Analysis of the Distributional Effects of COVID-19 and State-led Remedial Measures in South Africa

Figure 2 shows the distribution of household per capita disposable income by decile for March, April, May, and June 2020, in rands. Disposable income refers to incomes after the deduction of simulated personal income tax payments and UIF contributions, and having added all relevant simulated benefits. The deciles are deciles of disposable income for March 2020, and are held constant for the other three months.

Figure 2

Download asset Open asset

Mean disposable income fell for the wealthier deciles and ultimately increased for the poorer deciles. The first column in each decile (in green) shows the situation in March 2020 prior to the shock, while the results for the other three months are based on the shocked input dataset in which those in employment prior to the shock had been assigned different statuses. Here, the most notable change is the reduction in mean monthly household disposable incomes in the top (richest) deciles.

Figure 3 shows the change in mean monthly household disposable income by decile in rands. As can be seen, deciles 7–10 experienced a fall in disposable income in April, May, and June when compared with the baseline in March. The wealthiest (tenth) decile experienced the largest fall in disposable income. In contrast, in May and June deciles 1–6 experienced a rise in disposable income.

Figure 3

Download asset Open asset

The increase in disposable income in the lower deciles that is observed in Figure 3 is small in rand amounts, but when expressed as a percentage of March’s mean disposable incomes the change is more striking. This is shown in Figure 4, which shows that the mean disposable income of those in the first (poorest) decile increased by just over 100 per cent in April and by almost 200 per cent in May and June compared to March. As will be elaborated below, the notable increases for the lower deciles are a result of the introduction of social assistance for people of working age.

Figure 4

Download asset Open asset

Although the wealthiest (tenth decile) loses the most in absolute terms, the eighth and ninth deciles lose slightly more in relative terms.

Table 2 shows how the poverty rates changed across the four time points, using Statistics South Africa’s three poverty lines. It should be recalled that the input dataset for April, May, and June is held constant in the simulations and so the only drivers of any changes are the simulated policy responses to the pandemic and associated lockdown.

For each of the three poverty lines, the first row shows the poverty headcount ratio with all policies switched on (that is, taking into account both the set of policies that existed prior to the pandemic and the set of policies that were introduced as COVID-19 policies to mitigate the impact of the pandemic), but with receipt of the COVID-SRD benefit dampened to match reported numbers of beneficiaries. Using all three poverty lines, poverty is higher in April and May when compared to March. Poverty reached its height in April when only the TERS had been introduced and no changes had been made to the benefit system. For example, using the food poverty line, poverty rose from 0.206 in March to 0.263 in April, at which point over one-quarter of people in South Africa were below the food poverty line.

Notably for all three poverty lines, in June poverty fell to a level lower than in March. This is reflected in the poverty headcount and the poverty depth summary measure (Table 2). The two policy differences between May and June were the switch from CSG top-up payments to a dedicated Caregiver-SRD, and an improved roll-out of the COVID-SRD from 4.4 million beneficiaries in May to 5.1 million in June.

In Table 2, for each of the poverty lines the poverty headcount ratio is shown for a hypothetical scenario in which the COVID-19 policies are switched off (the row ‘All policies apart from COVID-19 policies’). These results are the same for April, May, and June as the input dataset and non-COVID-19 policies remain constant, but are shown for each month for completeness. In a hypothetical situation with no COVID-19 policies, it can be seen that poverty would have risen to 0.321 each month using the food poverty line (a 56 per cent increase from the baseline in March), and to 0.452 using the lower-bound poverty line (a 37 per cent increase from the baseline in March), and to 0.593 using the upper-bound poverty line (a 23 per cent increase from the baseline in March).

The COVID-19 policies played a particularly vital role for female-headed households, and households containing children or older people. Table 3 shows the poverty headcounts for three particularly vulnerable subgroups. The same overall pattern is observed as for the population as a whole: that is, poverty increases between March and April and then falls to levels lower than in March, though for these subgroups the fall to a level lower than in March occurs sooner (May) than for the population as a whole (June). For households containing one or more older people, poverty (as measured using the food poverty line) is almost obliterated. This will be driven by the R250 increase to the Old Age Grant from May onwards. The fall in poverty between the months of May and June will be due to the combined impact of the transition from an increase to the Child Support Grant payment (which occurred only in May) to a Caregiver SRD benefit (which started in June), and an improved role-out of the COVID-SRD benefit, as the number of beneficiaries increased from 4.4 million in May to almost 5.1 million people in June.

