The informal sector wage gap: accounting for misreporting amongst the informally self-employed.

Transcription

1 The informal sector wage gap: accounting for misreporting amongst the informally self-employed. Murray Leibbrandt, Neil Lloyd and Patrizio Piraino Southern Africa Labour and Development Research Unit 1 University of Cape Town, South Africa INCOMPLETE DRAFT -- NOT FOR CITATION ABSTRACT: The paper contributes to a growing literature on the sectoral wage gaps in emerging countries. We estimate both linear and quantile sectoral wage gaps in South Africa controlling for time-invariant individual heterogeneity using data from the Quarterly Labour Survey (QLFS). Our main contribution is to match detailed earnings information from the Survey of Employers and Selfemployed (SESE) to the QLFS data. This enables us to investigate the impact of over-reporting amongst the informally self-employed on estimates of the wage gaps. We argue that earnings from the SESE data are less prone to misreporting and use them in the matched sample. We find that the positive premium to informal self-employment at the upper end of the wage distribution, which is present in the QLFS (and in the literature), falls away. I. INTRODUCTION The South African labour market is characterized by high chronic unemployment, high wage inequality, and relatively low informal sector participation, even in comparison to other middleincome economies, such as Brazil and Mexico (Rodrik 2008). Studies on entrepreneurship and business development find that South Africa has the lowest level of enterprise dynamism in the region (GEM, 2012). Labour market inequality, through both unemployment and wage inequality, is the foremost driver of income inequality in the post-apartheid South Africa, one of the most unequal countries in the world (Leibbrandt et al. 2012). In the post-2008 financial crisis period, the official unemployment rate (strict definition) has remained steady around 25% (Statistics South Africa, 2014) while real economic growth has 1 This research has received funding from the NOPOOR project ( under the FP7 of the European Commission.

2 stagnated. 2 Banerjee et al. (2008) have argued that this high unemployment rate may even be an equilibrium rate suggesting that structural factors explain its persistence. Why is it then, that when so many are unable to find employment, informal sector activity remains so low? This apparent puzzle of the South African economy is addressed in a number of South African labour market papers (see Kingdon & Knight, 2003; and Rodrik, 2008; for seminal contributions to this debate). The low level of informal sector activity is not only a research puzzle for economists, but remains a major policy question. The current South African government, through its National Development Plan, is trying to look beyond `big business' and to the employment capacity of Small and Medium Enterprises (NPC, 2012). Notably, this includes the informal sector (Fourie, 2013). Depending on the informal sector, and in particular its business owners, as sources of future employment creation demands a better understanding of the linkages between the informal and formal economy, as well as of the returns to self-employment. Much of the literature has focused on the estimation of wage-gaps (earnings differentials) between the formal and informal sector, using this as evidence of the segmented labour market hypothesis. A large wage gap, which drives wages in the informal sector below workers reservation wages may then explain the low level of informal sector activity alongside high unemployment (Kingdon and Knight 2003). However, a recent study by Bargain and Kwenda (2011) finds that after controlling for individual fixed effects there is a positive premium to informal self-employment at the top end of the wage distribution (comparing to the formal wage employment). This would suggest that some business owners in the informal economy fair better than their formal counterparts. Our methodology draws on the recent South African literature (Bargain and Kwenda, 2011, 2014 and El Badaoui et al., 2008). We estimate linear and non-linear wage gaps using the third quarter of the 2013 QLFS data. Moreover, using the rotating panel of the QLFS we construct a two-wave panel (2013:Q2-Q3) and estimate a series of fixed-effects models. Our primary deviation from the existing literature comes in the attempt to integrate more reliable earnings data on informal sector enterprises. To achieve this, we use Statistics South Africa s Survey of Employers and Self-Employed (SESE). This is a survey administered in conjunction with the QLFS every four years (most recently in Q3:2013). Our results suggest a more dismal picture on informal self-employment to that of previous studies. We find a negative premium to informal self-employment across the entire wage distribution, including at the top. The rest of the paper proceeds as follows. First, we offer a brief overview of the international literature on labour market segmentation, before focusing in on studies specific to South Africa. We then give a detailed account of the data and in particular focus on the construction and distribution of the earnings series we estimate from the SESE. Our econometric analysis begins with a set of OLS and quantile regressions using the cross-sectional data. We then use a two-stage fixed effects estimator to control for time-invariant individual heterogeneity. Finally, we replicate these estimations substituting the informal self-employed earnings data from the QLFS with that of the SESE. Various approaches taken to ensure the robustness of our findings are discussed throughout the analysis. 2 The South African Reserve Bank's projection for 2015 is 2.2% (SARB, 2015).

