Overall analysis

The two variables of interest are reported incomes for February and April. However, some respondents reported bracket information rather than their actual income. For these respondents, their income is replaced with the mean of actual reported income that falls within their reported bracket. Figure 1 shows the kernel density of the log of income in February and April. The result shows that, while the upper part of the income distributions is similar, there are notable differences in the lower tail. Recall that our analysis is restricted to those who reported working in February. Figure 1 shows a significant increase in the proportion of people who reported zero or close to zero income in April relative to February.

Figure 1. Weighted kernel densities of log of reported income in February and April, 2020. Source: Weighted Wave 1 of NIDS-CRAM.

This can be attributed to job losses and workers who were furloughed as a result of the pandemic and lockdown. The result also suggests a possible change in the median as a result of the impact of the pandemic.

Figure 2 displays the relative density that compares the income distribution in February to the one in April (total difference). The lines on the bars represent the 95% confidence intervals. All the relative data estimations make use of survey weights and confidence intervals were computed using the bootstrap approach with 500 replications. Note that, when the confidence interval includes 1, we conclude that there is no significant difference between the relative density and 1. The result suggests that workers at or above the 7^th decile (of the February distribution) were barely affected in terms of income, since their relative densities are not significantly different from 1. Below the 7^th decile, the relative densities are significantly less than 1, which implies that there is more mass in these deciles in February relative to April 2020. It is also apparent that the 2^nd and 3^rd deciles were more affected than other deciles. As one would expect given Figure 1, the 1^st decile of the relative density is significantly greater than 1. This indicates that the comparison distribution (April) had more mass than the reference distribution (February) in this decile.

Figure 2. Weighted relative density using February income as the reference distribution. Source: Weighted Wave 1 of NIDS-CRAM.

This picture is consistent with the hollowing out of the middle of the distribution in April relative to February. Specifically, it means the income distribution has changed with movement of mass being largely from the 2^nd and 3^rd deciles to the 1^st decile.

Next, we examine the location and shape difference between February and April 2020. Figures 3 and 4 show the location and shape components of the decomposition (respectively).

Figure 3. Location component of the overall decomposition. Source: Weighted Wave 1 of NIDS-CRAM.

Figure 4. Shape component of the overall decomposition. Source: Weighted Wave 1 of NIDS-CRAM.

In terms of location difference, the pattern in Figure 3 suggests that (after adjusting for difference in shape) April’s income distribution was to the left of the February’s income distribution.

This suggests that the median in February was larger than the median in April, that is, in terms of location difference, the February distribution was better. This is consistent with the fact that many workers had lost some or all of their income by April. In terms of shape difference, the pattern in Figure 4 suggests that job loss between February and April had resulted in polarisation of the income distribution, that is, net of the median difference, the April income distribution is more polarised than the February distribution. This is confirmed by the U-shaped relative density.

Furthermore, Figure 4 suggests that polarisation was stronger at the lower tail. Compared with the result of Machema (Reference Machema2019), this suggests that gains in terms of lower polarisation have been eroded by the pandemic.

The information provided by the graphs can be summarised with the polarisation indices. A positive index is evidence of polarisation, while a negative index suggests the opposite (note this is only relevant for the shape component of the relative density decomposition). The polarisation indices in Table 1 (Panel 1) confirm the result presented in Figure 4 (shape component); all the polarisation indices MRP, URP, and the LRP are positive and statistically significant. The positive sign indicates that these coefficients represent polarisation from the median towards the tails of the distribution. Consistent with the description of how the pandemic has affected workers in South Africa – that is, the description of Jain, et al., (Reference Jain, Budlender and Zizzamia2020), Ranchhod and Daniels (Reference Ranchhod and Daniels2020), and Rogan and Skinner (Reference Rogan and Skinner2020) – LRP is larger in terms of magnitude, suggesting that the polarisation is stronger at the lower tail. On the other hand, workers at the top deciles (in February) are less affected, since these workers are more likely to have employment contracts or/and are able to perform their work tasks from home. As noted earlier, to check the robustness of the result, we restrict the analysis to individuals who reported monthly income only (this is about 58% of the unweighted sample). The result shows that there is significant polarisation from the median (MRP), and this is towards the bottom deciles (LRP). However, polarisation toward the upper deciles (URP) is not statistically significant. These results show that our substantive result is valid in the subsample (see table S1 in the supplementary file for details).

Table 1. Polarisation indices

MPR = median polarisation index; LRP = lower polarisation index; URP = upper polarisation index.

