Measurement is crucial to the implementation of output-based incentive schemes. This paper uses models to study the performance of teacher quality estimators that enter teacher incentive schemes. I model an administrator tasked with (i) categorizing teachers with respect to a cutoff, (ii) retaining teachers in a hidden type environment, and (iii) compensating teachers in a hidden action environment. The preferred estimator would be the same in each model and depends on the relationship between teacher quality and class size. I use data from Los Angeles to show that simple fixed effects would almost always outperform more popular empirical Bayes.