Abstract
dc:description.abstractThis dissertation consists of three chapters that explore the intersection of causal inference and distributional analysis using observational data. The first two chapters develop and apply econometric tools to measure heterogeneous treatment effects of policy interventions and the third chapter proposes a unified approach to distribution regression using machine learning methods. As policy interventions such as welfare or regulatory policies can have differential impacts across the outcome distribution, quantifying heterogeneous treatment effects are of central importance in many economic applications. Chapter 1 develops a flexible yet computationally straightforward approach to quantifying distributional causal effects in a quasi-experimental setting. This Chapter extends the identification argument of Athey and Imbens (2006) and proposes estimation and inference procedures based on distribution regression. This approach significantly widens the applicability of changes-in-changes by enabling scalable estimation controlling for observed confounders, while allowing for the effect of these confounders to vary across the outcome distribution. I derive the asymptotic distribution of the estimator as well as its functionals such as quantile and distributional treatment effects, and provide a bootstrap validity result for these quantities. Lastly, I apply the method to a health economics application analyzing the impact of the Earned Income Tax Credit on the entire birth weight distribution, finding that the policy had positive effects for at-risk infants, but had more muted effects for healthier babies. Chapter 2 extends the method developed in Chapter 1 to censored outcome variables. Distribution functions that are otherwise strictly increasing on the support but have a jump at the censoring threshold, arise frequently in many empirical settings in labor, education and health. Such discontinuities in the observed distribution functions complicate the construction of counterfactual distributions, which are essential for identifying and estimating treatment effects. Because the changes-in-changes can yield partially identified counterfactual distributions in this case, I propose sharp bounds along with the estimation procedure for the counterfactual distribution of interest. I also consider an extension of how to conduct inference in practice and provide Monte Carlo simulation evidence to demonstrate the numerical performance of the estimator. Chapter 3, co-authored with Victor Chernozhukov, Ivan Fernandez-Val and Victor Quintas-Martinez, proposes a spectrum of modeling choices for distribution regression that leverages machine learning tools to discover smoothness patterns in the conditional outcome distribution function. When economic theory or institutional knowledge can inform the smoothness or structure of the conditional distribution function, parametric modeling of the coefficient function can provide efficiency gains. Semiparametric modeling is done through Lasso regularization of the degree of smoothness between threshold indexes of the distribution. Theoretical results on the choice of optimal penalty levels and convergence rates allowing for cluster dependence are provided. Furthermore, we demonstrate a connection between distribution regression and feedforward neural networks and explore ways to uncover complex nonlinearities through multi-layered neural networks as an extension. Extensive simulation study is conducted to evaluate and compare various specifications of neural networks as well as regression-based methods. Lastly, the method is applied to the Current Population Survey 2015 dataset to investigate the gender wage gap in the US for the entire wage distribution.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Economics
- Grantor
- Boston University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Hong, Matthew M.
- Advisor dc:contributor.advisor
-
- Fernández-Val, Iván
Subjects
dc:subject × 8Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial 4.0 International
- Licence dc:rights.uri
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/2144/53361