3 II. LITERATURE REVIEW i. International Wage-Gap Literature The presence of both a formal and informal sector in the economy, and by extension labour market, is often understood to be the outcome of labour market segmentation. The traditional case for segmentation, as presented by Harris and Todaro (1970) (later extended by Fields 1975 and Dickens and Lang 1984), argues that wages above the market clearing rate in the formal sector, often as a result of regulation, create unemployment in the formal economy. This has a spill-over effect on the supply of workers in the informal sector, thereby driving down the informal market clearing wage rate. Moreover, regulation and unionisation may drive wages up in the formal sector, while a lack of representation reduces the bargaining power of employees in the informal sector, adding to the downward pressure on the informal wage rate (Carneiro & Henley 1998). This explanation creates a dualistic view of the labour market, dividing it between the regulated (formal) and unregulated (informal) sectors, where regulation constitutes a broad set of market interventions. One of the cases made against the dualistic view of labour market segmentation is given by Maloney (1999). Maloney suggests that informal sector employees (self-employed) may enjoy certain benefits (such as flexibility) and avoid implicit costs associated with regulation (i.e. tax). Workers may therefore prefer to work in the informal sector (or be self-employed). This differs from Lucas's model of selection into self-employment based on managerial and entrepreneurial competence or Tybout's notion of efficient entrepreneurs avoiding an overly regulated formal sector (Lucas 1978; Tybout 2000). Indeed, Fajnzylber et al. (2006) find that in Mexico salaried individuals are more likely to become self-employed relative to the unemployed, contrasting the classical view of the informal sector as a safety net for the unemployed (see also Hart 1973; de Soto 1989; Maloney 2004). Blanchflower and Oswald (1992) provide evidence to show that in a developed economy setting there are real non-pecuniary gains to self-employment, in terms of job and life satisfaction (see also Hamilton, 2000). Wage gaps could therefore be an efficient compensating differential for any pecuniary or non-pecuniary gains/losses associated with a particular employment agreement. Firm size also plays a role in this discussion, being related to both wages (i.e. larger firms pay more) and formality. 3 El Badaoui et al. (2010) show that in a model where large firms are more likely to be audited for tax purposes the efficient outcome is for large firms to enter the formal sector and small firms to stay in the informal sector (see also Bargain et al., 2013). In this case, any wage premium is simply a size premium, which they confirm empirically using Ecuadorian data (see also Pratap & Quintin 2006). The size premium is consistent with the classical efficiency wage model if large firms employ more productive individuals; suggesting that correctly controlling for unobserved skill should negate the size effect (Oi and Idson, 1999). If controlling for unobserved skills does not eliminate the size effect, it can be argued that structural differences must explain the size effect (Bulow and Summers 1985; Ringuedé 1998; Bertola and Garibaldi 2001). 3 In part, this is due to the way informality is defined by statistical agencies (Bargain et al. 2013; Söderbom & Teal 2004).

4 Söderbom et al. (2005) use panel data from Ghanaian and Kenyan manufacturing firms to show that firm size effects remain significant (and larger in magnitude than developed world estimates) after controlling for time invariant characteristics (individual heterogeneity). Therefore, structural differences are likely to be a part of the story, especially in a developing economy setting. Falco et al. (2011) expand on this work using the same data from Ghana and Tanzania. They show that unobserved ability is far more important than observed human capital in determining earnings. At the same time, unobserved market ability does not appear to explain the full extent of the informal sector divide, confirming the results of Söderbom et al. (2005). Thus, using firm size or selfemployment as a proxy for informality is likely to be misleading. ii. South African context and literature The labour market segmentation literature has developed in South Africa out of a need to understand the country's chronic high unemployment. Why do the unemployed not enter the informal sector and why is the informal sector so small in comparison with other developing countries? This is the question posed by Kingdon and Knight (2003), which has initiated a debate around the nature and extent of segmentation in the South African labour market. Cichello (2005) finds that sector change is the most important variable for explaining earnings growth in the post-apartheid South Africa between 1993 and An individual who switched from a formal sector job to the informal sector had a lower predicted earnings growth than someone who stayed in the formal sector. Kingdon and Knight (2003), using 1993 PSLSD survey data find evidence of a large formal sector premium, which they argue can help explain the high level of unemployment in South Africa. Furthermore, given that the unemployed, relative to the informally employed, are less likely to transition into formal employment, they argue that there must be barriers to entry in the informal sector. Alternatively, the reservation wages of the unemployed must be too high for them to enter the informal sector. However, Kingdon and Knight define the `informal sector' to be all those not in regular employment - the casually employed, domestic service, or agricultural/nonagricultural self-employed - which is a broad, heterogeneous group and differs from the current ILO definition of the informal sector. El Badaoui et al. (2008) re-investigate Kingdon and Knight's (2003) initial estimate of the informal sector wage gap. Using data from the Labour Force Survey ( ) they estimate both OLS and difference-in-difference earnings equations using the 20% rotating panel. Their sample is restricted to wage employed males, excluding all self-employed individuals. In addition, they define the informal sector using more detailed information from the LFS, including the respondents own view as to which sector they work in. Their results show that after controlling for unobservable time invariant characteristics, as well as taxes, there is no informal sector wage gap. However, they only estimate linear earnings equations, which does not allow the informal sector premium to vary along the wage distribution. Since only earners at the upper end of the wage distribution qualify for income tax in South Africa, the mean and median tax effects may differ substantially. Bargain and Kwenda (2011; 2014) present a detailed investigation into wage gaps in South Africa (alongside Brazil and Mexico). Their analysis makes use of both linear and non-linear fixed-effects estimates to show that the sectoral wage gaps differ across the wage distribution. Using rotating

