RRpage7-dp101

“Heteroscedastic Sample Selection and Developing Country Wage Equations”

Julie Anderson Schaffner

IED Discussion Paper 101, June 1999

Appropriate strategies for harnessing manpower and developing human capital hinge upon accurate estimates of the returns to schooling. Econometric estimates of this crucial parameter are, however, subject to a number of statistical biases. One of the most important of these results from selectivity problems, wherein a significant fraction of the relevant population do not secure employment at all, and are thus excluded from datasets employed in estimation.

It has become standard practice to deal with potential selectivity bias in developing country wage equations by employing Heckman’s two-step method or related techniques. This procedure may, however, produce misleading results if the statistical assumptions underlying this method are incorrect. In her paper, Schaffner demonstrates that taking a more flexible approach to the estimation of wage equations in endogenously selected samples is possible even in the context of easy-to-implement, parametric methods. This can be done by using model selection

tests and sensitivity analysis to discern the empirical importance of departures from three standard assumptions: normality of the selection rule error, normality of the wage equation error, and homoscedasticity of the selection rule error. In particular, the paper provides an illuminating discussion of the economic reasons to suspect selection rule heteroscedasticity, in which individuals with more schooling are prone to larger prediction errors. For example, unobserved school quality and years of schooling might be expected to have interactive effects upon the propensity to be a wage earner, generating heteroscedasticity. She provides econometric reasons to believe that the non-linearities introduced by heteroscedasticity might aid identification.

In an application to urban Peru, homoscedasticity is strongly rejected in favour of more flexible alternatives. Allowing for the alternative hypothesis of heteroscedasticity causes estimates of key parameters to become more robust to changes in other statistical assumptions. Moreover, the paper demonstrates the importance of the more flexible approach to produce believable estimates of other parameters such as the covariation of selection rule and wage equation errors which describes the nature of labor market sorting.