The COVID-19 polices greatly reduce the extent of poverty that would otherwise have existed: without them, poverty in female-headed households would have risen to 0.351 (a 44 per cent increase from the baseline in March), and poverty in households containing one or more older people would have risen to 0.156 (a 62 per cent increase from the baseline in March), and poverty in households containing one or more children would have risen to 0.339 (a 51 per cent increase from the baseline in March).

Table 4 shows the Gini coefficient for each month. From the first row it can be seen that inequality increased very slightly in April compared to March, but in May and June it fell to levels lower than in March. This was due to the reduced earnings in the top deciles, and increased incomes (mostly from the COVID-SRD and Caregiver-SRD) in the bottom deciles (as reflected in Figures 1–3). In the absence of any COVID-19 policies, inequality would have increased to 0.676.

As the COVID-19 benefit changes only commenced in May, it is possible to attribute the reduction of inequality in April from 0.676 (in the hypothetical situation of no COVID-19 policies) to 0.648 wholly to the TERS income received by furloughed workers. Similarly, as TERS is applied in a constant way in April, May, and June, the further reductions in inequality in May and June can be attributed to the COVID-19 benefits.

Figure 5 shows the overall change between March and June 2020 in mean monthly household disposable income, by decile and for South Africa as a whole. The figure decomposes the changes into three parts: changes due to loss of earnings caused by the pandemic (shown in dark grey); the cushioning effects of automatic stabilizers—that is, the tax–benefit system in place prior to the pandemic (shown in light blue); and the additional effects of the newly introduced COVID-19 policies (shown in dark blue). Overall (the final column), mean household disposable income fell between March and June by 11.0 per cent. If this change is decomposed, the change in earnings accounts for a 24.7 per cent drop in disposable income; the change in automatic stabilizers accounts for a 4.1 per cent rise in disposable income; and the introduction of new COVID-19 policies (including TERS) accounts for a 9.6 per cent rise in disposable income.

Figure 5

Download asset Open asset

These numbers can be used to calculate the so-called income stabilization coefficient.¹⁰ When comparing March and April it amounts to 40 per cent;¹¹ this measures the extent to which the automatic stabilizers and the TERS policy protected households from declines in market income. When comparing March and June, the income stabilization coefficient rises to 53 per cent, meaning that more than half of the drop in market income was avoided by the combined effects of automatic stabilization, TERS, and the COVID-19 benefit changes that were in place in June.

As can be seen, the mean disposable income increased in the lowest five deciles, remained largely unchanged for decile 6, and fell for deciles 7–10. The changes were driven by a combination of the introduction of the COVID-19 policies, a fall in earnings, and to a much lesser extent the role of the automatic stabilizers. Effects of the pre-COVID tax–benefit system (shown in light blue) at the top of the income distribution are mostly driven by the reduction in contributions to UIF and payments of personal income tax after earnings shocks.

The figure shows the important redistributional effect of the COVID-19 policies. However, although the percentage increases in mean disposable income are highest in the lowest deciles, the actual increases in rand amounts are very low (Figure 2).

Table 5 provides more detail about the profile of households in each of the deciles shown in Figure 5 in respect of earnings in March 2020 and April (and May and June) 2020. Only 13 per cent of households in the first (poorest) decile had earnings prior to the pandemic in March 2020, and this fell to under 9 per cent of households in April 2020. The mean earnings of those in the first decile are very low at both time points, which explains why the change in disposable income is so great in Figure 5 for the first decile between March and June: a COVID-SRD benefit of R350 per month is more or less equivalent to the mean monthly earnings of this decile.

It should also be kept in mind that SAMOD simulated many fewer recipients of TERS than received the benefit in practice (Department of Employment and Labour, 2020); as explained above in more detail, it is assumed that a subset of those who reported earnings in NIDS-CRAM Wave 1 were actually reporting income derived from TERS. As a consequence, the findings about the impact of the pandemic on earnings (shown in dark grey) is likely to be understated (that is, there will have been greater drops in earnings), and the counterbalancing impact of TERS (shown in dark blue) is also likely to be understated for those working in the formal sector (that is, TERS will have played a larger role than shown in protecting people’s incomes). Nevertheless, this should not affect the combined impact of the shock and all policies on mean distributional incomes (the white dots in Figure 5) as the income sources are not differentiated.

