Problems with SWIID: the case of South Africa

The information contained in databases of summary statistics should look plausible when viewed in context. Judged by that criterion the Standardized World Income Inequality Database (SWIID) comes up short with its South African data. Figure 1 contains the series as extracted from the SWIID web site (Solt 2014b). The 95% confidence bands suggest that inequality in 1965, at the height of apartheid, was significantly lower than in the 2000s. This, however, flies in the face of much other evidence. For instance it is well-known that Black mine workers' wages were static in real terms from the early twentieth century right up to the 1970s (e.g. van der Berg 1989). The wages of white miners, by contrast, increased, so that the ratio of White to Black mine wages reached its maximum of twenty to one in 1969 (Devereux 1983, p.18). Simkins (1979) estimated the Gini coefficient in 1970 at 0.71, which seems more in line with the political and social realities. Trying to understand how the SWIID may have arrived at such a misleading estimate is instructive about the types of problems that may be lurking elsewhere in the database.There are four potential sources of error in SWIID: measurement error, model error, imputation error and sampling error.