We note that the change in the income distribution can be interpreted as the impact of the pandemic on the income distribution. This is because the pandemic and associated lockdown can be thought of as an exogenous shock to the labour market. In other words, job losses between February and April were not influenced by actors in the labour market. The pandemic can therefore be argued to be the cause of the change in the income distribution.

Subgroup analysis

In this section, we explore the effect of the pandemic on the income distribution in different subgroups across time. This is motivated by the notion that workers with certain characteristics are disproportionately affected. In this section, we focus on the shape component. Figures 5 and 6 disaggregate the analysis in Figure 4 (i.e., the shape component) by gender to see if the polarising effect differs by gender. It is evident that, while the polarising effect is applicable to both gender groups, this effect is stronger for female workers.

Figure 5. Weighted relative density using February income as the reference distribution (male). Source: Weighted Wave 1 of NIDS-CRAM.

Figure 6. Weighted relative density using February income as the reference distribution (female). Source: Weighted Wave 1 of NIDS-CRAM.

This is because the relative density in Figure 6 (for women) appears to have been impacted more, relative to the one in Figure 5 (for men). Specifically, for the male distribution, individuals above the 5^th decile of the reference distribution are not affected (the relative densities are not statistically different from 1). However, in the female distribution, one must be above the 7^th decile of the reference distribution to experience the same protection. Table 1 (Panel 2) shows the associated MRP, LRP, and URP for Figures 5 and 6. Specifically, the polarisation indices are higher in the female distribution relative to the male distribution: 0.53 versus 0.28 for MRP; 0.67 versus 0.53 for lower relative polarisation (LRP); and 0.39 versus 0.03 for upper relative polarisation (URP). This result confirms that female workers are disproportionately affected.

Panels 3 and 4 of Table 1 show the polarisation results by race (Blacks vs. non-Blacks) and age (workers who are 30 years of age or younger vs. workers who are older than 30 years of age). The results are consistent with the gender result, that is, polarisation in the distribution of income in vulnerable groups (i.e., female and younger workers) is stronger than their counterparts. Specifically, the values of the polarisation indices are higher for the vulnerable groups. It is also evident that the polarisation is in the form of the downgrading of the income distribution.

Explaining gender differences (in April) using decomposition analysis

The analysis so far suggests that net of the pandemic-induced income losses, characteristics like being young, female, or African/Black was associated with higher income penalties in April (note that these groups are also more likely to work in the informal sector). However, the preceding analysis is not multivariate in nature, we therefore consider decomposition analysis that allows us to assess a more comprehensive list of characteristics that explains differences in the distributions. The advantage of the Oaxaca-Blinder decomposition discussed earlier is that one can identify covariates that explain these differences at the quantiles. However, unlike the previous analysis where income distribution was compared over time, this analysis compares the male and female distribution in April. This is because gender is one of the major factors that defines those who were disproportionately affected. Furthermore, as noted earlier, we only have the distribution of characteristics at one point in time.

The controls used in the decomposition analysis include location (traditional, urban, or farms), employment status in February and April, tertiary education, self-reported health, age, race, type of employment (regular, casual, self-employed, and running a business), a dummy for employment contract, and type of job (managers, professionals, and others). Table A1 in the Supplementary file presents the summary statistics. Table 2 below presents the Oaxaca-Blinder decomposition of the location difference. Following Clementi, et al., (Reference Clementi, Molini and Schettino2018), it is noted that the densities being compared under the location effect only differ in the measure of location (median in this case), that is, they have the same shape, so performing the decomposition at the median will suffice.

Table 2. Oaxaca-Blinder decomposition of location (median) difference

Standard errors are in parentheses

* p < 0.1,

** p < 0.05,

*** p < 0.01.

^a Employee has an employment contract.

^b This is a dummy that is 1 if the respondent is a manager or professional and 0 otherwise.

^c Omitted category includes casual work, self-employment, and running a business.

^d Omitted category is traditional areas.

Table 2 shows that male workers had a higher median wage than female workers, and the difference between the median was statistically significant. Furthermore, the endowment, coefficient, and interaction effects are statistically significant in explaining the median gap (note that both coefficient and endowment effect increase the gapFootnote ⁵). Three variables explain significant variation in the media gap: employment status in April, having a regular job, and age. Consistent with results in the last section, this shows that, apart from employment status, type of employment and age explain the difference in median income between male and female workers in April.

