Earnings and Earnings Inequality in South Africa: Evidence from Household Survey and Administrative Tax Microdata from 1993 to 2020

2.1 Trends in wage and earnings inequality

Wage inequality in the rich world has generally increased over the last 30 years (Acemoglu & Autor (2011); OECD (2011); Autor (2019)). But there has been a recent decline in wage inequality in China in urban households (Zhang, 2021) and wage inequality has also declined in Latin America over the last 20 years (Messina & Silva, 2021) and urban India in the 2010s (Khurana & Mahajan, 2020).

The key driver of increasing inequality in rich countries in the last 50 years was thought to be skill biased technological change (SBTC) (Acemoglu & Autor, 2011). But Autor et al. (2006), Goos and Manning (2007) and many subsequent papers have highlighted increasing employment and wage polarization as other drivers of increasing inequality. Polarization means rising employment and wages in low and high skilled jobs and declines in middle skilled jobs, which cannot be explained by the canonical SBTC model (Acemoglu & Autor, 2011). Institutional factors, such as declines in minimum wages and unionization rates, have also been found to be partially responsible for increases in wage inequality in rich countries (Farber et al., 2021; Fortin et al., 2021).

There are important links between wage and earnings inequality in rich countries and developing countries. Khurana and Mahajan (2020) note that job polarization in rich countries results from both automation in rich countries and off-shoring, meaning that some middle-skilled jobs are destroyed whilst others are shifted to developing countries, where they might actually be fairly high up the earnings distribution. Middle-skilled jobs could also be destroyed by automation in developing countries, although in low-income countries most jobs are unlikely to be automated (Maloney & Molina, 2019).

The recent literature on job polarization in developing countries finds little evidence that job polarization is occurring (Maloney & Molina, 2019; Martins-Neto et al., 2021). Thus, there is no reason to think that there should be wage or earnings polarization in most developing countries either. But evidence on wage polarization is more limited, due to a relative lack of comparable wage or earnings data across countries. Two recent papers do shed some light on this—Messina and Silva (2021) find no evidence of wage polarization in 13 Latin American countries between 1995 and 2015 and Khurana and Mahajan (2020) find no evidence of earnings polarization in urban, non-agricultural employment in India between 1983 and 2017.

2.2 Challenges when measuring earnings inequality

Researchers have identified several challenges that arise when measuring earnings or wage inequality in household surveys and have shown that these matter for the conclusions that can be drawn. These challenges include item and unit non-response, measurement error and under sampling or under representation of high-income individuals (Acemoglu & Autor, 2011; Bollinger et al., 2019; Lustig, 2020; Meyer et al., 2015; Ravallion, 2022). One check and possible solution to some of these issues is to use administrative data from tax or social security records, which Meyer et al. (2015, p. 201) argue ‘provides one potentially fruitful avenue for improving the quality of survey data.’ Solutions to household survey non-response, including reweighting (for unit non-response) and imputation (for item non-response) are important, and their efficacy has been interrogated using administrative data. In this Section I focus on item non-response, imputation and the use of administrative data as checks on survey results, given their relevance to measuring earnings inequality in South Africa using household survey and administrative data, which I show in my analysis in Section 4.

Item non-response, meaning that no information is provided by a respondent to a specific question, is a concern in household surveys. Bollinger et al. (2019) highlight that item non-response to the US Current Population (household) Survey (CPS) earnings questions doubled between the mid-1980s and mid-2010s to around 25 percent and Campos-Vazquez and Lustig (2019) show that the item non-response rate for the earnings questions in the Mexican labor force survey rose from 5 percent in 2002 to 30 percent by 2017. Linking the CPS responses to Social Security administrative data, Bollinger et al. (2019) show that item non-response to earnings questions in the CPS is not missing at random, i.e. that item non-response is not only a function of individual observable characteristics, but of the level of earnings itself. They show that item non-response is highest in the tails of the distribution.

Bollinger et al. (2019) note that the usual solution to item non-response to earnings questions is to impute earnings for non-responders, which in the CPS is done using hot deck imputation. Hot deck imputation means that the earnings of those who do respond are “donated” or “allocated” to observably similar individuals who do not give their earnings. Using the linked data they find that the census bureau imputation results in underestimates of inequality and that “between one-third and one-half the difference in inequality measures between the CPS and administrative data is accounted for by [item] nonresponse in the CPS” (Bollinger et al., 2019, p. 2148).