5 panel data from the Labour Force Survey they show that both observed and unobserved skills explain a large portion of the unconditional wage gap in South Africa, and that firm size cannot explain everything. Using quantile regression techniques, they find an informal wage gap of between 25% and 35% at the lower end of the wage distribution (Bargain & Kwenda 2014). In their earlier study (Bargain and Kwenda, 2011), which included the informal self-employed, the authors find a significant informal self-employment premium at the top of the distribution, of approximately 12% (similar to Marcouiller et al. 1997). To our knowledge, this is the only paper in the South African literature that includes both the informal wage- and self-employed separately in the analysis. III. DEFINITIONS, DATA, SUMMARY STATISTICS, AND WAGE DISTRIBUTIONS i. Survey description Our analysis is based on data from the 2013 Quarterly Labour Force Survey (QLFS) of Statistics South Africa. The QLFS replaced the LFS (a biannual survey) in 2008 as Statistics South Africa s primarily labour market survey. As with its predecessor, it is sampled from a household survey framework and contains a 25% rotating panel (LFS had a 20% rotating panel). In addition to using the 2013:Q3 crosssection (n=14,686 employed individuals) we construct a 2-wave panel using the Q2 and Q3 samples, which should at best be a 75% sample of the original 2013:Q2 cross-section. As we do not explicitly model labour market entry and exit, our sample only includes individuals who were employed in both waves (n=8,037 individuals). Despite the availability of QLFS data from 2008 to 2015, we deliberately limit ourselves to this short time frame in order to compare results from the most recent Survey of Employers and Selfemployed (SESE) which took place in 2013:Q3. SESE re-interviews all individuals in the QLFS who selfidentify as the owner of a business (employer or own account worker) which is not registered for Value Added Tax (VAT). The primary purpose of this secondary survey is to estimate the extent of the informal economy in South Africa (StatsSA, 2014). As the mandate of SESE is to measure the extent of the unregistered commerce in South Africa, it only captures data on employers and own account workers and is, therefore, not a comprehensive labour force survey of the informal economy. In particular, it excludes all employees of non-vatregistered businesses. SESE is therefore best understood as an informal enterprise survey. This is evident in the questions it asks, which focus on the characteristics of the enterprise, including detailed business accounts, and not on the characteristics of the business owner. Nevertheless, the data can be matched one-to-one with the business owner s survey responses in the QLFS, thereby providing more detailed individual and household information. ii. Defining labour market sectors What constitutes the informal sector has changed over time and differs within both the academic and policy spaces. Even within the South African literature there is a lack of consistency, with a different definition of the informal sector used on a case by case basis. Certain studies even focus on informal employment or the informal economy in place of the informal sector (Devey et al. 2003;

6 Devey et al. 2006). These definitions refer to a broader group than the informal sector, which is an enterprise (or employer) based division of the labour market: an employee works in the informal sector if the enterprise that employs them fits the profile of an informal enterprise. In general, the informal sector refers to a division of the economy - and by extension of the labour market - which is not regulated and does not contribute to the fiscal system in the form of sales tax (VAT) and income tax (labour or corporate income). It is common to conflate the notion of self-employment with the informal sector (e.g. Kingdon and Knight 2003). Indeed, data restrictions often make it impossible to create a more nuanced division of the labour market. This is not the case with the current QLFS. Statistics South Africa, in accordance with the International Labour Organisation, defines the informal sector as: (1) [e]mployees working in establishments that employ less than five employees and do not deduct income tax from their salary or wage ; or (2) [e]mployers, own account workers and persons helping unpaid in their household business who are not registered for either income tax or value-added tax (QLFS, 2008). 4 Using the above definition, we divide the labour market into five sectors: the formal wage employed, public sector (including civil servants), formal self-employed, informal wage employed, and informal self-employed. Excluded from these groups (and our analysis in general) are all unpaid workers, agricultural workers and private household employees (predominantly domestic workers). We include both male and female workers in our sample. A concern in including females is that they may have other factors determining their labour supply. However, as a large share of the informal sector is made of females, we opted to include them in our analysis. Moreover, these gender differences might be an important component of the labour market segmentation. Including individuals who work for private households, which in South Africa predominantly concerns female workers, would create a stronger correlation between employment sector and gender. Our segmentation of the labour market allows for more heterogeneity than in other studies where the informal sector variable is included as a binary variable. It also makes the important distinction between informality and self-employment, recognising that in South Africa there are many selfemployed professionals and small business owners who do not fit the presumed profile of an informal sector enterprise. In addition, we use a more contemporary definition of the informal sector based on tax registration and not self-reported answers. iii. Summary Statistics Table 1 gives, amongst other things, the distribution of the employment sector variable for the third quarter of Approximately 21% of the employed labour force is employed in the public sector, about 57% are wage employed and 3.5% self-employed in the formal sector. About 7% are wage employed and 11.5% are self-employed in the informal sector. 4 Given the ILO definition of the formal sector, which includes self-employed individuals who are registered for income tax (even if their business is not registered for VAT), the SESE sample does not fall directly within our definition of informal self-employment: approximately 7.35% of the matched SESE-QLFS enterprises are considered to be operating within the formal sector by this definition.

