Adding Value to Value-Added: Theory and Applications to Teachers and Bureaucrats

Abstract

Value-added estimators are extensively used to study teachers and other groups. In addition to estimating an individual's causal impact on outcomes, value-added models describe the dispersion of individual effects: For example, variation in teacher quality contributes about 2% of the variance in their students' same-year test scores. In this dissertation, I explore the use of value-added estimators in different contexts and discuss the statistical properties of these estimators. In the first chapter, I model teacher value-added as a vector encompassing teacher effects on many outcomes. The covariance matrix of teacher value-added reveals that teachers have large effects on high school graduation and future test scores, but that immediately-observable teacher effects are poor proxies for future effects. Teacher effects on attendance are about as persistent and as predictive of future achievement as teacher effects on test scores. In the second chapter, I discuss statistical issues that arise in finite samples. In the third chapter, joint with Jonas Hjort and Gautam Rao, I take an in-depth look at one context in which relatively small data induces finite-sample biases: the effects of bureaucrats on local economic outcomes in India. Although point estimates suggest that bureaucrat quality is an important determinant of variation in night light intensity and project completion, randomization inference shows that point estimates are biased upwards and insignificant.