Bollinger et al. (2019) helpfully compare administrative and survey data to learn about the potential weaknesses of household survey data. However Meyer et al. (2015, p. 216) note that “Comparing weighted microdata from surveys to administrative aggregates … also has some important limitations including possible differences between the survey and administrative populations.” One form of this problem is that tax data often include all individuals with earnings at any point in a tax year, whereas some household surveys only allow an estimate of the earnings distribution for those employed at the time of the survey—what Acemoglu and Autor (2011, p. 1049) call a “point-in-time” measure. I show below that this is an important issue in my comparisons of survey and tax administrative data in South Africa.1 This is because the low employment rate and high worker flows in South Africa (Kerr, 2018) imply that there are many fewer individuals counted as employed when they are asked about employment in the last week (as South African household surveys generally ask), compared to the number of individuals who worked at any point during the year.

3.1 Household survey data

The household surveys used in this paper were all conducted by Statistics South Africa, except the 1993 Project for Statistics on Living Standards and Development (PSLSD), conducted by the Southern African Labour and Development Research Unit (SALDRU). The first source of survey microdata used is a version of the Post-Apartheid Labour Market Series (PALMS) (Kerr et al., 2019). PALMS is a publicly available compilation of nationally representative household surveys (The PSLSD, October Household Surveys, Labour Force Surveys and Quarterly Labour Force Surveys) with harmonized data on employment and earnings from 1993 to 2019.

For this paper I created my own version of PALMS using the publicly available code but include non-public Quarterly Labour Force Survey earnings data for 2011, 2012 Quarter 1, 2018, 2019 and 2020 Quarter 1. This non-public earnings data is comparable to the 2007 and earlier surveys in PALMS and to the other source of household survey data I use, the General Household Surveys, because it does not have the earnings imputations present in the public data. I also analyze the public QLFS earnings with imputations separately and compare to both the non-public QLFS in PALMS and the GHS.3

The other source of household survey data I use is the General Household Surveys (GHS) from 2002 to 2019 (Statistics South Africa, 2002–2019). Surprisingly the GHS has been unused by researchers as a source of earnings data in South Africa. It allows comparisons with the potentially unreliable public QLFS, it has very similar questions about earnings to the LFS and QLFS, and in some years has the same or overlapping sample clusters as the QLFS. In Appendix S1 I provide description of the earnings questions asked in the PALMS and GHS surveys, details about item non-response and how I process the earnings data, take into account the sample design and impute for item non-response.

3.2 Administrative tax data

The administrative data I use comes from the South African Revenue Service (SARS) for the 2011–2020 tax years (National Treasury and UNU-WIDER, 2022). The tax year runs from March in the previous calendar year until the end of February.4 Any individual earning more than R2,000 per year who works in a firm registered for Pay as You Earn (PAYE) tax should be issued an IRP55 certificate for each employer they worked for during the year. Individuals can thus have multiple certificates. The microdata contains information from each IRP5 certificate issued to all individuals. There is no top coding of earnings and all individuals are included (i.e. there is no sampling in the data I use). In Appendix S1 I provide further details of how I process the tax data.

I use two different groups of individuals in my analysis of the tax data. The first is only those employed in the first 2 weeks of the tax year (the first 2 weeks of March). I use this group because it closely aligns with the survey definition of employment, which is derived from questions that ask about employment in the week before the survey took place. The second group I use is all those employed at any point during the tax year, a much larger group. I show in my analysis that these two groups give quite different results about the extent of differences between the surveys and tax data, because one is a point in time measure similar to the household surveys and one is not (Acemoglu & Autor, 2011).

One important issue is that around two thirds of all pension contributions were not included on IRP5 certificates before the 2017 tax year, as they were not tax deductible (Redonda & Axelson, 2021). This means that earnings that includes pension contributions is not comparable before the 2017 tax year when compared to 2017 and later tax years. For most of my analysis I thus use earnings that includes all parts of earnings except pension contributions. In Appendix S1 I provide some key results that include pension contributions.

