets" include both bonds and corporate shares, which generate in- terest and dividends respectively. We follow Orthofer (2015) and assume that the composition of other financial assets held by households is similar to that reported by unit trusts.6 This implies that between 80 per cent and 95 per cent of other financial assets consist in corporate shares over the 1975-2018 pe- riod, the remaining being classified as interest-bearing deposits.7 Finally, we separate currency, notes and coins from interest-bearing deposits by using published data from the Money and Banking Statistics 6 As discussed by Orthofer (2015), “A breakdown by asset class can be estimated by applying the portfolio composition of the respective counterparties (monetary institutions, pension funds and long-term insurers as well as unit trusts) to the total of household assets held with these institutions. In practice, we consider all assets with monetary institutions as cash equivalents and apply the portfolio composition of unit trusts to the other financial assets component." 7 More precisely, we estimate the share of corporate shares in other financial assets by comparing the market value of ordinary shares held by unit trusts (KBP 2412) to the sum of the market values of security holdings of public sector entities, stocks and debentures held by unit trusts (KBP 2410 + KBP 2411) in the capital market statistics published by the SARB. 8 Table 1: The level and composition of household wealth in South Africa in 2018 Market value (R billion) % of national income % of net wealth Non-financial assets 4504 111.4 % 42.4 % Owner-occupied housing 3020 74.7 % 28.4 % Tenant-occupied housing 988 24.4 % 9.3 % Business assets 497 12.3 % 4.7 % Financial assets 8294 205.1 % 78.0 % Pension assets 2944 72.8 % 27.7 % Life insurance assets 1412 34.9 % 13.3 % Bonds and interest deposits 1798 44.5 % 16.9 % Currency, notes and coins 87 2.2 % 0.8 % Corporate shares 2053 50.8 % 19.3 % Total liabilities 2170 53.7 % 20.4 % Mortgage debt 1022 25.3 % 9.6 % Non-mortgage debt 1148 28.4 % 10.8 % Net household wealth 10629 262.9 % 100.0 % Offshore wealth 575 14.2 % 5.4 % Net wealth incl. offshore wealth 11204 277.1 % 105.4 % Notes: The table shows the level and composition of household wealth in South Africa in 2018. The market value of each component is expressed in current billion rands. Source: Own estimates combining available data sources from the SARB. of the SARB (series KBP1312).8 These amount to 0.8% of net wealth, which is relatively comparable to figures from other countries where balance sheets data are available: in the US, for instance, non- interest financial assets amount to about 1 per cent of personal wealth in recent years (Saez and Zucman 2016). Pension assets and life insurance Pension assets correspond to the assets accumulated by wage earners through contributions to pension funds throughout their career, so they should in large part be distributed to wage earners and pensioners receiving pension income or annuities. Life insurance assets, by contrast, corresponds more to a form of savings device, but they do not directly generate interest income, so they cannot be categorised with interest deposits or bonds and have to be distributed differently. As explained above, the share of interest in pension funds and long-term insurers corresponding to assets held by long-term insurers is recorded in the Capital Markets Statistics published by the SARB under series KBP2215, "liabilities of long-term insurers under unmatured policies from the pension business", so we can measure directly total life insurance assets held by households in the economy accordingly. Offshore wealth Offshore wealth corresponds to the assets held abroad by South African residents, mainly for tax avoid- ance purposes. By definition, these assets are not recorded in official records and are therefore not included in the household balance sheets. Alstadsæter, Johannesen, and Zucman (2018) combine a number of macroeconomic data sources to measure the total amount of financial assets held in offshore tax havens and distribute it to specific countries. They estimate that the equivalent of about 11.8 per cent of South African GDP was held offshore in 2007. We add this quantity to total household wealth in 2007 and extrapolate it to other years by assuming that it has remained a constant fraction of GDP. Given that offshore wealth is known to have grown globally, this is a relatively conservative assumption: if anything, wealth inequality could have increased more since 1993 than what our estimates suggest (see 8 This variable, "Monetary sector liabilities: banknotes and coins in circulation", includes currency, notes and coins held by all institutions, not only households. We assume that 70% of the total can be attributed to households.Given the small share of this component in total wealth, especially at the top of the wealth distribution, our results are not affected by alternative scenarios. 9 Table 2: Ownership rates and coverage of household balance sheets by asset class in NIDS % of adults with asset or debt % of balance sheets covered Wave 4 Wave 5 Wave 4 Wave 5 Household assets Owner-occupied housing 72.3 % 65.2 % 151.7 % 220.8 % Tenant-occupied housing 3.3 % 3.5 % 122.4 % 97.2 % Business assets 5.6 % 5.0 % 135.4 % 59.6 % Pension and life insurance 25.7 % 24.4 % 110.0 % 104.3 % Bonds and stock 1.5 % 1.3 % 3.9 % 3.8 % Household debts Mortgage debt 8.0 % 7.0 % 71.0 % 56.8 % Other debts 36.3 % 33.7 % 54.5 % 37.0 % Notes: The table shows the share of South Africans who declare having a particular type of asset or debt, along with the share of the total value of this asset or debt in the economy captured by the NIDS survey. Source: authors’ computations based on data. The unit of observation is the adult individual aged 20 or above. Calculations based on weighted sample using design weights. section 5), as offshore wealth is well-known for been concentrated at the very top end of the distribution (Alstadsæter et al. 2018). The level and composition of household wealth in 2018 Table 1 shows the detailed composition of household wealth in 2018 after breaking down the balance sheets categories. Pension assets and owner-occupied housing are the largest component of household assets and liabilities, each amounting to about 28 per cent of net wealth. The next most important categories are corporate shares (19 per cent), bonds and interest deposits (17 per cent) and life insurance assets (13 per cent). Business assets are equal to less than 5 per cent of net wealth. Tenant-occupied housing and currency and non-interest deposits represent 9 per cent and less than one per cent of net wealth respectively. Accounting for offshore wealth increases net household wealth by 5 per cent and brings the net personal wealth to national income ratio to more than 275 per cent. 4 Data sources on the distribution of household wealth This section reviews the data sources available in South Africa that can be used to inform a distribu- tion function of household net assets, and ultimately be applied to the National Accounts’ aggregates, described above. 