Sample selection bias and the South African wage function

Conventional wage analyses suffers from a debilitating ailment: since there are no observable market wages for individuals who do not work, findings are limited to the sample of the population that are employed. Due to the problem of sample selection bias, using this subsample of working individuals to draw conclusions for the entire population will lead to inconsistent estimates. Remedial procedures have been developed to address this issue. Unfortunately, these models strongly rely on the assumed parametric distribution of the unobservable residuals as well as the existence of an exclusion restriction, delivering biased estimates if either of these assumptions is violated. This has given rise to a recent interest in semi-parametric estimation methods that do not make any distributional assumptions and are thus less sensitive to deviations from normality. This paper will investigate a few proposed solutions to the sample selection problem in an attempt to identify the best model of earnings for South African data.

C14, C15, C34, J21

Semiparametric and nonparametric methods; Simulation methods; Truncated and censored models; Labour force and employment, Size, and structure