Working Paper: NBER ID: w20657
Authors: Andrew Bacher-Hicks; Thomas J. Kane; Douglas O. Staiger
Abstract: In a widely cited study, Chetty, Friedman, and Rockoff (2014a; hereafter CFR) evaluate the degree of bias in teacher value-added estimates using a novel "teacher switching" research design with data from New York City. They conclude that there is little to no bias in their estimates. Using the same model with data from North Carolina, Rothstein (2014) argued that the CFR research design is invalid, given a relationship between student baseline test scores and teachers' value-added. In this paper, we replicated the CFR analysis using data from the Los Angeles Unified School District and similarly found that teacher value-added estimates were valid predictors of student achievement. We also demonstrate that Rothstein's test does not invalidate the CFR design and instead reflects a mechanical relationship, given that teacher value-added scores from prior years and baseline test scores can be based on the same data. In addition, we explore the (1) predictive validity of value-added estimates drawn from the same, similar, and different schools, (2) an alternative way of estimating differences in access to effective teaching by taking teacher experience into account, and (3) the implications of alternative ways of imputing value-added when it cannot be estimated directly.
Keywords: Teacher Effectiveness; Value-Added Estimates; Student Achievement; Education Policy
JEL Codes: I21; J45
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Teacher assignments (Y40) | Student achievement (I24) |
| Teacher value-added estimates (C13) | Student achievement (I24) |
| Top quartile teacher assignment (I24) | Student achievement (I24) |
| Teacher effectiveness differences (I24) | Achievement gaps (I24) |
| Missing value-added data treatment (C22) | Predictive validity estimates (C52) |
| Teacher value-added estimates across different contexts (C13) | Predictive validity (C52) |