Reducing Bias in Teacher Value-Added Estimates by Accounting for Test Measurement Error

• In Event: Issues in Clustered Data Structures

Abstract

The specification of value-added (VA) models for estimating teacher effects that are minimally biased by omitted student characteristics is a pressing issue in the VA research and policy communities. The ANCOVA method in which current test scores are regressed on past test scores, student characteristics and current teacher indicators is gaining popularity, but measurement error in the test scores may erode its ability to adequately adjust for student factors. We develop an enhanced version of the ANCOVA model which accounts for test measurement error using Bayesian latent regression modeling. Applying the new method to longitudinal data from a large suburban school district demonstrates that its VA estimates appear less biased than those from ANCOVA on metrics of potential bias.

Authors

• J. Lockwood, RAND

• Daniel McCaffrey, RAND Corporation