Skip to main content

Statistical Methods to Control for Covariates in Teacher Value-Added Assessment

Abstract

This project examines four models used to estimate value-added teacher effects: (1) no controls for student characteristics; (2) student fixed effects; (3) student characteristics with teacher fixed effects; (4) student characteristics without fixed effects. The first part of the project comprises a theoretical analysis of the properties of these methods. It establishes that none is guaranteed to be free of bias and sets out the conditions under which the bias is less in one method than in others. The empirical work has two components: a Monte Carlo investigation focused particularly on the trade-offs between methods (3) and (4), which have not been explored heretofore in the literature; and estimation of teacher value added using panel data on mathematics instructors in Dade County, Florida. Because the trade-offs just mentioned depend on the distribution of variables that are usually unobservable, it is particularly valuable to have a data set that contains a richer set of student characteristics than is typically available to researchers, permitting us to examine the distribution of variables that usually are not observed. The Dade County data meet this requirement. Preliminary analysis of the data has already shown that the choice of methods has non-trivial effects on the evaluations of individual teachers.