Social Distress and (Some) Relief: Estimating the Impact of Pandemic Job Loss on Poverty in South Africa

3.1. Income forecasting

We use an income measure close to disposable income, per capita, as our welfare indicator. The NIDS survey collects incomes and earnings net of taxes on wages and salaries, and the LCS collects a mix of both, so we adjust the LCS to similarly reflect net incomes and earnings.⁶ While this is not exactly disposable income – other sources of taxable income such as interest or dividends are recorded in gross terms – for the purposes of poverty analysis this is unlikely to make a notable difference, as these issues are overwhelmingly relevant only for the wealthy. Note that this aggregate is based on income, rather than consumption, and therefore not the same as the official welfare aggregate. As a result our results do not provide an update to the official poverty statistics but provide estimates of the change in the poverty level. We also do not conduct a full tax and benefit microsimulation model, which means we do not take into account the changes in indirect taxes due to the impact of the loss of income and Special COVID-19 SRD benefits on indirect taxes, and the resulting impact on the estimates of the change in poverty. In general, our focus in this paper is more on the benefit side. We forecast these “disposable” household incomes from the original survey year to Statistics South Africa (2022a)

As per a number of studies similarly estimating the impact of COVID-19 on poverty (for example Lustig and Pabon, 2020; Younger et al., 2020), to forecast income increases for the pre-pandemic period (from the original survey year up to December 2019) we use the growth in nominal per capita GNI from the national accounts.⁷ However, while the above-mentioned studies assume no growth in income in sectors over the subsequent mid-pandemic period, the evidence from South Africa is that for those individuals who retained employment, earnings kept pace with inflation (Bassier et al., 2021). From January 2020 to December 2021 we therefore assume wage and income growth equivalent to the growth in the consumer price index (CPI) of 8.85 per cent (authors’ calculations based on Statistics South Africa, 2022a; see Appendix B for more details).

One of the well-known challenges for forecasting survey incomes is that not all the growth in the national accounts is passed through to growth in household welfare in surveys (Ravallion, 2001), and studies typically adjust for this by applying a ‘pass-through’ factor. We use a pass-through factor of 1; this choice and other options are discussed in Appendix B.

3.2. Demographic updating

The second stage of the updating process is to account for population growth, which varies across demographic groups. We reweight the underlying dataset to reflect changes in the demographic profile of the South African population that occurred between the time of the original survey and the updating year.

Specifically, we reweight the underlying data so that weighted population proportions for each province and each age-race-gender interaction match those given in the relevant year’s Statistics South Africa mid-year population estimates (MYPE). 17 age categories, four racial groups and two genders implies 136 interacted age-race-gender cells (17 × 4 × 2 = 136). Including the nine provincial categories this gives 145 dimensions along which population proportions are adjusted plus the requirement that the weighted national population total in the survey match the total given in the MYPE.⁸ This type of demographic reweighting is very commonly used in the production of South African household survey data. We use the minimum cross-entropy estimation technique of Wittenberg (2010) to generate the weights.

3.3. Employment updating

The centrepiece of our dataset-updating methodology is to account for the dramatic employment changes associated with the pandemic. We do this by imposing individual employment status changes in the (pre-pandemic) base data (e.g. NIDS 2017), which induces changes in individual earnings and thus household incomes. We use the QLFS as the benchmark indicator of the contemporaneous state of the labour market, which provides regular quarterly surveys both before and during the pandemic.⁹

4.1. Employment rates by demographic characteristics

Reassuringly, the percentage change in employment rate in NIDS between 2017 and 2021 matches the employment change in the QLFS reasonably well (19 per cent and 16 per cent respectively). The slight divergence is likely due to lumpiness inherent in matching on survey weights and the issues associated with matching on many groups discussed in Section 3.3. The first row of Table C1 shows that in 2017, the overall employment rate (for those aged 25–55) in the underlying NIDS base data (63 per cent) was higher than that of the contemporaneous QLFS data (58 per cent). As expected (and indeed implicit) when matching on changes, this is reflected in the updated NIDS 2021 data, which show a marginally higher (51 per cent) employment rate than the QLFS (49 per cent).

Given that the NIDS 2021 data produce larger estimates of the aggregate employment loss, the implied employment loss for each demographic group is typically slightly higher in the NIDS 2021 data than in the QLFS. Table C1 shows changes in employment rates for each demographic group.