7 In the forthcoming sections, we estimate a series of Mincer-type earnings equations controlling for individual and job characteristics. The individual and household level controls include age, age squared, race, years of education, education squared, indicator for married (including cohabiting couples), household size, geographical location (urban-rural), and province. The job related variables include hours worked (last month), tenure, tenure squared, firm size, and occupation. Unfortunately, we do not have information on experience (only tenure) and are limited to a categorical firm size variable. On the other hand, we have firm size information for both the wage and self-employed and therefore do not define a separate firm size variable for the wage and selfemployed (as in Falco et al. 2011). Table 2 reports the mean for each of the covariates used in our model with the exception of province and occupation. Table 1: Table of Means: Individual and job characteristics Formal Sector Informal Sector Wage Public Self Wage Self INDIVIDUAL CHARACTERISTICS Male Female Race African Coloured Asian/Indian White Age (years) Education (years) Household Size Married Geographical location Urban Rural JOB CHARACTERISTICS Actual Hours (week) Tenure Firm size 0 employees employee employees employees employees employees employees Observations 8426 (56.7%) 3170 (21.3%) 516 (3.5%) 1033 (6.9%) 1723 (11.6%) Regarding gender, we find that female labour market participation is much higher among the informal self-employed in comparison with the formal self-employed. The relatively larger

8 participation of women in the informal sector - especially in a self-employed capacity - is a welldocumented characteristic of developing economies (see Maloney 2004). However, an overview of the past four SESE surveys shows a dramatic change in the gender profile of informal enterprises from 60% to 45% female ownership between 2001 and 2013 (StatsSA, 2013). To our knowledge, this trend is not well documented and explanations may range from high unemployment in the formal sector to more significant structural changes within the informal sector. Recall that our exclusion of private household employees from our measure of the informal sector will also decrease female labour market participation, as private household employment consists primarily of female domestic workers. Informal self-employment also displays a significantly higher proportion of Africans (91%), while white South Africans make up a disproportionate share of formal self- and wage employment. Given South Africa's history, this is an important distinction between the formal and informal selfemployed, which in turn will have implications for other socio-economic factors such as education and occupation. Indeed, average education among the formal self-employed is 12 years (this equivalent to a matric certificate) while amongst the informal self-employed is 9 years. Here we find an important justification for our deviation from including all self-employed within the category of informal sector. This is further highlighted by occupation differences. While just under 40% of the informal self-employed have a recorded occupation of elementary occupation, the majority of formal own account workers are legislators; senior officials and managers (see Appendix Table A1). Thus the 29% of the formal self-employed who are own account workers are likely very different firms to the 82% of informal own account workers. By definition, the majority of the informal wage employed work in firms with 2 to 4 employees, while approximately 44% of the formal wage employed work in firms of over 50 people. This strong correlation between employment sector and firm size is discussed elsewhere in the literature, and has been shown to explain a large share of the employment sector wage gaps (El Badaoui et al. 2008). For this reason, it is essential that in estimating simple earnings equations we control for firm size consistently across all sectors. The wage employed are on average younger than the self-employed in both sectors. This is consistent with the notion that entry into self-employment is driven primarily by entrepreneurial drive which may take place after an individual has gained work experience in a wage-employed position. The distribution of geographical location shows that a higher proportion of informal sector (both wage and self-employed) exists in Tribal Authority Areas. This is not surprising given the lack of formal sector businesses in these areas. In contrast, a larger proportion of the formal self-employed are in urban formal areas. With regard to hours worked, private sector employees on average work longer hours per week in comparison with public sector workers, while average tenure in the public sector is higher than in the private sector. This would suggest a higher degree of job security within the public sector. iv. The QLFS wage distribution The absence of formal, or even informal, accounting practices among informal enterprises makes it difficult to collect accurate earnings data. Indeed, 77% of the SESE sample of informal sector

