A Test for Autocorrelation in Models with Lagged Dependent Variables

Abstract

The strength of the relationship between schooling and the logarithm of earnings at different levels of experience can be measured by coefficients of determination or by residual variances. For tracking unobserved post-school investments in cross-sections, the latter are clearly preferable; the greater reliance on the former in the literature is misplaced. If the fraction of earning capacity devoted to postschool investment is uncorrelated with earning capacity at school-leaving, the relationship between schooling and the logarithm of earnings should be strongest at overtaking, which would plausibly be placed in the first decade of work experience. Mincer reported coefficients of determination which followed this pattern, at least for full-year workers. An analysis of more recent data failed to detect this pattern for coefficients of determination, but found more favorable evidence for residual variances using weekly (but not hourly)