3.3 IRP5 administrative tax data and household survey sample comparisons

To make comparisons between the household surveys and tax data one needs the same hypothetical population in both sources (Meyer et al., 2015). South African tax legislation requires that IRP5 tax certificates are issued by employers to all employees in all firms registered for Pay as you earn (PAYE) tax, if the employee earns more than R2,000 a year (around $115 in 2022), a very small amount. There are no direct questions on employer tax registration in the QLFS but there are enough questions to produce a very comparable population of employees—QLFS employees who report that their employers contribute to the Unemployment Insurance Fund (UIF) on their behalf or who are in the public sector (public sector employers do not make UIF contributions), who have a written contract and who are not domestic workers. All private companies are required to register all their employees for the UIF so all the employees responding that their employers do make UIF contributions should also be issued tax certificates. I check this below.

From 2010 to 2019, the period over which I undertake survey and administrative data comparisons, the GHS only asks a question about whether the organization an employee works for is “in the formal sector (registered to perform activity)”, so I am forced to use this somewhat broad definition.

Figure A1 in the Appendix shows the number of individuals in each source of data using these definitions of the formal sector. The IRP5 group including all individuals employed at any point in the year is about 20 percent larger than the group employed only in the first 2 weeks. The QLFS formal sector and IRP5 population employed in the first 2 weeks of each tax year are similar from 2013 onwards, as they should be based on the closely comparable hypothetical populations present in each.6 This means that comparisons of the earnings distribution in these two populations should be validly comparing the same population. The GHS formal sector is much larger than the QLFS formal sector—by around 20–30 percent. This means that the comparisons between the GHS formal sector and either the QLFS formal sector population or the tax data population employed in the first 2 weeks of the tax year are not comparing the same two populations. Unfortunately, this is the best I can do given the limited questions asked in the GHS. Tables A1 and A2 in the Appendix show the sex and age breakdowns in the QLFS, SARS and GHS. All three show broadly similar levels and all show declines in the male share of employment. This suggests that the three groups are fairly similar, although as noted the GHS formal group is too large.

Once I undertake the data-cleaning discussed in the Appendix and use the QLFS formal sector definition, I should be left with a comparable population in the QLFS surveys and the IRP5 tax data population employed in the first 2 weeks of the tax year.

4.1 Earnings and earnings inequality in the household surveys

I first use the household surveys to describe trends in earnings and earnings inequality for all employed individuals (employees and self-employed, formal and informal) between 1993 and 2020 quarter 1, just before COVID-19 lockdowns and a shift from in-person to telephone surveying. The surveys conducted in the 1990s are not as reliable as those from 2001 onwards for several reasons. The sample frame used to select the sample until 1998 was from the (pre-democracy) 1991 census, where in some areas the census could not be conducted due to political violence (Christopher, 2001). The 1990s surveys also did not probe as carefully about informal employment and thus undercounted such employment (Casale et al., 2004), meaning that the earnings distribution is likely too optimistic in these surveys. The first two LFSs in 2000 and the third in 2001 counted too many workers in subsistence agriculture and informal self-employment, respectively (Kerr & Wittenberg, 2015).

The analysis includes earnings data from non-public versions of the QLFS from 2011, 2012 Q3, 2018, 2019 and Q1 2020. Because this non-public data includes unimputed earnings I can process this data in a similar manner to the surveys pre-2008 and the GHS—see Appendix Section 6.1.3. The results using the non-public QLFS data are very similar to the GHSs conducted at the same time but very different to the rest of the public QLFS, after 2012. I show in the next section that this is likely a result of poor-quality imputations undertaken by Stats SA.

In my analysis I show changes in five percentiles (10, 25, 50, 75, 90), and then several measures of inequality: percentile ratios, Gini coefficients and the variance of log earnings.7 Because of the under capturing of informal, low paid work in the surveys done in the 1990s, I believe the trends from the early 2000s onwards in PALMS and the GHS are more reliable but that the results from the 1990s are not comparable to the later surveys. Figure 1 shows that the 10^th, 25th and 50^th percentiles rose substantially from the early 2000s until the mid-2010s in PALMS and the GHS. Both data sources differ markedly to the public QLFS from 2012 onwards.8 Figure 2 shows increases in the 75^th and 90^th percentiles from the early 2000s until 2012 in the GHS and PALMS and are stagnant over that but the public QLFS shows large declines after 2012.

