The Effect of Teacher Gender on Students of Differing Ability: Evidence from a Randomized Experiment
Abstract: Gender dynamics may play an important role in the determination of student outcomes in education. To date, the study of student/teacher gender dynamics has focused mostly on average effects. Exploiting random assignment of students to teachers in a field experiment, I study heterogeneity in the impact of teacher gender on the math and reading test scores for primary school students of differing ability. I find that assignment to a female teacher is generally positive for male students while having no significant effect for female students. In addition, I find very little heterogeneity in the effect of teacher gender on the ability axis, suggesting that average effect estimates do not mask significant heterogeneity. My results are consistent with differential teacher behavior based on gender stereotypes, and somewhat inconsistent with differential student behavior based on gender stereotypes.
Matching as Weight Selection: A Framework for Evaluating Matching Estimators
Abstract: Matching estimators provide an intuitively appealing approach to program evaluation. Due to the non-smooth nature of matching algorithms, the bias/variance trade-off from switching between different matching methods is opaque, leaving practitioners with little guidance when choosing a matching algorithm or smoothing parameters. I cast matching estimators as a subset of a larger class of weighting estimators and consider infeasible optimal weights. I use the insights derived from this to identify areas in which certain matching procedures have clear advantages or disadvantages.
A Note on Bootstraps for Matching Estimators
Abstract: Matching estimators are a popular approach to program evaluation in empirical literature. Abadie and Imbens (2008) showed that the naive bootstrap fails to provide valid inference on such matching estimators, and conjectured that a wild bootstrap might solve this problem. Otsu and Rai (2016) confirm this conjecture for the general case. I consider a simpler bootstrap and demonstrate that even with significantly stronger assumptions, their bootstrap cannot be improved. Simulations provide further support for this conclusion.