4.1 Household surveys Surveys can provide information either that directly measure household assets, and/or have information about incomes and outflows. The National Income Dynamics Survey, mentioned in section 2, is the only publicly available survey that explicitly collects data on wealth. Out of the five waves of the survey, wave 2, 4 and 5 have wealth modules that can inform both households’ and, for wave 4 and 5, individuals’ net worth. For the purposes of this paper, we only consider the variables that allow us to build a net wealth concept consistent with the System of National Accounts guidelines (United Nations 2009) for comparability and consistency. The surveys have information on business wealth, housing properties and land, life insurance, pension and other retirement assets, equity wealth, debt and bonds. However, we faced several challenges to extract reliable wealth estimates from this source. Firstly, we uncovered issues in coverage and reliability in each of the five components of wealth. Look- ing at housing first, approximately 40 per cent of debtors do not know the house bond value. With pension and other retirement funds, the information is quite inconsistent. For example, in wave 5 of the 10 Table 3: The coverage of selected national accounts components in South African surveys Gross wages Mixed income Rental income Interest and dividends PSLSD, 1993 87.7 % 51.7 % 38.4 % 11.5 % IES, 1995 76.9 % 55.0 % 9.9 % 8.8 % IES, 2000 70.9 % 37.2 % 23.1 % 3.4 % IES, 2005 80.5 % 64.2 % 21.7 % 3.8 % IES, 2010 80.2 % 71.9 % 13.5 % 4.5 % LCS, 2008 77.7 % 75.8 % 16.3 % 8.4 % LCS, 2015 74.6 % 86.8 % 21.6 % 12.6 % NIDS, wave 1 62.7 % 12.0 % 65.4 % 7.3 % NIDS, wave 2 67.6 % 4.1 % 13.0 % 0.8 % NIDS, wave 3 65.7 % 20.6 % 20.7 % 7.3 % NIDS, wave 4 73.5 % 12.9 % 43.9 % 2.5 % NIDS, wave 5 72.1 % 14.1 % 41.0 % 5.5 % Notes: The table shows the ratio of total income reported in surveys to the total corresponding income reported in the national accounts published by the SARB. PSLSD: Project for Statistics on Living Standards and Development. IES: Income and Expenditure Survey. LCS: Living Conditions Survey. NIDS: National Income Dynamics Study. Source: authors’ computations based on data. The unit of observation is the adult individual aged 20 or above. Calculations based on weighted samples using weights calibrated by the authors’ (see appendix). survey, 61 per cent of individuals declaring contributions to pensions funds declare having no "pension or retirement annuity", while 77 per cent of individuals declaring income from a pension or provident fund declare no "pension or retirement annuity". We correct for these gaps by imputing all missing values using predictive mean matching. For housing wealth, the market value of the house is modelled by number of rooms, number of adults, province and household income. Similarly, we impute missing business wealth for self-employed individuals, as well as pension funds and corporate shares based on factor income, contributions to pension funds and pension income. The comparison of household assets and liabilities reported in NIDS to macroeconomic statistics show important inconsistencies (see table 2). The market value of owner-occupied housing wealth is between 50 per cent and 120 per cent higher in NIDS than in the balance sheets, while tenant-occupied housing is closer to the macro aggregate. This most likely reflects the different methods in measuring market values.9 Business assets are covered very differently in the two waves: they are overestimated in wave 4 and underestimated in wave 5. Pension and life insurance assets, after the correction, seem to be relatively close to balance sheets figures, and they even slightly overestimate them. Other financial assets are extremely badly covered: the total reported in NIDS does not exceed 4 per cent of households’ bonds and stock reported in the balance sheets by the SARB. Household debts are slightly better covered, but still fall significantly below macroeconomic statistics. An alternate method to estimate wealth distribution consists of capitalising incomes (usually the method used for estimating the top end of the distribution from tax data). As more surveys deal with incomes, and generally income reporting is seen as more credible, this provides alternate sources of information for the wealth distribution. The method is discussed in more detail in section 5.1. In this section, we compare incomes from surveys to the corresponding totals recorded in the national accounts. For our 9 It is beyond the scope of this paper to discuss and evaluate these methods. However, this issue is not one specific to South Africa - in the US, survey values have also been found to be higher than in balance sheets figures, and which source of information provides the more accurate estimate of market values is contested (Blanchet 2016; Henriques and Hsu 2014; Dettling, Devlin-Foltz, Krimmel, Pack, and Thompson 2015). As a robustness check, we show in appendix figure A11 that our estimates of the wealth distribution are only marginally affected if one assumes that the balance sheets underestimate housing assets by a factor of two. Another potential issue is how to treat RDP housing, a government-funded social housing project in South Africa, due to compexities around owneship. However, given the typical low market value of these properties, it is unlikely to affect our distributional estimates. 11 Table 4: The coverage of owner-occupied housing, mortgage debt and other debt in South African surveys Owner-occupied housing Mortgage debt Other debt PSLSD, 1993 143.5 % 86.5 % 37.4 % IES, 1995 121.7 % 27.2 % 16.5 % IES, 2000 40.3 % 34.9 % IES, 2005 105.9 % 67.9 % 41.5 % IES, 2010 193.9 % 16.4 % 20.5 % LCS, 2008 145.4 % 13.9 % 18.4 % LCS, 2015 179.5 % 51.0 % 22.2 % NIDS, wave 4 122.3 % 74.3 % 57.4 % NIDS, wave 5 258.8 % 56.8 % 37.0 % Notes: The table shows the ratio of total assets or debts reported in surveys to the corresponding totals reported in the household balance sheets. PSLSD: Project for Statistics on Living Standards and Development. IES: Income and Expenditure Survey. LCS: Living Conditions Survey. NIDS: National Income Dynamics Study. Source: authors’ computations based on data. The unit of observation is the adult individual aged 20 or above. Calculations based on weighted samples using weights calibrated by the authors’ (see appendix). purposes, the components we consider are gross wages (to capitalise pension wealth), mixed income (income from unincorporated enterprises, to capitalise non-financial assets), rental income (to capitalise tenant-occupied housing) and interest and dividends (for equity and bonds). The surveys we consider were designed to capture information about consumption, expenditure and earnings: these are the Project for Statistics on Living Standards and Development (PSLSD) conducted in 1993, the Income and Ex- penditure Surveys (IES) from 1995 to 2010, the Living Conditions Surveys (LCS) of 2008 and 2015, and the NIDS.10 The labour force surveys which provide wage data only cover labour incomes, so would not be appropriate for this exercise. As table 3 shows, rental income, interest and dividends are very poorly covered in household surveys. This is due to this sort of income being concentrated by those at the upper end of the income distribu- tion, who are typically underrepresented in surveys due to issues of sampling and non-response. This motivates our use of the tax microdata to better cover top incomes. Gross wages and mixed income are much better covered in the PSLSD, IES, and LCS than in NIDS. Owner-occupied housing seems to be over-stated relative to the balance sheets in these surveys as in NIDS, echoing the earlier discussion (see table 4). Debts are always below balance sheets totals, but with important fluctuations across surveys. All these limitations justify the need to correct for discrepancies between micro and macro totals. In- deed, the households balance sheets have the advantage of tracking the evolution of wealth consistently, in contrast with surveys which show much greater fluctuations in reported aggregates. By mapping the surveys with macroeconomic statistics, we are at least able to get estimates of the wealth distribution which are consistent with what we know of the level of aggregate wealth and its composition over time, which is what we do in the next section. 4.2 Tax data The tax data, for purposes of this paper, refers to two data sources - the IRP5 data, and the ITR12 data. The IRP5 forms are income tax forms submitted to the South African Revenue Service (SARS) by employers on behalf of their employees, hence covers incomes related to the employment relationship. Specific variables of interest include gross wages as well as contributions to retirement assets (pension, annuities, etc). The ITR12 forms are self-assessment forms, that require taxpayers to disclose income 10There are concerns about whether these surveys are comparable: see for instance Berg and Louw (2005); Leibbrandt, Woolard, and Woolard (2009); Yu (2005); Pauw and Mncube (2007). 