9 businesses keep no business accounts. Nevertheless, the QLFS asks the self-employed a number of separate questions, including a distinct question on their earnings. While wage earners are asked a standard question on their wage, the self-employed are asked for their earnings after expenses (including tips and commission). Both groups are given the option to choose the time frame in which they report their wage/earnings (i.e. monthly, weekly, annual), and both are given the option of providing a numerical or discrete (bracketed) answer. However, the wage/earnings series which are made publicly available (those used in this paper) are reported after imputations and rescaling to provide two comparable monthly wage/earnings series. As information is only collected for an individual s primary source of employment, there is no overlap between the two distributions. We use the log of this variable as our primary variable of interest in our analysis. A data challenge related to including the self-employed in our analysis is tax. El Badaoui et al. (2008) show that taxes may explain a large share of the conditional wage gap between the informal and formal sector wage employed. In fact, they argue that tax and unobserved time-invariant heterogeneity may explain all of the wage gap reported in Kingdon and Knight (2003). Given our definition of the informal sector, the informal sample does not pay income tax. On the other hand, formal sector individuals are asked to report their wage/salary prior to taxation. Supporting our decision to only use gross income is the stylised fact that only a relatively small portion of South Africa's labour force qualifies for income tax. With the 2013 tax threshold at R67,111 for individuals younger than 65, only 45% of our sample would be required to pay tax, suggesting that the any quantile estimates at the median should not be affected. However, switching from net to gross income may change the wage gap at the top end of the distribution. For this reason, we re-run our analysis using net income in place of gross income for the formal sector in the Appendix. Figure 1 shows kernel density plots of the QLFS wage variable for each employment sector as well as the full sample. It is clear that the distributions are distinct on a number of dimensions, including mean, variance, and mode. While both informal sector distributions have a relatively symmetrical shape, the variance of wages among the informal self-employed is higher. This suggests a wider dispersion of self-employment outcomes. Furthermore, the mean of the informal self-employed distribution appears to be in line with the lower mode of the formal wage employed distribution. The formal sector self-employed distribution is skewed to the left, most likely capturing the high returns to self-employed professionals who make up a large share of this sector (see Appendix 1). The public sector wage distribution is bi-modal and also skewed to the left. Note the large gap between the primary modes of the public and formal wage sectors.

10 Figure 1: Distribution of monthly QLFS wages/earnings 2013:Q3 Table 2 gives the mean and median QLFS gross monthly wage/earnings (level and natural log) for each employment sector. Alongside the mean, we report the unconditional OLS wage gap with the formal wage employed as the base category. We trim the wage distribution at the 0.5th and 99.5th percentiles [R65, R per month]. The formal self-employed earn the highest average wages (on average 60% more than the formal wage employed), followed by the public sector and formal wage employed. Both informal sectors earn significantly less than the formal wage employed with an unconditional wage gap of -84% to informal wage employment and -65% to informal selfemployment. The large gap between the average earnings of the informal and formal self-employed once again emphasises the need to recognise heterogeneity among the self-employed in South Africa. Appendix Table A1 shows that more than half of the formal self-employed hold management positions within a business, while 80% of the informal self-employed hold occupations within the elementary, services/retail or craft/trade categories.

11 Table 2: QLFS Wage variable by employment sector Mean Mean Median Proporti OLS Wage ln(wage) ln(wage) on N Formal Sector Wage employed Public sector *** Self employed *** Informal Sector Wage employed *** Self employed *** Total Note: OLS estimates are unconditional. All estimates use nationally representative weights. Cluster-corrected standard errors are reported in brackets. Stars: * p < 0.1, ** p < 0.05, *** p < 0.01 v. Constructing earnings series from the SESE dataset There are two main ways to think about self-employment earnings: the first is to use the net profit of the business and the second the amount of money withdrawn from the business by the owner (Parker, 2005). The Survey of Employers and Self-Employed (SESE) attempts to collect a set of accounts for each enterprise in addition to asking for the business s average profit and the profit for the last month. This allows us to construct both measures of the firm s profits as well as the earnings withdrawn by the business owner. While the QLFS collects only net profit (phrased as earnings after expenses ), we are able to construct 5 additional earnings series using the SESE data: # 1. Reported Profit: amount given in response to a question on the business s profit over the past month. # 2. Average Profit: amount given in response to a question on the business s average monthly profit. # 3. Own Wage: this is derived from a set of questions on wages withdrawn by the owner and on money withdrawn to cover personal expenses. # 4. Reported Earnings: for individuals where the own wage (# 3) is different from last month's profit (# 1), reported earnings are calculated as the sum of the two series under the condition that profits are returned to the household (as reported in a question of the survey). Some respondents may have equated the notion of wages withdrawn by the owner and profits withdrawn to the household which explains our caution towards observations where these two variables are equal. # 5. Calculated Earnings: the firm's profit is calculated using the full set of expenditure (including own wage) and revenue variables found in the survey. If this value is positive and reported to have

12 been withdrawn to the household, total calculated earnings are given by the sum of calculated profit and own wage. These five options give us five earnings series to which we can compare the wage gaps estimated from the QLFS. Since the sample for SESE is drawn from the respondents in the QLFS, one can match the firm details in SESE with the corresponding business owners in the QLFS (see Appendix Table A2 for a detailed account of this matching process). Figure 2 shows the distribution of each of the SESE earnings series alongside the QLFS values for the matched sample. 5 All five SESE distributions appear to be fairly similar in both mean and variance, while the distribution of QLFS wages is centred off to the right of the SESE distributions. This is preliminary evidence that the QLFS estimates may over report self-employment earnings. Figure 2: Distribution of SESE monthly wages 2013:Q3 vi. Motivating for the use of SESE data Despite a similar dispersion (variance) in the overall distributions between surveys, there is both a mean shift in the log of earnings as well as a low correlation between business owners reported earnings (see Table 3). A number of factors could contribute to this weak correlation. First, consider that the series arise from different questions measured over different time frames. The QLFS allows respondents to report their wages or earnings using the following time frames: hour, day, week, fortnight, month or year. In addition, those respondents who do not report a numerical value are given the option of specifying an earnings bracket, from which the final wage is imputed. The final QLFS earnings series reported in the publicly available data has therefore been both rescaled (to 5 The SESE earnings series were also trimmed with the same 0.5 and 99.5 percentiles used to trim the original QLFS wage variable [R65, R ].