In summary, the COVID-19 policies not only served to mitigate the impact of the pandemic and lockdown to a great extent, but also represent a long-overdue change in policy approach by providing social assistance to low-income adults of working age.

This study has examined the impacts on poverty and inequality of a package of COVID emergency policies that were implemented by the South African government in late March 2020 in response to the arrival of the pandemic in South Africa. In the weeks preceding the announcement of this package, the government engaged intensively with South Africa’s research community over a range of possible interventions. Bassier et al. (2021) is one such study, originally written for the presidency in March 2020. It sought to use information from the NIDS Wave 5 survey of 2017 to investigate a set of emergency policies to support informal workers whose employment and earnings would be halted by a lockdown but who would not receive any relief through systems such as the UIF that rely on formal registration of employment. The evidence from NIDS Wave 5 was used to show that a very large proportion of such informal workers reside in households in the bottom deciles of the income distribution in which there are recipients of existing social grants, in particular the CSG. This made a case therefore that this grant could be re-purposed to provide substantial emergency relief to these workers and their households. They also showed that, if it were feasible to implement, a ‘Special COVID-19 Grant’ broadly targeting the unemployed and those in informal employment would be very effective in assisting these informal workers and mitigating COVID-19-associated poverty.

Quantitative work such as that of Bassier et al. (2021) used the 2017 situation to inform policy options by using ex-ante simulations of the impacts of various policy options. In this study, we have worked hard to ground our assessment on a base situation that prevailed in the country in 2020 on the eve of the pandemic and the lockdown. Then, we have incorporated reliable information on actual labour market outcomes between April 2020 and June 2020 in order to ensure that our simulations are assessing as closely as possible the impacts for incomes, poverty, and inequality of what actually happened in the labour market over this period.

Another useful point of comparison is a study by Zizzamia et al. (2020) based on NIDS-CRAM that matched job losers with observably similar job retainers in order to estimate poverty effects of COVID-induced job loss. Stressing that their results are highly approximate, they estimated that one million job losers (and three million people accounting for dependents) fell into poverty in April as a result of the COVID employment shock. This estimate incorporates grant receipt on the basis of survey responses and differs sharply to the poverty findings of this paper, but it must be borne in mind that the time point of April means SRD receipt would still have had a minimal effect on ameliorating poverty in their estimate.

See https://cramsurvey.org.

See Appendix A for a description of the model.

See Appendix B for details of this part of the analysis.

See Appendix C for details of this part of the analysis.

See Table C5 for estimates of how employment transitions varied along the dimensions of race and education level based on NIDS-CRAM.

Although people of working age (including caregivers) are not eligible for social assistance unless they are disabled, the social insurance scheme (Unemployment Insurance Fund) does exist, but this is time limited and depends on sufficient contributions having been made.

The monthly R350 payment is USD 24.53, EUR 20.60, and GBP 17.66 (xe.com 28 June 2021), and is similar to the value of the mean monthly earnings of households in the poorest household income decile (see Table 5 below).

Although most of the main tax and benefit policies that affect people’s incomes at the individual level are simulated in SAMOD V7.3-COVID, certain policies are not: value-added tax, grant in aid, the War Veterans Grant, and the usual UIF (i.e. non-TERS) payments. The only COVID-19 policy response that is not simulated is the introduction of tax payment deferrals.

As mentioned earlier, TERS was generally distributed to workers through their employers, and so income derived from TERS would have been reported as earnings by many respondents. In addition, the NIDS-CRAM questionnaire makes no distinction between different sources of earnings and does not explicitly tell respondents to exclude TERS.

See Dolls et al. (2012) for a description of the methodology.

Calculated as one minus the change in disposable income divided by the change in market income.

