School Spending and Student Achievement : Using Specification Tests to Eliminate Inconsistent Estimators

This paper tests the specification of the typical (and a few atypical) models for estimating the relationship between educational resources and student achievement, using data from the National Education Longitudinal Study of 1988. Using specification tests which do not rely upon the existence of a known-to-be-consistent estimator or on the availability of valid instrumental variables, our analysis rejects the null hypothesis of no misspecification for all of the models used in the literature reviewed by Hanushek (1986). This finding also holds for value-added models which measure "inputs" at the classroom level and allow for nonlinear resource effects. Two-stage estimators using school, state and ZIP code level instrumental variables, inspired by the previous literature, are also unable to produce consistent estimates for the parameters of interest. It is possible that consistent estimates could be produced using panel data which cover a longer sampling period than anything currently available; in the interim, public policymakers should be made aware that nonexperimental estimates are subject to potentially significant biases