NIDS 2021 does not fully capture patterns of socioeconomic security and vulnerability in the labour market. For Africans, we match the employment loss reasonably well (21 per cent in NIDS 2021 vs 17 per cent in the QLFS), but our updating method overstates the extent of white unemployment loss (11 per cent in NIDS 2021 vs 1 per cent in the QLFS) because it does not fully capture the impact on job loss of systemic inequalities in the labour market.

This is likely a general issue: because the markers of labour market security are both too numerous to include in our matching and also sometimes simply unobservable, the updating procedure will tend to do poorly for extremely vulnerable or secure groups, assigning labour market changes more like the ‘group average’ and thus understating inequality. The updated data similarly misrepresent labour market stratification in the changes in employment rates by gender (to a limited extent), though in this case our updating method overestimates female employment loss relative to the QLFS and in fact suggests higher employment loss than for men, contrary to the QLFS.

4.2. Proportion employed by sector

The change in employment types in the NIDS data matches the QLFS changes very well. Table C2 shows the percentage of the employed who are informally employed, and separately their industry, as well as how these proportions change over time in the QLFS and in the updated NIDS data. Table C2 shows how closely the NIDS data matches the QLFS when it comes to employment sectors, and the changing composition of the workforce over the pandemic.

4.3. Employment rates by province

Table C3 shows employment rates and employment changes over time by province, a demographic characteristic not used in the matching procedure, and presents a mixed picture.

Overall, NIDS changes in employment rates by province match QLFS changes about as well as our matching on variables used in the updating process (Table C1). It is gratifying that the updating works well for the main economic centre of Gauteng as well as for the province with the lowest baseline employment rate, the Eastern Cape.

However, the NIDS updating substantially overestimates the employment decline in the Western Cape, Free State, and Northern Cape compared with the benchmark QLFS data. In the case of the Western Cape this may be explained by its racial composition: it is the province with the largest proportion of white people (for whom our updating procedure performs poorly) and the lowest proportion of African people (for whom the updating process does seem to work well). The Northern Cape is the most sparsely populated province and the difference may reflect noise and sample size issues.

Another cause may be that these are the three provinces with the lowest employment declines according to the QLFS. In general, the employment declines implied by our NIDS data are much more uniform across provinces than in the QLFS—the coefficient of variation of the NIDS employment changes is approximately one-third that of the QLFS changes. This again likely reflects the fact that our updating procedure cannot fully capture the stratification and heterogeneity of the labour market and thus understates inequality in employment vulnerability.

4.4. Household disposable income

There are substantial differences between the change in GNI growth observed in the national accounts from 2017 to 2021 and the change in incomes in the NIDS survey from 2017 to 2021 (Table C4). In our view this illustrates potentially serious problems associated with approaches which estimate pandemic poverty rates simply by updating incomes with national accounts data.

While nominal per capita GNI increased by 7 per cent between 2017 and 2019, per capita incomes in the NIDS data increased by only 2 per cent over the same pre-pandemic period, after we update 2017 NIDS income using both GNI growth and our employment updating procedure.

This divergence is even more marked over the 2020-2021 pandemic period, where per capita nominal GNI reportedly increased by 9 per cent whereas per capita income in our updated NIDS data decreased by 6 per cent.

The substantial 2020-21 discrepancy reflects a documented pandemic-period divergence in South African data sources. While the QLFS and the firm-surveyed Quarterly Employment Statistics (QES) have shown almost no employment recovery over the pandemic period, the NIDS-CRAM survey, national accounts, and industry-level sales data have shown quick recoveries or even some sector-specific growth over the period of the pandemic (Bassier et al., 2021; Simkins, 2021). In Simkins (2021) words: ‘When it comes to employment, the national accounts and NIDS-CRAM are in one corner, and the QLFS and QES are in the other’.

The reasons for these discrepancies are not clear, especially when it comes to divergences between NIDS-CRAM and the QLFS (Bassier et al., 2021). However, one possible factor when it comes to the national accounts and industry-level sales data is the 2020/21 commodity boom, which saw a substantial increase in prices and sales in sectors like mining, but which may not translate into equivalent employment or earnings increases for workers (Bassier et al., 2021). This points to the problem of reaching conclusions about poverty effects over the course of pandemic by extrapolating from national accounts growth rather than using real-time survey data, as has been used for other countries (Bhalla et al., 2022). The distribution of growth is of critical importance for poverty results, and applying a uniform rate of growth, especially over such a tumultuous period as the pandemic, may lead to spurious results.