12 Table 5: The coverage of capital income in the tax micro- data Rental income Interest income Dividends 2010 9.5 % 25.4 % 2.4 % 2011 11.7 % 25.0 % 5.3 % 2012 12.3 % 28.3 % 3.9 % 2013 13.4 % 28.8 % 5.2 % 2014 12.1 % 27.8 % 25.1 % 2015 12.3 % 27.8 % 10.6 % 2016 13.7 % 31.0 % 13.1 % 2017 6.9 % 18.3 % 15.8 % Notes: The table shows the ratio of total income reported in the tax microdata to the corresponding total reported in the national accounts published by the SARB. Source: authors’ computations based on data. from sources other than employment, so that taxable income can be calculated. Thus, data from this form provides information on business, rental, interest and dividend incomes, which can be capitalised to calculate the asset bases from which the incomes derive. These data sources have been combined into a panel that provides detailed information about all incomes, allowances and deductions (for an overview and discussion of the dataset, see Ebrahim and Axelson 2019). Due to its administrative nature, this data covers the full tax paying population, including individual observations at the top of the distribution. As it is not a sample, it identifies all individual taxpayers, which greatly increases the granularity of measured income flows. This is an advantage over surveys which often suffer from low sample biases. That being said, there are a number of limitations with tax microdata which should also be emphasised. The fact that the ITR12 forms are self-assessed implies that there may be tax evasion or under-reporting of income flows, especially if the likelihood of being controlled by tax authorities is low. More importantly, tax microdata only covers forms of incomes which are useful for tax collection and deductions purposes, which implies that other forms of non- taxable incomes are not reported in the data. This, as we show below, is particular problematic for the measurement of capital incomes. In order to combine the tax data with survey data at the bottom of the distribution and capitalise income flows, we categorise the source codes reported in the tax data into seven broad categories: gross wages, business income, pension contributions, pension income, interest income, rental income and dividends (see appendix table A2).11 Table 5 shows that when looking specifically at capital incomes in the tax data, the reported totals fall significantly below the national accounts. Interest income is better mea- sured than rental income and dividends, reaching between 25 per cent and 30 per cent of total interest received by households in the national accounts. Rental income and dividends are significantly lower and inconsistent, covering between 2 per cent to 25 per cent of national accounts totals. 12 This under-representation of capital incomes in the tax data is due to three main factors. First, the taxable incomes are different from incomes reported in the national accounts, due to filing rules and tax base. This is particularly problematic for dividends, which in the ITR12 relate to dividends from equities that form part of compensation packages, such as equity share plans. These sort of dividends are subject to income tax, and so part of this data set, whereas dividends from regular ownership of equity is subject to a separate dividend tax. Approximately 80 per cent of dividend information would 11The IRP5 and ITR12 data are presented in the form of source codes corresponding to specific taxable income concepts, exemptions and deductions. See Ebrahim and Axelson (2019) for a more complete discussion. 12The particularly low figures obtained in 2017 (fiscal year 2018) are mainly due to the fact that assessment was incomplete at the time of writing. 13 Figure 2: Share of financial assets held through trusts, 1975-2018 0 10 20 30 40 50 60 Sh ar e of a ss et s he ld th ro ug h tru st s (% ) 19 74 19 76 19 78 19 80 19 82 19 84 19 86 19 88 19 90 19 92 19 94 19 96 19 98 20 00 20 02 20 04 20 06 20 08 20 10 20 12 20 14 20 16 20 18 Currency, deposits, bonds, loans Corporate shares Total financial assets Notes: The figure shows the share of total household assets in the economy held by unit trusts. Source: authors’ compilation based on data from the SARB. be recorded through this dividend tax returns (DTR01/2 forms), and this information is urgently required to make our estimate more reliable. Secondly, there may be issues of misreporting of incomes by individual taxpayers. Interest income seems to be poorly covered as a result of incomplete tax filing by taxpayers. In principle, the South African Reserve Bank receives direct information from banks and financial services that they provide about interest. Banks and financial service providers separately supply customers with a tax certificate (IT3(b) certificate), which is meant to inform the interest income declared by the taxpayer. At the same time, the bank sends the South African Revenue Service a third-party submission about incomes its customers’ receive. However, given that interest income is typically low relative to total taxable income, it is possible that small interest income received go unreported. The misreporting of rental income received by individual taxpayers is likely to be more significant, given that rental income is self-reported and that there may be a significant amount of informal letting of fixed property. 13 Finally, the most important issue regarding the coverage of capital incomes in the tax microdata is likely to be due to the definition of the taxpayer. The tax data covers only individuals and does not account forms of capital incomes received through units trusts or investment funds. This is particularly problematic in the case of South Africa, both because wealth is highly concentrated at the top of the distribution and because top wealth groups rely extensively on unit trusts. As shown in figure 2, the share of financial assets held through trusts exploded around the time of, politically, the end of apartheid, and economically, liberalisation and financialisation. Over half of specifically interest bearing and dividend 13Notice here that total rental income paid to individuals in the economy is estimated by the authors based on data from the PSLSD, the IES and the GHS surveys on total rental income paid by households to individual landlords. Therefore, this includes informal rents paid, which may explain why the rental income the tax data is so low compared to the macro aggregate. 14 earnings financial assets are held in trusts. Trusts in South Africa are used more extensively, including housing mutual funds, as well as tax avoidance vehicles, and one mechanism of several to protect against wealth dilation (wealth loss across generations) (Ytterberg and Weller 2010). There is therefore a clear need to access data on trusts to gain clearer estimates of wealth at the top of the distribution, as well as to understand the mechanisms that results in the persistence of wealth concentration. Trusts are required to submit ITR12T forms, this is discussed in the appendix. 5 Bridging the micro-macro gap: the distribution of wealth in South Africa This section brings together micro and macro data sources on household wealth in South Africa and discusses several main methods available to estimate the distribution of personal assets and liabilities that are harmonised to the national aggregates. We compare the results from what we identify as three broad approaches to measuring wealth inequality: direct measurement of wealth, rescaling of reported wealth, and capitalisation of income flows. 