The relative changes for all five percentiles are shown in Figure A2 in the appendix, normalizing to one in 2002 for all three data sources. From 2002 onwards the GHS and PALMS both clearly show the same pattern—the largest increases have been at the 10^th and 25^th percentiles, with smaller increases for the median, 75^th and 90^th percentiles. The GHS show no growth or even declines in real earnings from around 2013 onwards, except for a little growth at the 10^th percentile, and PALMS shows growth between 2012 and 2020 only for the 90^th, 10^th and 25^th percentiles. The public QLFS data from 2010 onwards shows a very different trend to the GHS and PALMS data—very large declines at all five percentiles, with the largest being at the 10^th percentile—where the GHS and PALMS show strong increases. Bassier and Woolard (2021) find negative growth in the QLFS for most of the bottom 95 percent of the distribution between 2010 and 2015, which Figures 1 and 2 and A2 show is likely due to the public QLFS earnings imputations, since no such pattern is observed in the GHS and PALMS (which includes the non-public QLFS data).

One important question is whether these trends provide any evidence of earnings or wage polarization. In rich countries this is generally taken to mean that earnings or wage growth was highest at the top, lower at the bottom and that the lowest growth was in the middle (Autor et al., 2006). In PALMS since 2002 the largest increases for monthly earnings have been at the 10^th and 25^th percentiles, the median the smallest, the 90^th percentile growth higher than the median and the 75^th percentile between the median and 90^th percentile. The GHS results are similar, although the growth the median and 75^th percentile growth is about the same. I would thus argue that there is not strong evidence for earnings polarization since 2002. Figure A3 in the Appendix shows that broadly similar patterns hold when using PALMS data for monthly earnings for employees only and for hourly wages for employees only, implying that there is also not strong evidence for wage polarization between 2000 and 2012. This contradicts Bhorat et al. (2023), who use the public QLFS earnings data for 2000 and 2015 to argue for wage polarization, and then to look for explanations.

I now use the data shown in Figures 1 and 2 to describe inequality trends and levels by showing the p90/p50 and p50/p10 percentile ratios. Figure 3 shows that between the early 1990s and Q1 2020, the top of the distribution consistently moved further away from the middle, and that the bottom consistently moved towards the middle. The changes are large in magnitude—someone at the median in the early 2000s when the surveys became more comparable earned about five times more than the person at the 10^th percentile but only around 3.3 times more in Q1 2020—as a result of the much larger increases at the 10^th percentile compared to the median, as shown above. And someone at the 90^th percentile earned around four times someone at the median in the early 2000s but around 5.5 times more by 2020. Again, the public QLFS trends look implausible whereas the GHS and PALMS trends look more reasonable. Figure A4 in the Appendix shows the p50/25 and p75/p50 percentile ratios. The conclusions are broadly the same. The key result is that the bottom of the earnings distribution has moved closer to the middle throughout the post-Apartheid period, whilst the top of the distribution moved away from the middle, a finding consistent with Wittenberg (2017b), who used household survey data up to 2011.

Figure 4 shows the Gini coefficient for earnings and variance of log earnings from 1993 to 2020 using PALMS, public QLFS and the GHS. Both measures show that earnings inequality is extremely high in South Africa. Between the early 2000s and 2019 PALMS shows an increase in the Gini of around two points but the GHS Gini is more erratic, perhaps because sparsity in top incomes can result in volatility in estimates over time (Jenkins, 2017), and the increase to 2019 is larger than in PALMS. The variance in log earnings in both PALMS and GHS has declined since the early 2000s when the surveys became more reliable. The public QLFS Gini coefficient and the variance in log earnings both increase very dramatically between 2012 and 2013 and the subsequent trend is much more erratic and highly implausible.

Broadly speaking then, the analysis thus far has shown that earnings inequality in South Africa is still extremely high, making it one of the most unequal countries in the world. But the GHS and PALMS show that since the early 2000s, when the surveys became more comparable, inequality has declined when using the variance of log earnings and has increased only slightly when using the Gini coefficient (the GHS Gini is volatile, although not as much as in the public QLFS). During this period the bottom has caught up to the middle whilst the top has moved further away from the middle.

The implausible trends in earnings distributions, the Gini and variance of log earnings in the public QLFS, as well as the differences between the QLFS public data on the one hand, and the GHS and PALMS data on the other, suggest that the public QLFS earnings imputations may be to blame. In the next section I show further evidence that this is likely true, by merging the non-public and public QLFS earnings data for the same individuals in the years where I have non-public QLFS data.