5.1. Employment-loss-induced poverty

We now turn from a diagnostic exercise to what we can substantively learn from updating NIDS and LCS: poverty changes over the pandemic period, which are not available in the QLFS. In this section we look at employment-induced poverty in the absence of government pandemic-related income support or changes in household composition. We change incomes in response to employment loss or employment gain only. In the next section, we look at the capacity of the Special COVID-19 SRD to mitigate the poverty impact of the employment loss.

This section, continuing from the previous section, presents results from the NIDS data updated by matching on changes—which produce the largest poverty increases out of our various updating methods and changes.

We update the 2017 NIDS data to 2020 Q1 and 2021 Q4 and measure the change in poverty between these periods.¹⁵ The 2020 Q1 update mostly reflects the ‘pre-pandemic’ state of the labour market, while 2021 Q4 represents the end of the most severe lockdown policy thus far.

Our results, shown in Figure 1, suggest that over the pandemic period 2020 Q1-2021 Q4 there were substantial poverty increases at both the UBPL and FPL. An additional 3.1 million individuals fell below the UBPL (a 5.2 percentage point increase in poverty at the headcount ratio, or 13 per cent), while 2.8 million people fell below the FPL (a 4.7 percentage point increase in the poverty headcount ratio, or 30 per cent).¹⁶ The UBPL and FPL poverty gap increased by 3.8 and 2.6 percentage points respectively, or 21 and 45 per cent.

According to this set of results, poverty was also increasing in the pre-pandemic period, much more slowly, with a 1.5 percentage point increase in the UBPL headcount ratio over the 2017–2020 period (4 per cent) and a 1.2 percentage point increase in the FPL headcount ratio (8 per cent). The UBPL and FPL poverty gap increased by 1.3 percentage points (8 per cent) and 1.4 percentage points (33 per cent but off a low base) respectively. However, these pre-pandemic estimates need to be interpreted cautiously. As discussed in Section 3.3, we expect our updating method to work less well outside of periods with large employment shocks such as the pandemic, and indeed Section 6 shows that these particular results are highly sensitive to the updating method used.

5.2. Special COVID-19 SRD grant

The Special COVID-19 SRD grant is a ZAR350 per month (56 per cent of the FPL) digitally-administered conditional grant. Despite initially being conceived of as a temporary measure when introduced in May 2020, it was extended multiple times until April 2021, at which point it was suspended and then reinstated in August 2021, with all applicants required to reapply (Gronbach et al., 2022). As of the time of writing, it is still in operation.

Eligible applicants are meant to be older than 18, unemployed, and not receiving income from a variety of other government programmes (see Appendix E for additional details of the policy design, implementation, and our criteria for programme eligibility in the simulation). In reality, the employment test became a test of formal employment using government administrative data, with a means test of ZAR585 per month (the 2020 FPL) applied only to individuals who appealed their grant denial (Goldman et al., 2021).

We model the grant by means of a simulation, given that we do not observe Special COVID-19 SRD recipients in the QLFS. While some 21.5 million individuals are apparently eligible for the grant in the QLFS, 10.5 million were approved for payment in Statistics South Africa (2022a) only 9.7 million received a pay-out (SASSA, 2022). We select 10 million recipients from the pool of eligible individuals according to the proportional distribution observed in the National Income Dynamics Study—Coronavirus Rapid Mobile Survey (NIDS-CRAM) (Bassier et al., 2021; Bassier et al., 2021). Note that while we take into account actual disbursement in the simulation, the distribution of the grant may differ from actual distributional allocation of the grant, particularly given issues with the NIDS-CRAM data of reliability, granularity, and representativity, the fact that individuals had to reapply for the latest phase of the grant, and the timing of the NIDS-CRAM data collection (Wave 3, October 2020) compared with our Q4 2021 estimates.

Our simulations suggest the grant was effectively able to mitigate 35 per cent of the poverty headcount increase at the UBPL (61 per cent of the poverty gap) and 70 per cent of the poverty headcount increase at the FPL (73 per cent of the poverty gap) (Figure 2). While the size of the grant is small at R350 per month—only 26 per cent of the 2021 UBPL—it was received by close to 10 million people in December 2021. In comparison, our highest estimate shows an increase of 3.1 million additional poor individuals at the UBPL during the pandemic due to job loss. Of course, not all of these individuals will be lifted out of poverty due to the SRD – it depends on the size of the gap.