13 monthly earnings) and imputed where necessary. While one can observe the time frame of the original reported observation, the original earnings (as measured over an hour, day, etc.) is not provided. Moreover, one cannot identify observations that have been imputed. Table 3. Cross-correlation of QLFS and SESE earnings QLFS Earnings Reported Profit Average Profit Own Wage Reported Earnings Calculated Earnings QLFS Earnings 1 Reported Profit Average Profit Own Wage Reported Earnings Calculated Earnings See Appendix 4 for a mean test of between the QLFS series and each of the SESE earnings series. In comparison, the SESE data is collected using only a monthly time frame and brackets are not provided for respondents. Although there is no one question in the SESE survey that matches the QLFS questionnaire directly, the strong correlation between all the SESE earnings series (Table 3) suggests a greater level of internal consistency. Certainly, a large reason for the high correlation between the options 3, 4, and 5 is due to the repeated inclusion of own wage. Moreover, we show in a companion paper that the rescaling parameter may explain some the differences between the data in the two surveys (Lloyd and Piraino, 2016). Indeed, controlling for the time frame of earnings reporting in the QLFS results in a significantly higher correlation between the QLFS and SESE wage series. Earnings reported using a shorter time frame (e.g. hourly, daily, weekly, fortnightly) tend to be significantly higher (with this difference decreasing in the length of the time frame), while those who report using an annual time frame are significantly lower compared to figures reported for a month. Finally, given the higher propensity for informal self-employed individuals to report their earnings over a shorter time frame (only 44% of the informal self-employed report monthly earnings in the QLFS compared with 80% of formal sector wage earners), this rescaling error generates an upward measurement error bias in informal sector wages. This may result in an under-estimation of the formal-informal sector wage gap. Thus, while there may be reasons to believe that individuals may have an incentive to over- or under-report their earnings in either survey and there is no a priori reason to favour one survey over another, the internal consistency of the five wage series estimated using the SESE data and the evidence of rescaling bias in the QLFS suggests that the SESE earnings data may provide a more accurate picture of informal self-employment earnings in South Africa. IV. ANALYSIS Our empirical analysis proceeds in two steps. First, we estimate both linear and quantile specific Mincer-type earnings equations using the QLFS earnings data. The results of this estimations can be

14 compared to cross-sectional as well as panel wage gap estimates found in the literature. 6 Significant departures should in large part be explained by our choice of data, sample, and sector definitions. In the second step, we re-estimate the cross-sectional results replacing the earnings data for the informal self-employed with the more detailed earnings data from the SESE. We also discuss a number of steps taken to ensure robustness of our results. i. Cross-sectional and fixed-effects wage gaps using QLFS We estimate standard Mincer equations for all employed individuals, excluding agricultural workers, employees of private households and unpaid workers. Table 4 gives the output of both conditional linear and quantile earnings equations. We only report the coefficients of the employment sector categorical variable with the formal wage employed as the reference category. 7 All models include individual and firm-specific controls, including firm size. 8 Model 1 reports the OLS cross-sectional estimates, while Models 2-4 report the results for the cross-sectional quantile regression (QR) at the second, median and eighth decile. Both the OLS and QR results use the full 2013:Q3 cross section. To ensure that the results are not driven by sample composition and the exclusion of taxes, we reestimate these models for the sample of African males and for net wages (see Appendix Table A3 and A4). The sectoral wage-gap parameters reported in models 1-4 are only identified under very strict exogeneity assumptions, as well as model specification assumptions. However, selection into each of these sectors is almost certainly not exogenous. Under the assumption that the factors determining chioice of sector are time-invariant one can control for selection by using individual fixed effects. This approach, which alleviates the need to find an instrumental variable in what is otherwise a fairly limited dataset, is used by Bargain & Kwenda (2014). Canay (2011) provides a consistent quantile fixed-effects estimator with a simple two-stage sample procedure. In the first stage, the fixed-effect component is estimated using a linear fixed-effects estimator. In the second stage, a quantile regression is performed using the difference between the original dependent variable and the fixed effect (estimated in the first-stage) as the new dependent variable. 9 Canay (2011) shows that this estimator is consistent. Models 6-8 in Table 4 report the results of the fixed-effects quantile regression (FE-QR) alongside the first stage regression in model 5. 6 We use a two-period panel (Q2-Q3 of the 2013 data) to estimate linear fixed effects as well as a two-stage fixed effects quantile regression model introduced by Canay (2011) and used in Bargain & Kwenda (2014). 7 A methodological challenge in this regard is that by pooling the labour market sectors and jointly estimating Mincer-type earnings equations (allowing only for a change in the constant by employment sector), we have implicitly assumed that the earnings function is the same in each sector, an assumption which is seldom made in the literature (Dickens & Lang 1984; Heckman & Hotz 1986; Günther & Launov 2006). Heintz and Posel (2008) reject the assumption that the earnings equation is the same across sectors in South Africa. However, after estimating the equations separately, they still find significant evidence that switching from informal to formal sector is associated with an increase in wages. 8 In line with the literature we find that firm size explained a large portion of the informal sector wage gap. 9 Standard errors are obtained using bootstrap techniques.