4.2 QLFS earnings imputations

Earnings imputation is the usual solution to earnings item non-response. Bollinger et al. (2019) highlight that accounting for item non-response is a crucial part of measuring earnings inequality, and that not doing so can result in under-estimates of inequality. Public QLFS earnings data is imputed but never discussed in public documents. Kerr and Wittenberg (2021) obtained a non-public Stats SA document that indicated that hotdeck imputation was used, with two sex, four race, seven education and three occupation categories used to find donors. Kerr and Wittenberg (2021) obtained and used unimputed earnings data for 2011 and 2012 from Statistics South Africa and concluded that the imputations were of poor quality- because the estimated union wage premium was much lower in the public data and the variance of earnings in the union and non-union sectors in the public data was highly implausible. The analysis in the previous section showed that in 2010–2011 the earnings distribution of the public QLFS data was not too dissimilar to PALMS and the GHS, but that it got further apart and that by 2020 the two sources were very different, suggesting the imputation may have got worse over time. I investigate this further in this section.

In Tables 1 and 2 I show the earnings response types in the public and non-public microdata in 2011, 2012 Q3, 2018, 2019 and 2020. According to the QLFS questionnaire, individuals can either give an “amount,” a bracket, a “don't know” or a refusal. There are also some “other missings” who are employed but are in none of the other four categories. Tables 1 and 2 show that in the non-public data there are (as expected) “amounts,” brackets, “don't knows” and refusals, but also that 3 percent of the employed have “Other missing.” In 2011/2 only around 56–57 percent of respondents gave an “amount,” 19 percent gave a bracket and 24 percent had no earnings information.

In 2018, 2019 and Q1 2020 the “amount” response rate dropped to 50 percent and roughly 20 percent gave a bracket, meaning that there was no earnings information for about 30 percent of the employed sample. Vermaak (2012) reports that there was no earnings information for around 6–10 percent of the employed in the LFSs that ran in the 2000s until 2007. This means that non-response roughly tripled from the early 2000s to 2012 and increased a further 25 percent between 2012 and 2020. This increase is larger than the earnings item non-response increase in the US CPS reported by Bollinger et al. (2019, p. 2145) for a nearly 30-year period, but similar to the increase in the Mexican labor force survey reported by Campos-Vazquez and Lustig (2019) between 2000 and 2017.

Tables 1 and 2 show different patterns in the public data for the different years. In 2011 there are almost all amounts, no “bracket onlys,” “don't knows” or refusals- meaning full imputation was undertaken for these categories. There is a small group of “other missings.” In 2012 Q3 in the public data 7 percent of responses are indicated as refusals but this percentage is lower than in the non-public data. In 2018, 2019 and Q1 2020 there are also no don't knows in the public data, which suggests that “don't knows” are imputed but that refusals are not. But this conclusion is incorrect, as I show next.

To undertake further analysis, I merge the public and non-public earnings data for each respondent using the household ID and person number. In 2018–2020 the merge is perfect- the person and household ID numbers merge perfectly, all individuals are merged and they have the same sex, race, age, industry, occupation and union status in the public and non-public datasets. In 2012 about 0.25 percent of employed individuals do not merge into the public data, but all those in the public data are merged and all those that were merged have the same values for the variables sex, race, age, industry, occupation and union status. In 2011 all observations in both non-public and public data merge but I was only provided with the earnings variables, the household ID and the person number, so I cannot check whether the merge results in the same sex, race etc. The imputed amounts in the public data are in monthly earnings, so I create a monthly earnings value from the non-public data using the period of reporting earnings (monthly, weekly etc.) and the earnings amount reported by the respondent in that period.9

To explore whether the imputations are to blame for the differences between public and non-public QLFS I first rule out that the earnings values in the public data for those who the non-public data shows gave actual amounts are problematic. Figure A5 in the Appendix shows that the levels and trends for the amount responses in the public data are very similar to the amount responses in the non-public data for 2011, 2012, 2018 and 2019. This implies that the differences between the QLFS public and non-public earnings distributions are driven by the imputations for bracket responders, refusals, “don't knows” and other missings.