6.1. LCS vs NIDS changes

As noted above, our main results are presented using just the NIDS dataset, updated to match percentage changes in employment in the QLFS. In this section we present and compare results when using different updating methods or datasets.

Table 1 presents a summary of the pandemic-period poverty changes using the different methods. We show both the FPL and the UBPL results, but report only on the UBPL results in the text for simplicity. Appendix F provides additional figures and discussion of both the LCS vs NIDS results (Figure F1) and the NIDS levels vs NIDS changes results (Figure F2), as well as discussion of poverty changes leading up to the pandemic.

Updating for employment loss using the NIDS dataset and matching changes in the QLFS employment rates produces the highest estimated increases in poverty. Table 1 shows our baseline method estimates that from 2020 Q1 to 2021 Q4 there was a 5.2 percentage point increase in the poverty headcount (3.1 million individuals or an 11.4 per cent increase), and a 3.8 percentage point increase in the poverty gap (a 17.3 per cent increase). This decreases to 3.4 percentage points (2.0 million individuals and 8.4 per cent) in the poverty headcount and 1.5 percentage points (8.2 per cent) in the poverty gap once we take the simulated Special COVID-19 SRD into account.

When we vary the method, matching QLFS employment levels in 2017 and 2021, we find that the estimates are substantially lower, at 3.0 percentage points for the poverty headcount (1.8 million individuals, 6.4 per cent increase), and 2.5 for the poverty gap (10.5 per cent increase). This decreases to 1.1 percentage points (0.7 million and 2.5 per cent) in the poverty headcount and 0.2 percentage points (0.9 per cent) in the poverty gap once we take the simulation of the Special COVID-19 SRD into account.

Applying the changes method to the LCS dataset produces estimates in-between the NIDS changes and NIDS levels method. There is a 4.0 percentage point increase in the poverty headcount (2.4 million individuals, 8.2 per cent increase), and a 3.0 change in the poverty gap (11.9 per cent increase).¹⁷

In summary, whether we look at the results with or without the simulated Special COVID-19 SRD, there is a substantial level of uncertainty in these poverty estimates, with a grey area of about 1.3–1.4 million individuals who may or may not have been impoverished by the pandemic. The upper-bound poverty headcount increases by between 7 and 13 per cent and the poverty gap between 12 and 21 per cent taking only employment loss into account. Once we take into account the simulated Special COVID-19 SRD, we find that the upper-bound poverty headcount increases by between 2.5 and 8.4 per cent, the poverty gap between 0.9 and 8.2 per cent.

6.2. Comparison with existing estimates in the literature

We further compare our poverty results with two other estimates of mid-COVID-19 poverty in the literature (Table 2). Barnes et al. (2021) estimate job-loss-induced poverty for April 2020, and Van de Heever et al. (2021) present an estimate of employment-change-related poverty in Q4 of 2020 (we compare with their estimates without the Special COVID-19 SRD policy included). We consider these estimates to be broadly comparable: these papers use related but different methods, and in both cases their methodologies also differ from our own in important ways. Appendix G presents a detailed discussion of their methods and potential strengths and weaknesses. A general issue which must be kept in mind is that substantial COVID-19 related social assistance apart from the Special COVID-19 SRD was implemented in 2020, which we do not include in our 2020 Q2 and 2020 Q4 poverty estimates. This will have a limited effect on our comparisons, however, as we compare to the Barnes et al. (2021) April estimate, which is before these policies came into effect, and the Van de Heever et al. (2021) 2020 Q4 estimate, when these policies ended in October 2020.

The Barnes et al. (2021) estimates are for April 2020 (they update NIDS 2017 using NIDS-CRAM) and therefore are not directly comparable to any estimates we can produce using the quarterly data of the QLFS. Our closest comparison is updating NIDS 2017 (by matching on levels) to QLFS 2020 Q2, but Jain et al. (2020a) show that there was non-negligible employment recovery and poverty reduction over the second quarter of 2020, from April to June. We would therefore expect the Barnes et al. (2021) April estimates to show higher poverty than our 2020 Q2 equivalent, because our quarterly data incorporates some April–June recovery.