5.1 Methodological approaches: direct measurement, rescaling and capitalisation For our purpose, it is interesting to compare three different ways of estimating the distribution of house- hold wealth. The first one, henceforth direct measurement, consists in using reported data on the market value of the assets and liabilities of households. In South Africa, the only publicly available data source enabling such measurement for the entire spectrum of household wealth components is the NIDS survey. This approach is likely to suffer from strong under-estimation of wealth inequality due to non-response and under-sampling issues at the top of the distribution. In particular, the direct measurement approach implies that figures are not consistent with macroeconomic statistics, both in terms of levels and compo- sition of household wealth. In the case of the NIDS we showed in section 4 that the direct measurement approach implies overstating the importance of housing assets and understating the significance of non- pension financial assets. A second way of measuring the distribution of wealth, which we coin as “rescaling" in what follows, con- sists in assuming that the distribution of recorded wealth components and their correlation is relatively well measured by the household survey, but that it is mainly the average amounts of each component which are understated or overstated. In this case, one can obtain an estimate of the wealth distribution by effectively blowing up individual-level assets and liabilities to match the totals recorded in the na- tional accounts. The core identifying assumption is that individuals overestimate or underestimate the value of the assets and liabilities that they report, but that this misreporting is uncorrelated to rank within each asset class. This approach, as we show below, is problematic in our case because it tends to create a number of outliers, both at the top and at the bottom ends of the distribution. This is in large part because debts are very badly measured in survey data, so that rescaling reported values leads to giving unrealistic levels of debt at the very bottom of the distribution. A third approach to measuring wealth inequality is the income capitalisation method. This approach consists in using the capital income flows corresponding to the assets and liabilities of households to approximate the distribution of wealth. In practice, this involves multiplying the income flow of a given asset class by the inverse of the rate of return of this type of asset. Just as in the case of rescaling, the capitalisation of income flows has the advantage of leading to figures which are consistent with aggregate household wealth. The identifying assumption in this case is that of constant rates of return by asset class. If the return to a given asset increases with wealth, for instance, then the income capitalisation method will lead to overestimating wealth concentration. The capitalisation method is only possible for types of assets and liabilities generating income flows. In this paper, we rather propose a “mixed approach" as our preferred methodology to estimate the distribu- 15 Table 6: Estimating the distribution of personal wealth in South Africa: a mixed approach Asset / liability Variable Measurement method Non-financial assets Owner-occupied dwellings Value of home (GHS) Rescaling Tenant-occupied dwellings Rental income Capitalisation Business assets Business income Capitalisation Financial assets Pension assets Pension contributions and pension income Mixed method Life insurance assets Factor income Mixed method Currency, notes and coins Bank account balance (NIDS) Rescaling Bonds and interest deposits Interest income Capitalisation Corporate shares and equity Dividends Capitalisation Liabilities Mortgage debt Reported debt and house value Mixed method Other debts Reported debts and consumption Mixed method Notes: The table shows the methodological approach used to estimate distribution of the different assets and liabilities reported in the household balance sheets. Direct measurement corresponds to reported data on the market value of assets or liabilities. Capitalisation corresponds to assuming that the distribution of an asset follows that of one or several corresponding income flows. GHS: General Household Survey. NIDS: National Income Dynamics Survey. Source: authors’ elaboration. tion of wealth, by combining income capitalisation for available income flows with rescaling when no flow counterpart data is available. As shown in table 6, two types of household assets, owner-occupied dwellings and currency, notes and coins cannot be capitalised and have to be measured directly from available household surveys. We choose to capitalise six types of assets: tenant-occupied dwellings from the rental income received by individual landowners; business assets from the business income received by the business owners of unincorporated entreprises; pension assets from the pension contri- butions and pension income of formal wage earners and pensioners; life insurance assets from factor income; bonds and interest deposits from interest income; and corporate shares and equity from divi- dends received.14 Mortgage debt and other forms of debts have been recorded consistently in the NIDS and other house- hold surveys, but as we showed in section 2, the coverage of liabilities remains partial and inconsistent. As a result, rescaling debts to balance sheets totals may result in overestimating the number of individu- als with negative net worth and extrapolating implausibly high debt values. Instead, we follow a mixed method: we assume that the mortgage debt from the huosehold balance sheet is distributed proportion- ally to the value of the house of mortgagors in the surveys, and that other forms of debts are distributed proportionally to the consumption of those declaring having contracted debts. These are conservative assumptions, as mortgages and other forms of debt are likely to be more unequally distributed than house values and consumption respectively. 14In the case of pension assets, we follow the approach proposed by Saez and Zucman (2016) in allocating them to wage earners and pensioners so as to match their distribution recorded in the NIDS. In our case, we assume that 75 per cent of pension assets belong to formal wage earners proportionally to pension contributions paid, and 25 per cent belong to pensioners proportionally to pension income received. As we show in the appendix (figure A5), this capitalisation technique applied to the NIDS data yields results which are very similar to those obtained from direct measurement. For life insurance assets, we assume that 50% belong to wage earners proportionally to factor income – the sum of wages, self-employment income and pension income – and that 50% belong to all other adults proportionally to factor income. This again reproduces well the distribution of life insurance assets reported in NIDS (see figure A6). 16 Table 7: Shares of household wealth held by groups in South Africa: survey-based results Bottom 50% Middle 40% Top 10% Top 1% Top 0.1% Direct measurement NIDS, wave 4 -3.3 % 18.4 % 84.9 % 41.3 % 9.7 % NIDS, wave 5 -0.5 % 16.9 % 83.6 % 40.2 % 8.6 % Rescaling NIDS, wave 4 -8.2 % 10.9 % 97.3 % 58.3 % 24.6 % NIDS, wave 5 -7.0 % 8.0 % 99.1 % 63.9 % 29.3 % Mixed approach NIDS, wave 4 -4.5 % 14.5 % 90.0 % 58.5 % 25.2 % NIDS, wave 5 -3.3 % 12.5 % 90.8 % 60.6 % 30.1 % PSLSD, 1993 -1.3 % 12.0 % 89.3 % 51.7 % 20.6 % IES, 1995 -5.1 % 15.3 % 89.8 % 50.6 % 23.7 % IES, 2000 -1.8 % 14.9 % 86.9 % 52.8 % 26.0 % IES, 2005 -0.2 % 13.6 % 86.6 % 54.2 % 28.6 % LCS, 2008 -8.0 % 14.0 % 94.0 % 52.3 % 22.4 % IES, 2010 -7.3 % 14.8 % 92.4 % 60.0 % 31.7 % LCS, 2015 -3.2 % 14.0 % 89.2 % 51.1 % 20.0 % Notes: The table compares estimates of the share of household wealth owned by the bottom 50 per cent (p0p50), the middle 40 per cent (p50p90), the top 10 per cent (p90p100), the top 1 per cent (p99p100 and the top 0.1 per cent (p99.9p100) obtained from household surveys using different methodological approaches. The unit of observation is the individual adult aged 20 or above. PSLSD: Project for Statistics on Living Standards and Development. IES: Income and Expenditure Survey. LCS: Living Conditions Survey. NIDS: National Income Dynamics Study. Source: authors’ computations based on data. 5.2 Measuring wealth inequality using survey data We start by looking at the distribution of personal wealth estimated from survey data. For all the fol- lowing results, we take the individual adult aged 20 or above as the unit of analysis.15 Table 7 compares estimates of the share of wealth held by the bottom 50 per cent (p0p50), the middle 40 per cent (p50p90), the top 10 per cent (p90p100), the top 1 per cent (p99p100) and the top 0.1 per cent (p99.9p100) ob- tained from direct measurement, rescaling and the mixed approach. The NIDS survey is the only survey collecting direct data on wealth and thus for which the results from the three methodologies can be compared. Other household surveys collect data on the value of owner-occupied housing and household debts, so they can be used to estimate the wealth distribution with the mixed approach.16 The first result which clearly stands out is that all approaches converge in revealing an extreme degree of wealth concentration. Regardless of the methodology, the bottom 50 per cent of the South African adult population is consistently negative, while the top 10 per cent is higher than 80 per cent in all surveys and 15We therefore provide “individual" wealth inequality series rather than series where wealth is divided among spouses (narrow equal-split), among adult household members (broad equal-split) or among both children and adult household members (per capita). The main motivation is that the tax microdata is only available at the individual level, so that applying equivalence scales to the survey data but not to the tax data would imply that the results are not comparable. Notice however that there are wealth components which are only measured at the household level – namely owner-occupied housing wealth, mortgage debt and non-mortgage debt. We split equally these components among adult members of the household. This is far from being a perfect solution, but overall wealth inequality is only moderately affected by changes in units of observation, especially at the top of the distribution. We report in the appendix (figure A9) how changes in equivalence scales affect survey-based top and bottom wealth shares. 16The PSLSD, IES and LCS surveys did not collect data on currency and non-interest deposits, so we impute their value from the NIDS by assuming that their distribution has remained constant, both in terms of overall concentration and conditionally to post-tax income. Given the small share of currency and non-interest deposits in aggregate wealth, this imputation does not affect our results. Also notice that the data on owner-occupied housing wealth in the IES and LCS surveys is very erratic, so we keep the rank of housing wealth reported in these surveys but force its distribution to match that observed in the GHS between 2002 and 2018. 17 Table 8: Shares of household wealth held by groups in South Africa: results from tax microdata and survey combined Bottom 50% Middle 40% Top 10% Top 1% Top 0.1% 2010 -6.8 % 16.6 % 90.2 % 57.3 % 30.0 % 2011 -6.4 % 16.7 % 89.8 % 57.0 % 29.3 % 2012 -5.3 % 16.5 % 88.9 % 57.2 % 33.5 % 2013 -4.0 % 16.0 % 87.9 % 56.3 % 32.1 % 2014 -3.0 % 16.2 % 86.8 % 54.5 % 29.9 % 2015 -2.9 % 16.0 % 86.9 % 55.0 % 29.2 % 2016 -2.9 % 16.2 % 86.7 % 53.5 % 27.5 % 2017 -2.5 % 16.9 % 85.6 % 54.7 % 29.8 % Notes: The table shows estimates of the share of household wealth owned by the bottom 50 per cent (p0p50), the middle 40 per cent (p50p90), the top 10 per cent (p90p100), the top 1 per cent (p99p100 and the top 0.1 per cent (p99.9p100) obtained from the income capitalisation method combining surveys and tax microdata. The unit of observation is the individual adult aged 20 or above. Source: authors’ computations based on data. methods. According to these results, wealth inequality in South Africa appears to be substantially larger than in any other country for which relatively reliable data is available (see below). The second result is that there are some important differences in the results obtained from the three different approaches, especially at the top of the distribution. Direct measurement in the NIDS implies a top 0.1 per cent share below 10 per cent, more than twice lower than most of the results obtained from rescaling or the mixed approach. This is mainly due to the very poor coverage of other financial assets in the NIDS, which are particularly concentrated at the top end of the wealth distribution. Rescaling financial assets to balance sheets totals or capitalising income flows corrects for this micro-macro dis- crepancy. Rescaling the value of assets and liabilities increases wealth inequality significantly compared to the mixed approach. This is mainly due to the fact that blowing up debts to balance sheets totals cre- ates a large number of households with strongly negative net worth (the bottom 50 per cent goes down by several percentage points). Finally, it is interesting to note that the mixed approach yields relatively close results across years and data sources: the top 10 per cent share lies between 85 per cent and 90 per cent and the top 1 per cent is estimated to be between 50 per cent and 60 per cent in most cases. This suggests that despite the fact that these households surveys were conducted using different sampling methods and questionnaires, capitalising reported income flows remains somehow an efficient method to broadly capture the structure of wealth concentration in South Africa. Yet, all these surveys are likely to suffer from misreporting or non-response, which implies a misrepresentation of income and wealth inequality at the top end. 5.3 Measuring wealth inequality using tax data We now turn to the estimation of the distribution of wealth obtained by combining surveys and tax data. As explained in section 4, tax microdata has the advantage of both better covering capital income flows and capturing with a greater degree of precision the levels and composition of incomes at the top end of the distribution. As a result, one may expect that our mixed method will lead to higher measured wealth inequality levels as compared to the capitalisation of income flows in household surveys. That being said, the important limitations of the tax data itself discussed above do prevent us from considering the tax-based estimates presented below as satisfactory. Income tax data in South Africa does not cover the full adult population: the matched IRP5-ITR12 panel only covers between 40 per cent and 42 per cent of adults over the 2010-2017 period. In order to get a reliable estimate of wealth inequality, we combine the tax data with household surveys in two steps. In a first step, we derive an income concept which is comparable between the two sources, which we 18 name “merging income", defined as the sum of gross wages, self-employment income, rental income, interest income and private pension income. We then merge the two data sources based on the exact rank of “merging income“ observed at the individual level. In a second step, we identify the quantile of the South African income distribution q starting from which reported merging incomes are higher in the tax data than in the survey data, and we assume that the tax data is more reliable than the survey data only above q. In practice, this implies keeping all variables from the survey data below q, and replacing all comparable variables from the tax data above q – namely wages, self-employment income, rental income, interest, dividends, private pension income, and contributions to pension funds. Between 2010 and 2017, we find q to be consistently located between the 70th and the 75th percentiles (see appendix figures A7 and A8).17 Table 8 shows the results obtained from combining the surveys and tax data and applying the mixed approach. Wealth inequality appears to be relatively similar when measured by combining surveys with tax data than when measured solely from the surveys available in similar years (the NIDS, the IES 2010 and the LCS 2015). The top 10 per cent wealth share stands at between 86 per cent and 90 per cent over the 2010-2017 period. The top 0.1 per cent share exceeds 30 per cent, compared to 20 per cent to 30 per cent in household surveys. This is a relatively surprising result, as one would expect the under- representation of top incomes in surveys to imply significantly lower levels of wealth concentration. A careful look at the particular structure of capital income concentration in South Africa can help solve this apparent paradox. The relative consistency between the two sources is mainly due to the fact that both in the surveys and the tax data, financial incomes (interest, dividends and rental income) are extremely concentrated, so that both sources imply attributing a substantial share of wealth – and in particular of tenant-occupied housing, bonds and shares – to the top 0.1 per cent of the distribution. 