The Return to Schooling in Structural Dynamic Models: A Survey

This paper contains a survey of the recent literature devoted to the returns to schooling within a dynamic structural framework. I present a historical perspective on the evolution of the literature, from early static models set in a selectivity framework (Willis and Rosen, 1979) to the recent literature, stimulated by Keane and Wolpin (1997), and which uses stochastic dynamic programming techniques. After reviewing the literature thoroughly, I compare the structural approach with the IV (experimental) approach. I present their commonalities and I also discuss their fundamental differences. To get an order of magnitude, most structural estimates reported for the US range between 4% and 7% per year. On the other hand, IV estimates between 10% and 15% per year are often reported. The discrepancy prevails even when comparable (if not identical) data sets are used. The discussion is focused on understanding this divergence. The distinction between static and dynamic model specifications is a recurrent theme in the analysis. I argue that the distinction between the IV approach and the structural approach may be coined in terms of a trade off between behavioral and statistical assumptions. For this reason, and unless one has very specific knowledge of the true data generating process, it is neither possible, nor sensible, to claim which approach to estimation is more flexible. More precisely, I show that structural and IV approaches differ mainly at the level of i) the compatibility of the underlying models with truly dynamic behavior, ii) the role of heterogeneity in ability and tastes, iii) the consideration of post-schooling opportunities, and (iv) the specification (and interpretation) of the Mincer wage equation.