A similar analysis for Ecuador (another upper-middle-income country) shows, for example, a dramatic increase in poverty despite the introduction of a new social benefit (Jara et al., 2021). In the UK, while the income losses were smaller, the role of government protection is at roughly the same level as in South Africa (with a mean disposable income drop of 7 per cent against a corresponding reduction in earnings of 13 per cent in the UK) (Brewer and Tasseva, 2020).

On the basis of the information in the 2021 budget review, taxpayers had used tax deferrals with a total value of R40 billion until mid-February 2021.

See www.nids.uct.ac.za/nids-data/documentation/overview-documentation/wave-5.

The pandemic was declared a national disaster on 15 March 2020 and a national lockdown was announced on 23 March 2020.

In practice, it was not possible to ascertain which respondents from NIDS Wave 5 were ‘farmers’ and so this category was not coded in either the SAMOD input dataset or the QLFS external controls.

All children aged 0–14 were assigned code 99 on the ‘les’ classification, irrespective of whether they had been assigned a different economic status in the raw NIDS data or raw QLFS data.

In practice, it was not possible to identify armed forces personnel in the QLFS data, so this category was excluded. There were 15 respondents coded as ‘armed forces’ occupation in the NIDS data and these were recoded to ‘elementary occupation’.

All children aged 0–14 were assigned code 999 on the composite ‘les_loc’ variable, irrespective of whether they had been assigned a different economic status in the raw NIDS data or raw QLFS data.

This restriction to positive earners follows the precedent of Brewer and Tasseva (2020).

A more sophisticated model that does not rely on the assumption of the ‘independent of irrelevant alternatives’ (IIA) property could be used in future work.

Due to small numbers in the Asian/Indian group, a three-category race variable was used for regressions that collapsed the Coloured and the Asian/Indian groups into one category.

Due to a lack of available data, marriage status was not included as a regressor.

Note that this classification means that a few individuals who reported zero earnings in April but still worked positive days will be placed in the reduced earnings group rather than the furloughed group.

Note that the earnings variables in the input data have no missing values, which means there is no distinction between zero-earners and missing data (in contrast with NIDS-CRAM). The multinomial logit is restricted to February positive earners, which attenuates the effect this difference has, while the inability to distinguish between zero-earners and missing in April is largely consistent with the coding in NIDS-CRAM, which treated missing as sufficient to classify people as having zero earnings or earnings reductions.

This is in line with the method of Brewer and Tasseva (2020). Once again, a 15 per cent threshold is used for what counts as an earnings reduction.

SAMOD is a static tax–benefit microsimulation model for South Africa (Wright et al., 2016). It was developed and is maintained by members of Southern African Social Policy Research Insights. The model uses the EUROMOD software developed by Professor Holly Sutherland and colleagues at the University of Essex to simulate policies for the countries in the EU (Sutherland and Figari, 2013; University of Essex, 2019). The EUROMOD software was designed to enable analysis across countries using harmonized concepts and methodology, and is sufficiently flexible to be applicable to countries outside the EU, with South Africa being the first developing country to use the software (Wilkinson, 2009).

The analysis presented in this paper is based on SAMOD Version 7.3-COVID, which uses Version 3.1.8 of the EUROMOD software. SAMOD draws on underpinning input micro-datasets which are stored within the model as text files. The two underpinning datasets that are used in the analysis in this note are modified versions of the 2017 National Income Dynamics Study Wave 5 (SALDRU (Southern Africa Labour and Development Research Unit), 2017).¹⁴

South Africa’s tax and benefit rules for each of March, April, May, and June 2020 were added into SAMOD using harmonized EUROMOD commands. These rules are applied by the model to all individuals in the relevant underpinning dataset, taking into account characteristics such as age, gender, marital status, and detailed information on income sources. The taxes paid and benefits received by each individual are calculated on-model (simulated), so there is no need to rely on reported receipt. This enables the ‘next-day’ financial impact of the simulated policies on individuals to be calculated.

The model output comprises an individual-level text file for each run of the model, which can then be analysed in STATA. The datasets and tax–benefit systems used in the analysis for this background note are summarized in Table A1.

This appendix summarizes the steps taken to prepare a baseline dataset for SAMOD for this study. The objective was to generate an input dataset that would reflect the situation in South Africa immediately prior to the pandemic. This was achieved by reweighting the SAMOD dataset to a ‘pre-COVD’ timepoint of March 2020.¹⁵ Selected diagnostic outputs are also included in this appendix to illustrate the impact of the reweighting exercise on the distribution of survey weights.