18 Thus the benefit of tax data in providing more reliability of estimates of wealth at the top end is undermined by the lack of data on capital incomes. 6 The distribution of wealth in South Africa: key results and comparative perspectives We now present some key figures on the levels, evolution and structure of wealth inequality in South Africa. Our preferred estimate of the wealth distribution is the one obtained from combining household surveys with the tax microdata and applying the mixed method. For the years preceding 2010, we use household surveys to estimate wealth inequality using the mixed method, and we assume that the under- representation of top wealth groups is similar to that observed during the 2010-2017 period. Finally, we combine the PSLSD, the IES and the LCS with labour force surveys to have a more consistent estimate of wage inequality and business income inequality – and therefore of pension and business assets –, as well as to have yearly estimates over the entire 1993-2018 period. We explain the methodology used to combine these various data sources in the appendix. We stress again that none of the results presented 17Our choice of a merging point based on an income concept differs slightly from the approach of Hundenborn, Woolard, and Jellema (2018), who rather derive a taxable income concept from survey data, and then keep the tax data above the filing threshold of taxable income. The main reason for merging our two datasets based on a broad income concept is twofold. First, our IRP5-ITR12 panel covers a large number of individuals who are below the filing threshold, given that all employers in South Africa are now required to file an IRP5 tax form for all their employees, regardless of their level of remuneration. However, as is emphasised in the SARS’ Tax Statistics, this rule was not followed strictly by all employers, so that the tax data cannot be considered to be representative of the universe of formal wage earners. In other words, our data covers relatively well the top of the distribution up to a certain point, below which it contains a mix of low- and middle-income wage earners. It seems therefore most useful to keep as many individuals as possible from the tax data, while removing those whose location in the distribution of income cannot be identified precisely, which is what our method does in a simple way. Secondly, defining taxable income remains a complex task, and it remains unclear whether this can be done with a sufficient level of precision and consistency, in particular given that surveys tend to not properly capture the top of the distribution. 18According to our matched survey-tax dataset, about half of rental income and 60 per cent of interest income were received by 0.1 per cent of the South African population in 2017. 19 Table 9: The distribution of personal wealth in South Africa in 2017 Number of adults Wealth threshold Average (2018 R) Average (2018 PPP $) Wealth Share Full population 35,400,000 R 326,000 $ 52,200 100% Bottom 90% (p0p90) 31,860,000 R 94,100 $ 15,100 14.4% Bottom 50% (p0p50) 17,700,000 R -16,000 $ -2,600 -2.5% Middle 40% (p50p90) 14,160,000 R 27,700 R 138,000 $ 22,000 16.9% Top 10% (p90p100) 3,540,000 R 496,000 R 2,790,000 $ 447,000 85.6% Top 1% (p99p100) 354,000 R 3,820,000 R 17,830,000 $ 2,860,000 54.7% Top 0.1% (p99.9p100) 35,400 R 30,350,000 R 96,970,000 $ 15,540,000 29.8% Top 0.01% (p99.99p100) 3,540 R 146,890,000 R 486,200,000 $ 77,920,000 14.9% Notes: The table shows the distribution of household wealth in South Africa in 2017. The unit of observation is the individual adult aged 20 or above. Wealth thresholds are in 2018 Rands. Source: authors’ computations based on data. below are fully satisfactory given the lack of proper data available to measure the distribution of wealth and in particular of financial assets in South Africa (see sections 3 and 4). The distribution of wealth in South Africa in 2017 Table 9 provides information on the number of adults, the entry thresholds, the average wealth and the share of wealth of various groups of the wealth distribution in 2017. Average wealth per adult in South Africa amounts to about 326,000 rands, or 52,200 dollars at purchasing power parity. This is three times higher than the national income per adult, which stands at about 110,000 rands (18,000 dollars) per year or 9200 rands (1450 dollars) per month. Average wealth varies hugely across the distribution. The bottom 50 per cent of the South African population have negative net worth: the levels of the debts that they owe exceeds the market value of the assets they own. The middle 40 per cent of the distribution – individuals located between the median and the 90th percentile – have a net worth more than twice lower than the average wealth per adult. Together, the bottom 90 per cent of the South African population own about 14 per cent of total personal wealth in the economy, while the remaining 86 per cent belong to the top decile. The average wealth of the bottom 90 per cent of the population is about four times lower than the national average, while the top 10 per cent has an average wealth about nine times higher than the average wealth per adult. Ownership is not only polarised between top and bottom wealth groups, it is also extremely concentrated within the top 10 per cent. The top 1 per cent of the South African adult population (350,000 individuals) own 55 per cent of aggregate personal wealth, and the top 0.1 per cent alone (35,000 individuals) own almost a third of wealth. The top 0.01 per cent of the distribution, amounting to some 3,500 individuals, own about 15% of household wealth, greater than the share of wealth owned by the bottom 90 per cent of the population consisting of 32 million individuals. They have an average wealth which is more than 1500 times that of the average South African adult, and 6000 times that of the bottom 90 per cent. The composition of personal wealth across the distribution The extreme degree of wealth inequality that we observe is in large part driven by the relative exclusion of poorer wealth groups from any form of wealth accumulation, and by the concentration of all forms of assets at the top end of the distribution. Table 10 provides some insights into this polarisation by showing the share of different types of assets held by wealth groups across the distribution. The top 10 per cent own more than 55 per cent of all forms of assets, including pension assets, housing wealth, business assets and currency, notes and coins. They own more than 99 per cent of all bonds and stock held in the economy. The top 1 per cent alone holds more than a tenth of all forms of assets and as much as 90 per cent of bonds and corporate shares. Currency and housing wealth are the least concentrated 20 Table 10: Share of total assets held by wealth group by asset class, 2017 Currency Business assets Housing Pensions / life insurance Bonds & Stock Bottom 90% (p0p90) 37.3% 40.4% 41.2% 36.2% 0.2% Bottom 50% (p0p50) 9.7% 1.4% 14.0% 5.3% 0.0% Middle 40% (p50p90) 27.7% 39.1% 27.2% 30.9% 0.2% Top 10% (p90p100) 62.7% 59.6% 58.8% 63.8% 99.8% Top 1% (p99p100) 10.6% 41.9% 27.8% 14.1% 95.2% Top 0.01% (p99.99p100) 1.5% 13.4% 8.5% 2.1% 62.7% % of total assets 0.6% 3.6% 28.8% 32.5% 34.6% Notes: The table shows the shares of different types of assets held by specific wealth groups in 2017. The unit of observa- tion is the individual adult aged 20 or above. In 2017, the top 1 per cent of South Africans in terms of net worth owned 95 per cent of the bonds and corporate shares in the economy. Bonds and shares represented 34.1 per cent of total household assets in the economy at this date. Figures may not add up