15 Table 4: Conditional, linear and non-linear earnings equations using QLFS Model (1) (2) (3) (4) (5) (6) (7) (8) Estimator OLS QR FE FE-QR Formal sector Mean Q=.2 Q=.5 Q=.8 Mean Q=.2 Q=.5 Q=.8 Public sec *** *** 0.242*** *** *** *** (0.0331) (0.0401) (0.0282) (0.0214) (0.046) ( ) ( ) (0.002) Self-empl *** *** *** *** *** Informal sector (0.0990) (0.149) (0.0967) (0.0794) (0.174) (0.0291) ( ) (0.025) Wage empl *** *** *** *** *** *** *** (0.0834) (0.106) (0.0455) (0.0716) (0.0862) (0.0140) ( ) (0.008) Self-empl *** *** *** *** 0.101*** (0.110) (0.0896) (0.0821) (0.0851) (0.134) (0.0355) ( ) (0.0311) N Note: In all models the base category is Formal Wage Employed. OLS & QR: Individual controls include age, age squared, gender, race, education, education squared, marriage status, household size, province, and geographical location. Job controls include occupation, log of hours worked, tenure, tenure squared. OLS standard errors are corrected for sample clustering and estimates are weighted. The coefficients displayed in Table 1 can be summarized as follows. On average, public sector workers earn 10.6% more than formal-sector wage-employed workers, holding constant observable individual and job characteristics. This premium appears to be higher for the median public sector worker, explained by the bimodal public sector wage distribution observed in Figure 1. While there is no significant difference between average formal self-employed earnings and formal sector wage earnings, this masks a large positive premium at the top end of the distribution and a large negative premium at the lower end of the distribution. This is further evidence of the larger dispersion of self-employment earnings, even after controlling for firm size. Contrasting this is the almost flat negative premium associated with informal wage-employment. The premium to wage employment in the formal sector over informal sector, estimated at 40%, appears to be relatively linear controlling for observable characteristics. This is comparable with the El Badaoui et al. (2008) informal premium of -37% (estimated using only the wage employed) using gross wage. This is smaller than the Bargain and Kwenda s (2014) estimate of -62%, which is also fairly flat along the distribution. The first set of FE-QR results resemble those of Bargain and Kwenda (2011). Models 5-8 suggest that there is a -11% wage gap at the mean, median and second decile to informal wage employment. An equivalent estimate from El Badaoui et al. (2010) and Bargain and Kwenda (2014) is -18%. Thus, individual heterogeneity explains a large share of the informal wage employed gap observed across the wage distribution in the models 2 to 4. The estimates also suggest that the there is no conditional mean difference between the informal self-employed and formal wage employed. However, this is again masking a positive earnings

16 premium at the top end of the distribution and a negative premium at the lower end. The coefficients suggest that there is higher variation among the earnings of self-employed individuals even after controlling for firm size and other individual characteristics. The dispersion of these wage premiums is suggestive of the risk carried by small business owners in comparison to their wage employed counterparts. In model 8 we find a 10% positive wage premium to informal self-employment at the eighth percentile using the QLFS data. This suggests premium at the top of the wage distribution is not evident among the formal self-employed. In fact, they are the only sector in which one finds this positive premium. This is a similar to Bargain and Kwenda (2011) who estimate a -11.6% gap to informal wage employment and +12% premium to informal self-employment at the upper end of the distribution. Our results are therefore relatively comparable, despite the important differences in sample restrictions and income definitions. ii. Cross-sectional estimates using SESE data As previously discussed, there are a number of reasons why the data in the QLFS may be a biased measure of true self-employment earnings. The approach taken here is a simple one: we propose that given the short time lap (of about one/two weeks) between the QLFS and the SESE, data from the latter can be used as a substitute for the QLFS data (StatsSA, 2014). Below we report the same set of estimates using a cross-sectional wage series where the informal self-employed s QLFS earnings are replaced by their corresponding SESE amounts. In Table 5 we substitute the QLFS wage series with the calculated earnings (#5) of the informal self-employed from the SESE. Table 5: Cross-sectional estimates with SESE Adjustment #5 Sample Full Cross-section Model (1) (2) (3) (4) Estimator OLS QR Mean Q=.2 Q=.5 Q=.8 Formal sector Public sector 0.113*** *** 0.261*** (0.0334) (0.0397) (0.0293) (0.0215) Self-employed *** *** (0.101) (0.139) (0.100) (0.0949) Informal sector Wage employed *** *** *** *** (0.0828) (0.125) (0.0522) (0.0754) Self-employed *** *** *** *** (0.107) (0.144) (0.0947) (0.112) N Note: In all models the base category is Formal Wage Employed. OLS & QR: Individual controls include age, age squared, gender, race, education, education squared, marriage status, household size, province, and geographical location. Job controls include occupation, log of hours worked, tenure, tenure squared. OLS standard errors are corrected for sample clustering and estimates are weighted.