The objective of the reweighting procedure was to recalibrate the survey weights in the SAMOD input dataset so that the weighted population totals corresponded to estimated demographic profiles and labour market profiles for March 2020. This was achieved through a two-stage process. First, the SAMOD input dataset was reweighted to match external demographic and labour market profiles for the timepoint at which NIDS Wave 5 was enumerated (mid-2017). Second, the reweighting was performed again, this time controlling to demographic and labour market profiles for March 2020. The advantage of adopting this two-stage approach is that it enables an assessment of the magnitude of weight changes at each stage. The demographic profiles were derived from Statistics South Africa’s population estimates by population group, age, and sex. The labour market profiles were derived from Statistics South Africa’s QLFS.

The reweighting process was undertaken using the technique of iterative proportional fitting (IPF; also referred to as ‘raking’). The Stata .ado file ‘ipfraking’ was utilized for this purpose.

The reweighting procedure consisted of five main stages:

• Stage 1: preparing the external population control totals relating to demographic profile.

• Stage 2: preparing the external population control totals relating to labour market profile.

• Stage 3: preparing the SAMOD input dataset prior to running IPF.

• Stage 4: running the IPF procedure, first for mid-2017 then for March 2020.

• Stage 5: performing quality assurance tests on the reweighted input datasets.

This step consisted of extracting the relevant population estimates from the file produced by Statistics South Africa and structuring the data into separate Excel worksheets for the relevant years. The population estimates produced by Statistics South Africa consist of estimated counts by age, sex, and population group. There are 34 separate quinary age/sex groups for each of the four population groups, resulting in 136 discrete demographic categories (i.e. 34 × 4 = 136). Population estimates are available for all 136 demographic categories at the mid-point of each year.

Following the NIDS weight calibration approach adopted by Branson and Wittenberg (2019), prior to implementing the IPF procedure the three most elderly Indian/Asian age groups (70–74, 75–79, and 80+) were collapsed into a combined ‘aged 70+’ category for males of the Indian/Asian population group and also separately for females ‘aged 70+’ of the Indian/Asian population group. This resulted in a final age/sex/population group classification consisting of 132 discrete categories derived from the Statistics South Africa population estimates.

To derive population estimates for the March 2020 timepoint, it was necessary to interpolate between the population estimates for mid-2019 and mid-2020. A simple linear interpolation approach was used to derive the estimates for March 2020.

This step entailed the derivation of labour market profiles from specific waves of the QLFS. Weighted population shares by labour market category were calculated for two separate classifications: (1) ‘current economic status’ (relating to the ‘les’ variable in the SAMOD input dataset); and (2) ‘occupation type’ (relating to the ‘loc’ variable in the SAMOD input dataset). A composite classification, named ‘les_loc’, was then derived by disaggregating the ‘self-employed’ and ‘employees’ according to their occupation type. These external statistics were calculated using QLFS waves 2017 Q2 and 2020 Q1.

The ‘les’ classification adhered to the following coding scheme:

• 1 = Farmer¹⁶

• 2 = Employer or self-employed

• 3 = Employee

• 4 = Pensioner

• 5 = Unemployed

• 6 = Student

• 7 = Inactive

• 8 = Sick or disabled

• 9 = Other

• 99 = Aged 0–14¹⁷

The ‘loc’ classification adhered to the following coding scheme:

• 0 = Armed forces occupations¹⁸

• 1 = Managers

• 2 = Professionals

• 3 = Technicians and associate professions

• 4 = Clerical support workers

• 5 = Service and sales workers

• 6 = Skilled agricultural, forestry, and fisheries

• 7 = Craft and related trades workers

• 8 = Plant and machine operators

• 9 = Elementary occupations

The ‘les_loc’ classification adhered to the following coding scheme:

• 201 = Employer or self-employed: Managers

• 202 = Employer or self-employed: Professionals

• 203 = Employer or self-employed: Technicians and associate professions

• 204 = Employer or self-employed: Clerical support workers

• 205 = Employer or self-employed: Service and sales workers

• 206 = Employer or self-employed: Skilled agricultural, forestry and fisheries

• 207 = Employer or self-employed: Craft and related trades workers

• 208 = Employer or self-employed: Plant and machine operators

• 209 = Employer or self-employed: Elementary occupations

• 301 = Employee: Managers

• 302 = Employee: Professionals

• 303 = Employee: Technicians and associate professions

• 304 = Employee: Clerical support workers

• 305 = Employee: Service and sales workers

• 306 = Employee: Skilled agricultural, forestry, and fisheries

• 307 = Employee: Craft and related trades workers

• 308 = Employee: Plant and machine operators

• 309 = Employee: Elementary occupations

• 400 = Pensioner

• 500 = Unemployed

• 600 = Student

• 700 = Inactive

• 800 = Sick or disabled

• 900 = Other

• 999 = Aged 0–14¹⁹

The QLFS labour market shares were then adjusted to take account of the appropriate population totals derived from the Statistics South Africa population estimate external control totals produced in Stage 1 of the reweighting procedure. The total population across the ‘les_loc’ classification derived from the QLFS was scaled to match the total population from the Statistics South Africa population estimates. Furthermore, the ‘les_loc’ category ‘999’ derived from the QLFS was fixed at the value of the population estimate for children aged 0–14 from Statistics South Africa. The QLFS labour market shares by ‘les_loc’ for people aged 15+ were then applied to the population estimate for people aged 15+ according to the Statistics South Africa estimates. Scaling the QLFS labour market profile to the respective population totals ensured that the sum of the labour market categories matched the sum of the age/sex/population group categories, thereby maintaining internal consistency between the two sets of external controls. Although this was not strictly necessary for the reweighting process, it provided methodological clarity as the adjustments to the control totals were explicitly defined rather than being implicitly generated during the reweighting procedure.

This step consisted of producing derived variables in the SAMOD input dataset to match the categories of the external control totals. A new variable, ‘pgagesex’, was derived in the SAMOD input dataset to categorize the survey respondents according to their population group, quinary age group, and sex.

As noted in the step above, in order to enable the labour market profile to be explicitly controlled during the reweighting process, the variables ‘les’ and ‘loc’ from the SAMOD input dataset (i.e. ‘current economic status’ and ‘occupation type’) were combined to form a composite variable called ‘les_loc’.

The ‘ipfraking’ Stata .ado file was used to operationalize the iterative proportional fitting procedure. The first round of iterative proportional fitting was configured to reweight the SAMOD input dataset so that the demographic and labour market profiles of the reweighted survey counts matched the demographic and labour market profiles in the external control totals for the 2017 year of survey enumeration.

The process commenced with the importation of the external control totals (both demographic profile and labour market profile) from the specified Excel files into Stata matrices. The ‘ipfraking’ command was then configured according to the specification of the SAMOD input dataset. Running a ‘summarize’ command on the original survey weight variable (dwt) in the input dataset revealed the range and distribution of the original survey weights. This information was then used to inform the setting of the trimming parameters in the ‘ipfraking’ command, which allowed the configuration of both absolute and relative trim limits within which the ‘ipfraking’ procedure was forced to operate. There is no hard rule in terms of how the trimming parameters should be configured, so these parameter values were specified according to the distribution of original survey weights and to ensure convergence of the ‘ipfraking’ procedure. The setting of trimming parameters inevitably involved an element of trial and error, with the parameter values gradually relaxed or tightened and the results from the ‘ipfraking’ procedure reviewed. The objective was to achieve convergence of the procedure to within the specified tolerances, while minimizing the magnitude of changes to individual weights and minimizing changes to the overall range and distribution of the weights.

The second part of the ‘ipfraking’ procedure entailed reweighting the survey to the demographic and labour market profiles for the March 2020 timepoint. This time, a ‘summarize’ command was run on the rebased 2017 survey weights, calculated as described above, and this weight distribution informed the configuration of the trimming parameters in the rebasing to March 2020. Again, having specified the initial trimming parameters based upon the results of the ‘summarize’ commands, a process of ‘trial and error’ was undertaken to refine the reweighting procedure in order to achieve convergence while minimizing distributional change to the weights. Figures B1 and B2 show distributional plots and trace plots illustrating the results of the ‘ipfraking’ procedure for SAMOD.

