Supervisor: Professor Servaas van der Berg
Institution: Stellenbosch University, Faculty of Economic and Management Sciences, Department of Economics
In order to evaluate the extent to which South Africa achieve the objectives of poverty and inequality reduction as well as job creation, up-to-date and reliable data are required. Since the transition, various survey data have been commonly used for these analyses, namely Census, Community Survey (CS) 2007, Income and Expenditure Survey (IES), October Household Survey (OHS), Labour Force Survey (LFS), Quarterly Labour Force Survey (QLFS), General Household Survey (GHS), Project for Statistics on Living Standards and Development (PSLSD), National Income Dynamics Study (NIDS) and All Media Products Survey (AMPS). However, these datasets are not fully comparable, due to differences in the sampling design, sample size, questionnaire structure, methodology to derive labour market status, as well as the way the income and expenditure information was collected. Hence, this dissertation begins by analysing these issues in each survey in Chapter 2. With regard to the income and expenditure information, it was collected differently in the surveys: the recall method was used in all surveys except IES 2005/2006, the only survey that adopted the diary method; respondents were asked to report the actual amount in some surveys but only asked to declare the relevant interval in others; for the former approach, respondents could either declare the single estimate amount or amounts for sub-categories that were then aggregated; for interval data, various methods can be used to determine the amount in each interval. Thus, Chapter 3 begins by discussing the merits and drawbacks of these approaches, as well as how they would affect the reliability and comparability of income and expenditure variables across the surveys. In some surveys (e.g., the two censuses and CS 2007), quite high proportions of households incorrectly reported zero income or expenditure or did not specify their income or expenditure. Poverty and inequality estimates could be influenced by either including or excluding these households from the analyses. Hence, various approaches to deal with these households are examined in Chapter 3. As the surveys typically under-captured income or expenditure when compared with the national accounts income, the validity of the resultant poverty and inequality estimates might be affected. Hence, arguments for and against adjusting the survey means in line with the national accounts mean (e.g. by shifting the survey distribution rightwards) are discussed. As the survey data are, strictly speaking, crosssectional and not designed for time-series labour market, poverty and inequality analyses, it is sometimes argued that the data should be re-weighted to be consistent with demographic and geographic numbers presented by the Actuarial Society of South Africa (ASSA) and Census data. This cross entropy re-weighting approach is discussed in Chapter 3. Finally, the chapter examines the labour market status derivation methodology in all OHSs, LFSs and QLFSs in greater detail, and investigates how the changes across the surveys could possibly affect the comparability of labour market estimates throughout the years. The dissertation then examines the labour market trends since the transition by using the OHS, LFS and QLFS data, and it is found that both the labour force and employment numbers increased in general since the transition, but the latter increase was not rapid enough to absorb the expanding labour force. In addition, the number of narrow unemployed doubled between 1994 and 2009, and the narrow unemployment rate showed an upward trend and peaked at just above 30% in 2003. It decreased between 2004 and 2007, before rising again in 2008- 2009 due to the impact of global recession. Application of the cross entropy approach does not substantially affect labour market trends, suggesting that the trends (including the abrupt increase in labour market estimates during the changeover from OHS to LFS) were either real or took place due to the improvement of the questionnaire to capture the labour market status of the respondents better. Furthermore, the application of the LFS 2000b-LFS 2007b methodology on the earlier surveys reduced the extent of the abrupt increase of the number of broad unemployed and broad unemployment rates during the changeover between OHS and LFS. Finally, the use of the QLFS methodology (which required minor revisions) on the LFSs greatly reduced the extent of the abrupt decrease of unemployment aggregates between LFS 2007b and QLFS 2008Q1, thereby improving the comparability of these aggregates across the surveys. In Chapter 5 poverty and inequality concepts are reviewed, followed by a detailed explanation of the sequential regression multiple imputation (SRMI) technique to deal with households with zero or missing income or expenditure, as well as the derivation of real income, expenditure and consumption variables in each survey. Poverty and inequality trends since the transition are examined in Chapter 6. With regard to poverty, with the exception of AMPS, the poverty trends were very similar across the surveys, that is, poverty increased since the transition, before a downward trend took place since 2000. As far as inequality is concerned, both the levels and trends in the Gini coefficients differed a lot amongst the surveys, as the estimates were very stable in the AMPSs, showed an upward trend in surveys like IESs and GHSs, but first increased until 2000 before a downward trend took place in others (e.g., the two censuses and CS 2007). The levels of inequality also differed when comparing the surveys. The abovementioned poverty and inequality estimates and trends could in part be affected by the various issues discussed in Chapter 3, thus there is a need for careful analysis. The impact of the number and width of intervals in which income or expenditure data are recorded on poverty and inequality estimates and trends are dealt with in greater detail in Chapter 6 by applying various intervals on the three IESs and NIDS 2008. It is found that the number and width of intervals only had some impact on these estimates and trends in some surveys. The effect of adjusting the survey means in line with the national accounts mean is also investigated. Finally, the application of the cross entropy re-weighting technique did not have any significant impact on the poverty and inequality estimates and trends.