due to rounding. Source: authors’ computations based on data. form of wealth, but low wealth groups only possess a small share of them: the bottom 50 per cent of the wealth distribution own about 10 per cent of currency, notes and coins, and less than 15 per cent of housing assets. Figure 3 gives another view of the link between forms of asset and wealth inequality by showing the portfolio composition of percentiles in the wealth distribution. Currency, notes and coins are the main form of assets held by poorest South African adults, while owner-occupied housing, pensions and life insurance form the majority of assets for most remaining income groups within the bottom 90 per cent. Business assets represent a small share of portfolios for the upper-middle class. Bonds and stock, finally, represent a large share of wealth for the top 1 per cent and the bulk of assets of wealth groups within the top 0.1 per cent. Wealth and age19 How does wealth change during the life cycle, and to what extent wealth accumulation and reduction throughout the lifetime account for wealth inequalities? Figure 4 shows a stable relationship between age and average wealth over the 2012-2017 period. Average net worth rises significantly and linearly between ages 20 and 55: individuals aged between 20 and 25 have an average net worth lower than 25 per cent of the national average, while those aged between 50 and 55 are between 50 per cent and two times wealthier than the average adult. Average wealth then stabilises between ages 50 and 65 and decreases slightly for older individuals, but still remains more than 50 per cent higher than the national average for individuals older than 75. Interestingly, this pattern is almost perfectly similar to that found in the case of France (see Garbinti et al. 2017, figure 5). While average wealth does vary significantly across age groups, age does not explain the observed levels of wealth concentration. Top wealth shares are almost perfectly similar within each age group than in South Africa as a whole: the share of wealth held by the top 10 per cent exceeds 85 per cent, and the top 1 per cent share is higher than 55 per cent, whether one restricts the analysis to those aged between 20 and 39, between 40 and 59, or older than 60 (figure 5). This apparent paradox can be better understood when directly comparing differences across age groups to differences across wealth groups in the distribution. The average wealth of those aged 20 to 25 was about 6.5 times lower than that of those aged 75 or above in 2017. In comparison, the average wealth of the top 10 per cent was about 30 times higher than the average net worth of the bottom 90 per cent of the distribution. Our results therefore point to inequalities in access to wealth accumulation across the life cycle – via income inequality, debt and savings patterns 19There are other important categories to investigate in the context of wealth inequality in South Africa. Although the tax data is more complete, it has less covariates than the surveys, therefore, given our methodology, we are restricting our decomposition to age. We leave gender, race and other related categories for future work. 21 Figure 3: The composition of assets by wealth group in 2017 0 10 20 30 40 50 60 70 80 90 100 Sh ar e of a ss et s (% ) 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 9599 99.9 99.99 Wealth group (percentile) Currency, notes and coins Owner-occupied housing Tenant-occupied housing Business assets Pension / life insurance Bonds and stock Notes: The figure shows the composition of assets of various groups in the distribution of household assets in South Africa in 2017. The unit of observation is the adult aged 20 or above. The results come the harmonised survey data file, and wealth is split equally among adult members of the household, except for the top 1 per cent and above for which the individual data built from the combined survey and tax microdata are used. Source: authors’ computations based on data. – as well as inequalities in access to inheritance as being the primary drivers of the high levels of wealth concentration observed in South Africa.20 Long-run trends and comparative perspectives We conclude this paper by bringing together our South African wealth inequality series with comparable data on other countries. For all the countries outlined below, corresponding studies followed the Distri- butional National Accounts methodology by combining all available micro and macro data sources to distribute household wealth (see Alvaredo et al. 2016). They are therefore directly comparable to our estimates. Figure 6 plots the evolution of the share of wealth accruing to the top 10 per cent in South Africa, China, Russia, India and the United States.21 The top 10 per cent wealth share has risen in all 20Notice that the estimates presented here correspond to individual series, rather than to “equal-split" series where wealth would be split equally among household adult members. In practice, splitting wealth among household members would imply redistributing wealth to younger individuals, thereby making the wealth-age profile less steep. If anything, this reinforces our argument that age is not a primary determinant of wealth inequality in South Africa. 21Notice that the wealth shares presented here for South Africa are based on individualisation of assets, given the difficulty to split wealth equally among household members or spouses after combining surveys and tax data. In the capitalised survey series, moving from individual series to broad equal-split series decreases slightly the top 10 per cent share by 4-5 percentage points, and has a more limited effect on top 1 per cent and top 0.1 per cent wealth groups (see figure A9). This implies that our series are not perfectly comparable to that of other countries, which generally split wealth equally among spouses (narrow equal-split). From available evidence on the differences between individual, broad equal-split and narrow equal-split series, we can except top shares in narrow equal-split series for South Africa to be lying between the broad equal-split and the individual- based results, so the top 10 per cent share would be lower by between 1 and 3 percentage points (see Blanchet, Chancel, and 22 Figure 4: Average wealth by age relative to average wealth per adult, 2010-2017 0 25 50 75 100 125 150 175 200 225 250 Av er ag e w ea lth (% o f n at io na l a ve ra ge ) 20 25 30 35 40 45 50 55 60 65 70 75+ Age 2012 2013 2014 2015 2016 2017 Notes: The figure shows the mean net worth of South African adults by age group relative to the national average. The unit of observation is the individual adult aged 20 or above. Source: authors’ computations based on data. Figure 5: Wealth inequality within age groups, 2010-2017 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110 Sh ar e of w ea lth (% ) 2010 2011 2012 2013 2014 2015 2016 2017 Top 10% (20-39 yr) Top 1% (20-39 yr) Top 10% (40-59 yr) Top 1% (40-59 yr) Top 10% (60+ yr) Top 1% (40-59 yr) Notes: The figure shows top 10 per cent wealth share and the top 1 per cent wealth share estimated when splitting the South African population into three age groups (20-39 years old, 40-59 years old, and 60+ years old). The unit of observation is the individual adult aged 20 or above. Source: authors’ computations based on data. 23 these countries, while in the long run it has remained more stable in South Africa, increasing between 2005 and 2010 but gradually moving back to its level of the early 2000s since then. Wealth concen- tration has remained substantially higher in South Africa since the beginning of the 1990s than in any other country for which comparable data is available. The South African top 10 per cent wealth share has fluctuated between 80 per cent and 90 per cent during the 1993-2018 period, while it has remained below 75 per cent in the US, 70 per cent in Russia and China, 65 per cent in India and 55 per cent in France or the United Kingdom. The same result holds for the top end of the distribution: the top 1 per cent wealth share was 55 per cent in South Africa