Tables 3 and 4 below also use merged data and compare the response types for the same individuals in the public and non-public data. Tables 3 and 4 show that the imputation methods are very poor.10 The main concerns are that most bracket responders are imputed earnings outside the bracket they reported in the public data, that after 2011 many bracket responders are “imputed” as refusals in the public data, a substantial share of don't know responses are “imputed” as refusals in the public data and only some true refusals are imputed, while others are shown as refusals in the public data. My conclusion from this analysis is that the bad quality imputations that I have described and the very large differences between the earnings distributions and the inequality measures in non-public and public data that result from this render the data unusable for inequality analysis.

The imputation weaknesses are very important for a range of other highly relevant policy issues. For example, the monthly earnings in a full time job paying the national minimum wage (R20 an hour for most jobs) is at the 33^rd percentile of the formal sector employee earnings distribution in the Q1 2020 public data whilst it is at the 21^st percentile when using the non-public data and my own hotdeck imputation, a massive difference that translates to around 1.2 million formal sector workers incorrectly assumed to be earning less than the full time monthly equivalent of the minimum wage.

4.3 Earnings and earnings inequality in the IRP5 data and household surveys

Having described earnings and earnings inequality for all employed in the household surveys and documented the weaknesses of the QLFS public earnings data, I now use administrative tax data to estimate earnings and earnings inequality distribution and trends over time for employees. The tax data is available from the 2011 to 2020 tax years, and so I limit the analysis of the General Household Surveys and QLFSs to this period. I include the public QLFS earnings data even though I have shown above that it is unreliable, since it has been used before to describe inequality and to show again that trends and levels differ dramatically compared to the tax data, the GHS and the non-public QLFS data.

To make the tax and household survey data more comparable I include only employees in the formal sector.11 As noted above, and shown in Figure A1, the population of employees in the tax data employed in the first 2 weeks of the year and the number of QLFS formal sector employees are very similar from 2013 onwards. This, and the results from Tables A1 and A2, means that the hypothetical populations in the tax and QLFS formal sector data are similar. However, the number of GHS formal sector employees is too large, due to the limited number of questions available to determine whether employed individuals would be in the hypothetical population that was issued a tax certificate. This means that the QLFS and GHS “formal sector” populations are not as comparable and may explain the larger differences between the results of the two surveys in this section compared to the results in Section 4.1.12

Figure A1 also shows that the population of employees in the tax data employed at any time in each tax year was much higher than the estimate from the QLFS of total formal sector employees, and so the earnings distribution from this population is not directly comparable to the QLFS population estimates (or the GHS), which Acemoglu and Autor (2011) call “point in time” estimates. Nevertheless, I include the earnings distribution for the tax data population employed at any point in the year to demonstrate that this distribution is very different to the one obtained for those employed in the first 2 weeks of the tax year, and that comparing a point in time measure from a household survey to the population employed in the tax data at any time of the year is incorrect.

Figures 5 and 6 show the 10^th, 25^th, 50^th, 75^th and 90^th percentiles in the surveys and the two different tax data populations. Figure A6 in the appendix shows the relative changes in these percentiles in the GHS and SARS data and Figure A7 shows the relative changes for the public and non-public QLFS data. Three key results stand out. The first is that the for the comparable populations—the non-public QLFS formal sector employees and the SARS group employed in the first 2 weeks of the tax year, earnings at the 25^th–90^th percentiles are much lower in the QLFS than in the tax data, and the gap has been widening at all five percentiles. This suggests that even the household surveys with a comparable population and reliable earnings data are under capturing earnings throughout the distribution, not just at the top, and that this has got worse over time. I previously compared the shares of formal sector employees earning less than the full-time monthly equivalent of the national minimum wage. The tax data implies that around 15 percent of employees working for tax registered firms earned less than the monthly full time equivalent of the hourly minimum wage in 2019, compared to 21 percent in the non-public QLFS with my imputations and 33 percent in the public QLFS in 2019.13

The tax data suggest that at the 25^th and 50^th percentiles formal sector employees saw real annual earnings growth of <1 percent, that at the 75^th and 90^th percentiles this was just over 1 percent and that annual real growth was around 3 percent at the 10^th percentile. In the GHS and non-public QLFS there were declines in annual earnings of around 1 percent a year at the 50^th percentile, roughly no growth at the 25^th percentile and then growth of less than 1 percent at the 10^th, 75^th and 90^th percentiles. The public QLFS implies declines of between 1 percent and 6 percent per year, with the largest decline for the 10^th percentile.