17 Comparing the results of Tables 4 and 5, we notice the stark difference between both the mean and quantile earnings-gaps. The insignificant mean effect estimated in Table 4 is now replaced by a significant coefficient of (also evident in Figure 2). Moreover, the dispersion of wage gaps along the wage distribution estimated in Table 4 which went from -0.4 to 0.35 is now -0.9 to This suggests that the substitution of wages from the SESE questionnaire has a larger impact on the wage gaps at the top end of the distribution. As a robustness check for this model we re-estimate them including only wage observations reported as monthly wages, in accordance with the potential for measurement error in scaled variables (see Talbe A5). The results are suggestive that the size of the rescaling error is independent across sectors. Excluding higher frequency reports from the base group reduces the wage gap from to at the eighth percentile. We also report the mean and eighth percentile results for the four other SESE earnings candidates (#1 to #4) in the Appendix. The profit and earnings results are not significantly different from those reported in Table 5, while the own wage estimates are larger in value. iii. Fixed-effect estimates using SESE data The FE-QR results of Table 4 (models 5-8) cannot be easily replicated using the SESE data, as SESE is a cross-section survey. Existing approaches used in the literature to control for individual heterogeneity, such as those used by Falco et al. (2011), can offer a solution. However, any approach to control for individual time-invariant heterogeneity relies on estimates of individual fixed-effects using QLFS and not the SESE data. Whether these estimates are consistent depends on the timevarying nature of the bias in the QLFS data. [TO COMPLETE ] [TABLE 6 HERE] V. CONCLUSION The seeming incompatibility of South Africa s sustained high levels of unemployment and relatively small informal sector has led to a wealth of research on labour market segmentation. Recently, the literature on the informal sector wage-gap in South Africa has highlighted the fact that there may be a positive premium to informal self-employment at the top end of the wage distribution. These estimates suggest a higher dispersion of informal self-employment earnings conditional on individual fixed effects as well as observable individual and firm characteristics.

18 This paper investigates further the informal sector wage gap by introducing into the conversation new data from the Survey of Employers and Self Employed. This survey was administered concurrently with the 2013:Q3 Quarterly Labour Force Survey and offers a more detailed look at informal enterprises in South Africa. While the simple substitution approach adopted in this paper may not be robust enough to refute previous findings in the ltieratire, the results certainly suggest th possibility of a different picture. Our estimates suggest a harsher reality for South Africa s informal self-employed, one in which there is no positive wage premium at the top of the distribution. [ADD DISCUSSION OF FIXED EFFECT RESULTS HERE] REFERENCES El Badaoui, E., Strobl, E. & Walsh, F., Is there an informal employment wage penalty? Evidence from South Africa. Economic Development and Cultural Change, 56(3), pp El Badaoui, E., Strobl, E. & Walsh, F., The formal sector wage premium and firm size. Journal of Development Economics, 91(1), pp Banerjee, A. et al., Why has unemployment risen in the New South Africa? 1. Economics of Transition, 16(4), pp Bargain, O. et al., The formal sector wage premium and firm size for self-employed workers, Bargain, O. & Kwenda, P., Earnings Structures, Informal Employment, and Self-Employment: New Evidence from Brazil, Mexico, and South Africa. Review of Income and Wealth, 57, pp.s100 S122. Bargain, O. & Kwenda, P., The Informal Sector Wage Gap: New Evidence Using Quantile Estimations on Panel Data. Economic Development and Cultural Change, 63(1), pp Bertola, G. & Garibaldi, P., Wages and the Size of Firms in Dynamic Matching Models. Review of Economic Dynamics, 4(2), pp Blanchflower, D. & Oswald, A., Entrepreneurship, Happiness and Supernormal Returns: Evidence from Britain and the US, Bulow, J.I. & Summers, L.H., A Theory of Dual Labor Markets with Application to Industrial Policy, Discrimination and Keynesian Unemployment, Canay, I.A., A simple approach to quantile regression for panel data: A simple approach to quantile regression for panel data. The Econometrics Journal, 14(3), pp Carneiro, F.G. & Henley, A., Wage determination in Brazil: The growth of union bargaining power and informal employment. The Journal of Development Studies, 34(4), pp Cichello, P.L., Hindrances to self-employment activity: evidence from the 2000 Khayelitsha/Mitchell s Plain survey, Centre for Social Science Research, University of Cape Town. Devey, R., Skinner, C. & Valodia, I., Informal economy employment data in South Africa: A critical analysis. Report prepared for the Employment Data Research Group, Human Sciences Research Council.