Measuring the returns to education nonparametrically

Abstract

This article uses a nonparametric model of earnings to measure the returns to education. Under very general smoothness conditions, a nonparametric estimator reveals the true shape of the earnings profiles up to random sampling error. Thus, the nonparametric model should provide better predictions than its parametric counterpart. We find that the nonparametric model predicts very different estimated returns than standard Mincer formulations. Depending on the experience and education level, returns measured in log earnings estimated from nonparametric model can be nearly twice those obtained from the Mincer model. Finally, this article examines what structural features parametric models should include.

Acknowledgements

I wish to thank Joel Horowitz, John Geweke, and Qi Li.

The views in this article are those of the author and do not necessarily reflect those of the Federal Trade Commission or any individual Commissioner.

Notes

¹ (Antecol and Bedard 2004) show that predicted experience may give different results than actual experience, but one cannot avoid using the former when the latter is not available.

² See (Blackburn et al. 1990); (Bound and Johnson 1992); (Katz and Murphy 1992); (Murphy and Welch 1992); and (Katz and Autor 1999).

³ See, for example, (Angrist and Krueger 1999); (Card (1999); Park (1999); (Ulrick 2005). The scheme in this article is also similar to one recommended by Jaeger and Page (1996), with the largest difference being that we distinguish between individuals with/without a H.S. diploma, while Jaeger does not.

⁴ See (Angrist and Krueger 1991, 1999) for applications and surveys.

⁵ OLS returns are often reported in applications (e.g. Psacharopoulos, 1992; Maxwell, 1999; among others).

⁶ In this test, the fitted values are obtained from the parametric regression. If the parametric model is properly specified a nonparametric mean regression of on y_i should result in a 45^° line. Uniform confidence bands can be generated from such a regression, as well as a 45^° line. If the model is properly specified, the 45^° line should lie entirely inside the confidence bands. None of the regressions of on y_i lie within 90% uniform confidence intervals, and thus all of the parametric models are rejected.