Are value-added models good enough for teacher evaluations? Assessing commonly used models with simulated and actual data

Teachers’ evaluations in many states include information about their students test score gains. In this paper, we describe the assumptions that are required for teacher value-added (TVA) estimates to be treated as unbiased causal effects. We compare commonly used TVA models on policy-relevant criteria using simulated data in which the assumptions of unconfounded assignment of students and teachers and no peer effects are violated and with actual data. The three-level hierarchical linear performs best when either assumption is violated. For year-to-year consistency, the dynamic ordinary least squares model performs best. A common policy goal – identifying the lowest performing quintile of teachers—can be done with reasonable accuracy but between 3.2 and 9.3 percent of all teachers are misclassified.