Indeed, this is exactly what we see: for 2020 Q2 our estimated headcount ratios are 22.5 and 46.7 per cent at the FPL and UBPL respectively, while the equivalent April 2020 estimates from Barnes et al. (2021) are 26.3 and 52.5 per cent. However, we are wary of attributing all of the substantial difference in our results to different time periods; as discussed in Appendix G, we believe that poverty may be overestimated in Barnes et al. (2021). It is notable that the Barnes et al. (2021) poverty estimates do reduce sharply after April 2020, as they simulate the various COVID-19 relief policies and automatic stabilizers; so much so that poverty in June 2020 is actually lower than in the pre-pandemic period.

Van de Heever et al. (2021) present poverty results for 2020 Q4 by updating NIDS to match the QLFS. While their method is substantively different from ours, in this case a direct comparison of the results to our own is possible (for this comparison we update NIDS to 2020 Q4 by matching on levels). Our headline poverty results are fairly similar. While we estimate UBPL and FPL headcount ratios of 21.6 and 46.0 per cent respectively in 2020 Q4, Van de Heever et al. (2021) estimate headcount ratios of 21.2 and 48.9 per cent.

Conducting this analysis provides us with a set of estimates over time. Our method suggests that in the absence of the Special COVID-19 SRD grant, there was no employment-induced poverty reduction over the period of the pandemic, with employment-related poverty essentially remaining constant. This is consistent with the QLFS data, which report no substantive employment recovery over the period of 2020 Q2 to 2021 Q4. This is a particularly stark finding, which must be considered alongside the active debate about whether the QLFS might understate employment recovery in the pandemic period (see Section 4.2).

1.

This was the National Income Dynamics Study—Coronavirus Rapid Mobile Survey (NIDS-CRAM). It used an income measure that was not comparable with that of previous surveys and was discontinued due to quality concerns. See Jain et al. (2020b) for discussion of the income measure and an attendant attempt to measure some poverty impacts in the first wave of the survey.

2.

Disposable income in a tax and benefit microsimulation model refers to an income aggregate that includes private income, net of direct taxes on income and salaries and direct transfers. Note that our aggregate differs from the official poverty aggregate in that it is based on income rather than consumption. This is discussed in more detail in Section 3.1 and in Appendix A.2.

3.

Note that we also run the method giving primacy to the QLFS data on the LCS 2014/15 dataset, and we find that the results are very similar to those of the LCS method. For simplicity we do not discuss the method or results here.

4.

See the appendix of the working paper version of this article for a full variable list used in the updating procedure.

5.

We do not simulate the impacts of other relief programmes such as the Presidential Employment Stimulus, the Unemployment Insurance Fund COVID19 Temporary Employment Relief (UIF COVID19 TERS), or various ‘top-ups’ to existing grants which were implemented during 2020. While this means that our poverty estimates will be somewhat overestimated during periods when these policies were in effect, the UIF COVID19 TERS and grant top-ups had been discontinued by the period of our main results, 2021 Q4.

6.

See Appendix A.2 for more detail on how income is captured in the NIDS and LCS.

7.

We use nominal values in order to produce a dataset that in 2021 will be comparable with 2021 poverty lines and administrative values for grants and income tax brackets, for example. The studies mentioned here use GDP while we use GNI; however, the differences are negligible. We also considered using private disposable income as an alternative, and again find the differences in growth rates to be small enough to ignore.

8.

We have 145 dimensions (136 + 9) because we separately impose that weighted age-race-gender population proportions match the national age-race-gender proportions per the MYPE (136 constraints) and that weighted provincial total population proportions match the provincial total population proportions per the MYPE (9 constraints). We do not directly impose that weighted age-race-gender proportions for each province match the equivalent provincial age-race-gender proportions in the MYPE; this would require demographic and provincial interactions, leading to 1224 constraints (136 × 9). We do require that recalibrated weights be constant within households, as in the underlying surveys, and do not allow missing item response in any of these demographic variables, as we need to use these same categories for the employment-updating process discussed below. If any individual is missing one of these characteristics, we impute it at random (with probabilities in proportion with the distribution of non-missing responses). This affects very few observations in practice.

9.

Questions have been raised about the accuracy of the pandemic-period QLFS (Paton, 2022; Simkins, 2021) and some analysts prefer the alternative NIDS-CRAM dataset,. In our view there are few technical reasons to prefer one dataset over the other; our use of the QLFS is necessitated by our requiring a consistent labour market series pre- and post-pandemic, which is available from no other source. The QLFS also has the advantage of being the official source of labour statistics in South Africa. We discuss potential issues with the QLFS and its divergences from NIDS-CRAM in Section 4.2, under ‘Household disposable income’.