The final part of the ‘ipfraking’ procedure entailed the following: first, reviewing the outputs from the reweighting exercise to ensure that the weighted counts by demographic profile and labour market profile corresponded to the external control totals; and second, assessing the magnitude of change in the weights generated through the reweighting process.

Figures B1–B4 show key results from the quality assurance exercise. The validation work indicated that the rebased survey weights for the baseline were in line with each of the external control totals and the distributions were plausible when considered in the context of the distributions of the original survey weights in the input datasets.

It can be seen from figures B3 and B4 that the most substantive adjustments to the survey weights occurred in the rebasing of the original weights to the mid-2017 timepoint, controlling to the external demographic and labour market control totals. Panel A of Figure B3 shows the overall composite effect of rebasing from the original survey weight to the rebased mid-2017 weight, controlling to demographic and labour market totals. The overall effect can be disaggregated to assess the relative contributions of the demographic and labour market controls separately, and these are shown in Panels B and C of Figure B3, respectively. These panels show that both the demographic and labour market elements of the rebasing had notable effects.

With regards to the demographic controls for mid-2017, it is evident from Panel B of Figure B3 that the original survey weights were typically increased through the rebasing procedure. This was a necessary step to account for the calibration approach adopted in the original NIDS data, whereby individual-level non-response cases (within enumerated households) were allocated a survey weight, despite no information being captured in the survey concerning these ‘non-response’ individuals’ incomes or labour market status. As SAMOD must be based on an underpinning dataset with no missing values, any ‘non-response’ individuals had to be excluded from the dataset. The weight rebasing procedure therefore adjusted the weights to ensure that the weighted total of enumerated cases in the NIDS survey equated to the total population estimate from Statistics South Africa.

With regards to the labour market controls for mid-2017, there were differences in the labour market profiles enumerated by the QLFS compared to NIDS. As one of the purposes of this reweighting procedure was to use the QLFS to account for labour market change between mid-2017 and March 2020, it was first necessary to reweight the NIDS data to reflect the QLFS labour market profile in mid-2017. Panel C of Figure B3 shows the effects of reweighting the NIDS data so that the weighted labour market totals equated to the QLFS external statistics.

Figure B4 shows the effects of moving from the rebased mid-2017 timepoint to the ‘pre-crisis’ timepoint of March 2020. It is evident that the magnitude of difference between the rebased mid-2017 weights and the ‘pre-crisis’ March 2020 weights is much less than the magnitude of difference between the original survey weights and the rebased mid-2017 weights.

Figure B1

Download asset Open asset

Figure B2

Download asset Open asset

Figure B3

Download asset Open asset

Figure B4

Download asset Open asset

This appendix describes the simulation of the COVID-19 employment shock and the accompanying lockdowns in the input data (after updating to March 2020 pre-lockdown levels). The employment shock is estimated using NIDS-CRAM Wave 1 (referred to as the ‘shock data’).

NIDS-CRAM is a broadly representative individual-level survey implemented using CATI and focusing on adult individuals’ responses to the COVID-19 pandemic and national lockdown (Ingle et al., 2020). Conducted in May, respondents were asked retrospectively about their employment in April (after the imposition of the level 5 lockdown) and in February (pre-lockdown).

Given the structure of this data and its similarity to that used in the UK, this methodology draws principally from that of Brewer and Tasseva (2020) in terms of regression design and dependent variables. Principally, this means that we model probabilities of the employed transitioning into different states, rather than probabilities of being employed or not in the COVID era.

In line with the UKMOD (Brewer and Tasseva, 2020) and ECUAMOD (Jara et al., 2021) studies, this methodology focuses on estimating employment shocks (and enabling poverty and inequality estimates) at the peak of the crisis (i.e. in April when the level 5 lockdown was in place), leaving estimates of further developments in June and beyond to future work.

The study was originally commissioned under the UNU-WIDER project Southern Africa – Towards Inclusive Economic Development (SA-TIED).

An earlier version of this paper was published as WIDER Working Paper 68/2021 https://doi.org/10.35188/UNU-WIDER/2021/006-1

1. Version of Record published: August 31, 2021 (version 1)

© 2021, Barnes et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.