Research on the impact of teachers on student achievement (e.g., Jonah E. Rockoff 2004; Steven G. Rivkin, Hanushek, and John Kain 2005) has established two stylized facts: (1) teacher effectiveness varies widely, and (2) outside of experience, qualifications that determine a teacher’s certification and salary bear little relation to outcomes. This provides motivation to understand how to identify effective and ineffective teachers, particularly early in their careers. Studies that examine how student achievement data can predict teachers’ impacts on student outcomes in the future (e.g., Robert Gordon, Thomas J. Kane, and Douglas O. Staiger 2006; Dan Goldhaber and Michael Hansen 2010) conclude that using such data to selectively retain teachers could yield large benefits. However, “value-added” measures of effectiveness are noisy and can be biased if some teachers are persistently given students that are difficult to teach in ways that are hard to observe. Thus, using other information may achieve more stability and accuracy in teacher evaluations. There is also a literature on subjective teaching evaluations (i.e., evaluations by the school principal or evaluations based on classroom observation protocols or “rubrics”), which also finds significant relationships between evaluations and achievement gains. However, these studies typically investigate how evaluations predict the exam performance of current, not

This paper presents findings from a national experimental evaluation of performance bonuses funded by the Teacher Incentive Fund grant program. The study finds no robust evidence of positive impacts of bonuses on student achievement, although some specifications suggest a small positive effect that compares favorably with other education interventions. When controlling for covariates, we find that offering bonuses of an average yearly cost of $100 per student had a small significant impact of about 0.04 standard deviations. However, these impacts are smaller (0.01 standard deviations) and become insignificant when not controlling for covariates or using an alternative method of inference.