Bassier and Woolard (2021) showed much higher rates of income growth in tax data between 2003 and 2018 for the top percentiles, around 4–5 percent per year. There are several possible explanations for the differences between my results using tax data and theirs. The first is that for the period in which the tax data I use overlaps with theirs, they find growth rates of only around 2.5 percent per year. The second is that their data includes all incomes, including capital income, capital income is more important at the top of the distribution and Bassier and Woolard (2021) show that share incomes increased by 15 percent per year between 2003 and 2015 for the top percentile and 10 percent per year for the 95^th percentile.14 The third is the difference in datasets—their data are derived from tax filings that include the self-employed whereas mine include all employees issued an IRP5 certificate. Bassier and Woolard's percentiles are also calculated for the entire adult population, whereas mine are only for the employed in the formal sector, a much smaller population. This may also contribute to differences in the results.

The second key result is that the earnings percentiles in the tax data of the population employed in the first 2 weeks of the tax year, the population more comparable to the household surveys, are much higher than the distribution of those employed at any point in the year. This means that researchers using tax data on the population employed at any point in the year and comparing this to household survey “point in time” estimates (Acemoglu & Autor, 2011, p. 1049) are understating the differences between the survey and tax data.

The third key result is that the publicly available QLFS data again seems unreliable- the declines in the 10^th percentile and median are unbelievable. The non-public QLFS earnings data seems much better although it still substantially under captures earnings compared to the tax data, despite the two having very similar hypothetical populations.

Figures 5 and 6 show only five percentiles of the earnings distribution. In Figure 7 I provide a comparison of earnings throughout the distribution between the non-public QLFS Q1 2019 and the comparable group in the tax data—those employed in the first 2 weeks of the 2020 tax year (the first 2 weeks of March 2019).15 I rank all individuals in the SARS data and create a rank for the QLFS respondents using the survey weights,16 and then comparing earnings at each rank at which there is a survey respondent, following Wittenberg (2017c). The points are then joined using a local linear regression. A value of 1 means that the amount earned by the person in the tax data at that rank was 100 percent more than the person at the same rank in the survey. Figure 7 again shows that the differences between the QLFS and tax data are substantial. Under capturing is highest at the top ranks, is then roughly constant for the rest of the top 60 percent of earners and then declines over most of the rest of the distribution.17 Figure 7 again suggests that under capturing of earnings is not limited to the top of the distribution.

Having described and compared trends and levels in earnings across the different data sources, Figure 8 shows inequality in earnings using the p50/p10 and p90/p50 percentile ratios.18 The GHS, non-public QLFS and tax data show broadly similar trends—the bottom catching up to the middle and the top moving away from the middle. The public QLFS data produces the opposite trend in the p50/p10 ratio and larger increases in the p90/p50 compared to the other data sources.

Figure 9 shows the variance of log earnings and the Gini coefficient for formal sector employees in the surveys and tax data. The tax data and non-public QLFS are supposedly the most comparable. But the variance of log earnings in the tax data is 80 percent higher than in the QLFS at the start and around 40 percent higher at the end, as a result of it declining in the tax data and rising slightly in the non-public QLFS. The Gini is around 3–5 points higher in the tax data, and the gap also declines. The Gini in the GHS is higher and more erratic than in the non-public QLFS whilst the level and trend in the variance of log earnings are more similar to the non-public QLFS than for the Gini. The non-public QLFS trends again seem implausible.

As discussed in Section 3, the tax data analysis thus far has excluded pension contributions from the definition of earnings because of a change in tax policy that meant that before the 2017 tax year roughly two thirds of total pension contributions were not included on the IRP5 certificates that are the source of data for the analysis. In Figures A9–A11 I redo some analysis including pension contributions in earnings-showing the 75^th and 90^th percentiles over time, a version of Figure 7 comparing SARS and QLFS across the entire earnings distribution and the Gini coefficient and variance of log earnings. There is a clear shift up at the 2017 tax year in the P75 and p90 when all pension contributions are included and this shows up in a larger difference between SARS and QLFS in Figure A10. Figure A11 shows that changes in inequality due to including pensions are not large—the Gini increases by half a point and the variance of log earnings by about 5 percent from 2017 onwards when all pension contributions are included on IRP5 certificates.