Tables 3 and 4 show the decomposition for the shape effect. For this analysis, the decomposition was estimated for $\tau = 10,20,30,70,80,90.$ Footnote ⁶ In terms of the coefficient and endowment effect, both effects are not significant in explaining the gap at the lowest quantile (i.e., $\tau = 10$ ); however, at the 20^th and 30^th quantiles, only the coefficient effect is significant. At the higher quantiles (i.e., $\tau = 70,{\rm{\;}}80$ ), both the coefficient and endowment effects are significant. While in the highest quantile (i.e., $\tau = 90$ ), only the coefficient effect is significant. The result shows that significant endowment effect can only be found at the upper quantiles (i.e., $\tau = 70,{\rm{\;}}80$ ). Apart from this, the size of the coefficient effect suggests that the polarisation effect becomes stronger at the higher quantiles. In other words, the coefficient (or unexplained) effect was more important and was stronger at higher quantiles.

Table 3. Oaxaca-Blinder decomposition of shape effect

Standard errors are in parentheses

* p < 0.1,

** p < 0.05,

*** p < 0.01.

^a Employee has an employment contract.

^b This is a dummy that is 1 if the respondent is a manager or professional and 0 otherwise.

^c Omitted category includes casual work, self-employment, and running a business.

^d Omitted category is traditional areas.

Table 4. Oaxaca-Blinder decomposition of shape effect (contd)

Standard errors are in parentheses

* p < 0.1,

** p < 0.05,

*** p < 0.01.

^a Employee has an employment contract.

^b This is a dummy that is 1 if the respondent is a manager or professional and 0 otherwise.

^c Omitted category includes casual work, self-employment, and running a business.

^d Omitted category is traditional areas.

The detailed decomposition shows that, like the median effect, employment status was important in explaining differences at the quantiles. Furthermore, job characteristics in terms of type of employment (regular as against casual, self-employment, and running a business), having an employment contract, and being a professional explain the shape effect. It is also noted that, while being a professional was not significant at the lower quantiles (i.e., $\tau = 10$ , 20, 30), it was important at the higher quantiles.

The implication of this result is that net of change in employment status, the type of employment in terms of having an employment contract, being a professional, and having a regular job are important in explaining the income differences. The result also shows that differences in returns are not only important in explaining the inequality, but also the significance of this factor increased at the higher quantiles. This is important because it is consistent with existing evidence that suggests the returns are more important in explaining income differentials in South Africa (Mosomi, Reference Mosomi2019). However, it further suggests that the unexplained effect that favoured male workers before the pandemic remained effective in mitigating the effect of the pandemic on male workers. The raw differences show that the net effect of these covariates was that they disproportionately protected male workers under the pandemic. This implies that the pandemic and the associated lockdown exacerbated existing income gaps like the gender and informal income gap.

These results are important for two reasons. First, the size of the informal economy in African countries is relatively large. The ILO estimate from 2018 showed that about 60% of the world’s employed population are in the informal economy, while in sub-Saharan Africa this figure is 85%.Footnote ⁷ Second, because of relatively high unemployment rates in developing countries, women and young workers (who happen to be the demographic groups in the majority) are more exposed to informal employment (ILO, 2018). This indicates that a larger proportion of the African population has suffered under health restrictions that were designed to get the pandemic under control. Therefore, the current (largely informal) labour-market structure in developing countries is not designed to cope with a pandemic like COVID-19. Workers who are referred to as vulnerable in normal times are even more vulnerable under the kind of lockdown restrictions that have been necessary to control the outbreak of COVID-19. These categories of workers and the households that depend on them face a decision between hunger and compromising their health under these restrictions. It is therefore not surprising that workers in this category may opt not to obey lockdown rules, especially in countries where they are not covered by social security.

The World Health Organization has warned that COVID-19 may not be the last pandemic (Gill, Reference Gill2020).Footnote ⁸ This raises the question of how low- and middle-income countries plan to deal with future pandemics if that were to occur. A fragile labour market structure coupled with an absence of social security systems in developing countries points to a lack of resilience in these economies to withstand the strain of a pandemic. South Africa is one of the few countries in Africa that has an existing and effective social security system, but it still struggled and is still struggling under the effects of the pandemic. The experience in South Africa suggests that only existing channels (i.e., old age pension, child support grant, etc.) were instrumental in getting needed help to the vulnerable under lockdown conditions. This is because the newly established Social Relief of Distress (SRD) grant, which was established to mitigate the effect of the lockdown on individuals who were not covered under existing grant systems, faced teething problems. The point is that most African countries were not prepared to cope under lockdown restrictions, both in terms of budget to make financial relief available and the logistics required to make sure such help gets to where it is needed the most.

In countries where social security systems are simply not in existence, one can imagine that the situation will be worse. Therefore, given that the threat of another pandemic remains a valid concern, the absence of a social security system coupled with a large informal economy predicts possible humanitarian crises going forward.