10.

When updating NIDS we use 2017 as the base year, and when updating LCS we use 2015.

11.

In order to reduce dimensionality, we define race as a binary variable for the purposes of our matching algorithm, collapsing the standard African, coloured, and Indian/Asian classifications in South Africa into one category, with white as the other category. In our diagnostic section we nonetheless present results for each of the standard four racial typologies.

12.

These categories are selected from a review of the NIDS-CRAM literature, in particular Casale and Posel (2021); Jain et al. (2020b); and Ranchhod and Daniels (2021).

13.

Industry classification is not available in the LCS, so we use only formal/informal and a non-employed category when it comes to employment sectors, leading to three employment groups. The formal/informal definition in the LCS is also different to the NIDS definition. For each matching process we create an analogous formal/informal variable in the QLFS.

14.

Note that an alternative implementation to the cell-by-cell matching would be a regression approach. Specifically, one could create a cell-level dataset with the changes in the probability of employment (positive for employment growth, negative for employment loss), and regress this on cell characteristics such as gender, age, urban and education. A fully interacted version of this would be equivalent to the cell-by-cell approach that we implement. The advantage of such a regression approach would be that one could include fewer interactions, which would increase precision especially since there is an issue of a small number of observations in many cells. On the other hand, the disadvantage relative to the fully interacted approach would be that any non-interacted covariates rely on extrapolation which may be outside the common support of the actual employment probabilities of these groups, and also rely on a particular (variance-based) weighting of the covariate effect which may be less desirable than a proportions-based weighting in a prediction context. Ultimately we opted for the fully interacted approach, but any future work comparing the approaches would be useful.

15.

We update the NIDS 2017 data to 2020 Q1 very similarly to how we update it to 2021 Q4, outlined in Section 3, but instead use the 2020 Q1 QLFS to match employment changes rather than the 2021 Q4 QLFS. There are however differences when it comes to the income forecasting and demographic adjustments. For the nominal income adjustment, when adjusting for income growth after December 2019, we use the same CPI inflation factor (ie 8.85%) calculated from January 2020 to Statistics South Africa (2022a) For the demographic adjustment, we still reweight the NIDS data using the 2021 MYPE. These are convenient choices because we can use the same nominal poverty lines and population totals to compare 2020 Q1 and 2021 Q4 poverty estimates, but more than that they help us isolate the employment-induced poverty effects which this paper is concerned with, by abstracting from effects of nominal income or demographic changes. It is worth explicitly noting, however, that our results do not incorporate poverty effects due to these factors.

16.

The poverty lines are published by Statistics South Africa. The 2021 FPL was 624 rands (ZAR) per month while the UBPL was ZAR1,335 per month.

17.

Note that unlike the NIDS dataset, whether we apply the changes or levels method to the LCS dataset produces fairly similar results, with a 4.0 percentage point increase in the poverty headcount for both methods (2.4 million individuals, or an 8.2 and 7.7 per cent increase respectively), and a 3.0–3.1 change in the poverty gap (11.9 and 11.2 per cent increase respectively).

18.

That is, countries using an income aggregate, with a median income above US$172 in 2011 purchasing power parity (PPP), with a Gini index higher than 32.246, and which are not in the Europe and Central Asia region.

19.

Note that one possible extension to this income inflation method would be to estimate an index of distributional changes in household income growth per decile or quintile based on previous survey datasets or secondary surveys, and apply these to the GNI growth factor used to inflate income (see, for example, Van der Berg et al., 2007, who use a similar technique for adjusting surveys to be consistent with the national accounts). For this paper, in the spirit of keeping our method simple enough to be able to isolate the changes due to datasets versus method versus additional assumptions, we do not apply this technique here.

20.

Rather than using Wittenberg (2010)'s minimum cross-entropy estimation, they calibrate weights using iterative proportional fitting (raking). We do not expect these methods to lead to substantively different results.

21.

It is worth noting that their employment status variable is much more detailed than our own, with particular detail on different kinds of non-employment which we coarsely group together.

22.

These types of employment and earnings states are again substantially more detailed than our own, for example including a category for furloughed workers.

23.

From 2017 to 2021 (or 2015 to 2021) the Old Age Grant value increased by 18 per cent (or 34 per cent), the Child Support Grant value increased by 21 per cent (or 39 per cent), while the headline CPI increased by 16.5 per cent (or 32.5 per cent).