in 2017, compared to 43 per cent in Russia, 39 per cent in the United States, 31 per cent in India, 30 per cent in China and less than 25 per cent in France and the UK (figure 7). In terms of trends, our results suggest that wealth inequalities in South Africa have remained stable at very high levels since 1993. Two facts are however worth noticing. First, wealth concentration seems to have rapidly increased between 2005 and 2008 before slowly coming back to its long-run level between 2009 and 2017. This short-run dynamic was in large part due to the strong fall in the bottom 90 per cent share driven by the boom and bust in mortgage advances in the 2000s, which temporarily drove a higher share of households into negative net worth. Between 2004 and 2008, in particular, mortgage debt increased from 9 per cent of net household wealth to almost 15 per cent, and decreased back to 9 per cent in 2018 (see figure A4). This temporary fall in bottom wealth shares driven by expanding debts mirrors that observed in the US at about the same period (figure A3). In the appendix, we discuss in greater detail the importance of household debt in South Africa and how it explains why bottom wealth shares have remained significantly negative throughout our period of interest. We also show that the concentration of household assets has remained remarkably constant (i.e. excluding debts from the analysis removes virtually all fluctuations across the period): the top 10 per cent share of assets has remained at about 80 per cent between 1993 and 2018 (see appendix figure A13). A second result which appears from our long-run series is that while the top 10 per cent share has remained broadly stable, there seems to have been a slight increase in inequality within the top 10 per cent. Between 1993 and 2017, the top 1 per cent share increased from 54 per cent to 57 per cent and the top 0.1 per cent share from 22 per cent to 31 per cent (figure A2). This is likely to be due to two factors: the increase in the share of non-pension financial assets from 19 per cent to 24 per cent of net household wealth between 1992 and 2018, and the increase in wage inequality in South Africa during this period – which indirectly affected the distribution of pension assets. That being said, we should stress that the low quality and the important issues regarding the comparability of the household surveys conducted during this period do not allow us to conclude to an increase in wealth inequality with a high degree of certainty. The best we can say with a certain level of confidence is that there is no evidence that wealth inequality in South Africa has decreased since the end of apartheid, and that South Africa remains significantly more unequal than any emerging or developed country for which good-quality data on the distribution of wealth is available throughout the world. Gethin 2019, for a longer discussion of the impact of different equivalence scales on inequality). Changing equivalence scale will therefore have no consequence on our main conclusions. 24 Figure 6: South African wealth inequality in comparative perspective: Top 10 per cent wealth share 30 40 50 60 70 80 90 100 Sh ar e of h ou se ho ld w ea lth (% ) 19 80 19 82 19 84 19 86 19 88 19 90 19 92 19 94 19 96 19 98 20 00 20 02 20 04 20 06 20 08 20 10 20 12 20 14 20 16 20 18 South Africa Russia United States India China France United Kingdom Notes: The figure compares the top 10 per cent wealth share in South Africa to that of other countries. The unit of observation is the individual adult aged 20 or above. Wealth is individualised (South Africa) or split equally among adult household members (other countries). Source: authors’ computations based on data for South Africa; World Inequality Database (http://wid.world) for other countries. Figure 7: South African wealth inequality in comparative perspective: Top 1 per cent wealth share 0 10 20 30 40 50 60 70 Sh ar e of h ou se ho ld w ea lth (% ) 19 80 19 82 19 84 19 86 19 88 19 90 19 92 19 94 19 96 19 98 20 00 20 02 20 04 20 06 20 08 20 10 20 12 20 14 20 16 20 18 South Africa Russia United States India China France United Kingdom Notes: The figure compares the top 1 per cent wealth share in South Africa to that of other countries. The unit of observation is the individual adult aged 20 or above. Wealth is individualised (South Africa) or split equally among adult household members (other countries). Source: authors’ computations based on data for South Africa; World Inequality Database (http://wid.world) for other countries. 25 http://wid.world http://wid.world 7 Conclusion This paper presented a first attempt to systematically compare income and wealth reported in surveys and tax data to official macroeconomic statistics in South Africa, and to assess several methods to correct the micro-macro gap in the measurement of households’ net worth. Our analysis has revealed two main findings. Firstly, the data sources available to measure wealth in South Africa remain largely unsatisfactory. Re- ported housing wealth is substantially higher in household surveys than in balance sheets statistics, while most surveys cover very poorly business assets, financial assets and household debts. More importantly, two major limitations will have to be addressed in future research: the absence of any reliable source of the distribution of dividends received by households, and the lack of distributional data on the wealth held through unit trusts, which represents a substantial part of the net worth of top end groups in recent years. While such data exist – through the trust forms and dividends tax forms reported by taxpayers to the South African Revenue Service –, they have unfortunately not yet been made available to academic researchers. Access to such data will be crucial to understand not only the distribution of wealth, but also the processes by which wealth gets accumulated, transmitted and redistributed in the economy. Secondly, all data sources do suggest that wealth inequality in South Africa is the highest among all countries for which data is available. The top 10 per cent own more than 85 per cent of wealth, and the top 0.1 per cent at least 25 per cent. Drawing on household surveys, we find no evidence that wealth concentration has decreased since apartheid. This finding calls for further research on the dynamics of asset transmission and on the factors accounting for the persistence of wealth inequality in the post- apartheid era. 26 References Alstadsæter, A., Johannesen, N., and Zucman, G. (2018). ‘Who owns the wealth in tax havens? Macro evidence and implications for global inequality’. Journal of Public Economics, 162(1): 89–100. Alvaredo, F., Atkinson, A., Chancel, L., Piketty, T., Saez, E., and Zucman, G. (2016). Distributional National Accounts (DINA) guidelines: Concepts and methods used in WID.world. (WID.world Working Paper 2016/2) Aron, J., Muellbauer, J., and Prinsloo, J. (2006). Estimating household-sector wealth in South Africa. (South African Reserve Bank Working Paper) Aron, J., Muellbauer, J., and Prinsloo, J. (2007). Balance Sheet Estimates for South Africa’s Household Sector from 1975 to 2005. (South African Reserve Bank Working Paper) Atkinson, A. B., and Harrison, A. J. (1974). ‘Wealth distribution and investment income in Britain’. Review of Income and Wealth, 20(2): 125–142. Berg, S., and Louw, M. (2005). ‘Changing patterns of South African income distribution: towards time series estimates of distribution and poverty’. South African Journal of Economics, 72(3): 546–572. Bharti, N. K. (2018). Wealth inequality, class and caste in India, 1961-2012. (WID.world Working Paper 2018/14) Blanchet, T. (2016). Wealth inequality in Europe and in the United States: estimations from surveys, national accounts and wealth rankings. (Paris School of Economics Master Thesis) Blanchet, T., Chancel, L., and Gethin, A. (2019). How Unequal Is Europe? Evidence From Distribu- tional National Accounts, 1980-2017. (WID.world Working Paper 2019/6) Boserup, S. H., Kopczuk, W., and Kreiner, C. T. (2016). Intergenerational Wealth Formation over the Life Cycle: