DataScience@BI seminar with Associate Professor Tom Boot

Inference on LATEs with covariates

Abstract
In theory, saturated two-stage least squares identifies a weighted average of covariate-specific local average treatment effects (LATEs) without parametric assumptions on the covariates. In practice, saturation leads to a large number of control dummies and arguably weak instruments, that induce a severe bias. We show that the estimator by Chao at al. (2023) (i) removes this bias, (ii) is asymptotically normal, yet (iii) the associated confidence intervals are invalid under treatment effect heterogeneity. We propose a new robust variance estimator with easy to compute confidence intervals. We find LATEs statistically significantly different from zero in applications in criminology, finance, health, and education.

arXiv: https://arxiv.org/abs/2402.12607

Key research areas:
His research focuses on developing econometric theory to improve macroeconomic forecasting and make effective use of high-dimensional data.

• Read more about the speaker
• Visit our website to read more about the Department of Data Science and Analytics