Skip to content

The Mystery of South Africa's Ghost Workers in 1996: Measurement and Mismeasurement in the Manufacturing Census, Population Census and October Household Surveys

Literature › Works

Martin Wittenberg (2004). The Mystery of South Africa's Ghost Workers in 1996: Measurement and Mismeasurement in the Manufacturing Census, Population Census and October Household Surveys. School of Economics, University of Cape Town (unpublished paper, dated 2 September 2004). School of Economics, University of Cape Town.
Type working paper
Scope Comparative and methods
Method descriptive quantitative
Period studied 1995–1997
Regions South Africa, Gauteng, Western Cape, KwaZulu-Natal, Eastern Cape, North West
Themes methods data sources, labour markets wages, transition post apartheid, manufacturing industrial policy
Citation key wittenberg2004

Argument

Wittenberg compares three 1996-era data sources on South African manufacturing employment: the Census of Manufacturing, the Population Census and the October Household Surveys (OHS). The Population Census records about 300,000 fewer manufacturing workers than the manufacturing census. He rules out explanations in turn: employer over-reporting, non-enumeration, and misclassification of employment status. He then shows that misclassification into 'industry not elsewhere classified/unspecified' cannot fully account for the gap. The household surveys are themselves noisy and inconsistent by sector. He concludes that the census figure is too low, the puzzle is unsolved, and the cross-source discrepancies make year-to-year comparisons of aggregates hazardous.

Evidence and method

Census of Manufacturing 1996 regional reports (Statistics South Africa), the 1996 Population Census community profile databases (extracted via SuperCross), and the 1995, 1996 and 1997 October Household Surveys with Statistics South Africa weights. Confidence intervals take clustering into account but disregard stratification. The text has four comparison tables by sub-region, province and industry, plus cross-checks against Klasen and Woolard (1999) and Bhorat (2001, 2003).

Findings

  • The manufacturing census and population census disagree sharply on 1996 manufacturing employment. (p. 2)

    The manufacturing census of 1996 reports a total of 1.4 million workers and an additional seven thousand “working proprietors”. By contrast the population census of the same year only records 1.1 million people employed in manufacturing.

  • Employer over-reporting in the manufacturing census is implausible. If anything the census is an underestimate, which deepens the puzzle. (p. 3)

    One would expect the total employment estimate in the manufacturing census to be an underestimate rather than an overestimate. The mystery of the disappearing manufacturing workers in fact deepens.

  • The OHS point to about 1.5 million manufacturing jobs, so the manufacturing census may itself have missed some workers. (p. 3)

    According to the 1995 and 1997 OHS it appears as though total manufacturing employment should be of the order of 1.5 million. This would suggest that the manufacturing census missed about eighty thousand workers.

  • Household survey totals are volatile and carry wide confidence intervals, so trend inference from aggregates is risky. (p. 3)

    Looking across the three surveys it is clear that not only are the confidence intervals large, but the point estimates jump around a lot also.

  • The population census has a large residual industry category, which at first suggests misclassification. (p. 4)

    it turns out that there are 1.1 million individuals in the category “Industry not elsewhere classified or unspecified”. There is therefore ample room to believe that the disappearance of the manufacturing workers may be more apparent than real.

  • Deficits in trade and services in the population census are large enough to account for the whole residual category, undermining the misclassification story for manufacturing. (p. 4)

    there is a deficit of about 500 000 retail and wholesale workers and about 460 000 service workers. Given that the “residual” category is never less than 200 000, these two deficits are sufficient to explain the entire residual category!

  • The household surveys under-record mining employment, which favours the census on that sector. (p. 4)

    mining employment in the population census is significantly higher than it is in any of the surveys speaks in favour of the census. It is well known that the surveys were badly designed to pick up mining employment

  • Gauteng's deficit would require an implausibly high share of the residual category to be manufacturing workers, and different misclassification processes from the other provinces. (p. 5)

    We would require fully 75% of everyone who was recorded under “industry not elsewhere classified and unspecified” in that province to have been a manufacturing worker

  • The puzzle is left unresolved, though the census number appears far too low. (p. 6)

    The evidence seems unambiguous that the population census number is far too low. It is much more difficult to pin down where those workers have gone to.

  • The manufacturing census total is a realistic lower bound, and the author urges stronger standard-error treatment for survey-based totals. (p. 6)

    The Manufacturing Census figure of 1,427,502 employees and working proprietors seems like a realistic lower bound on manufacturing employment figures for 1996.