Date of Award

Fall 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Statistics and Data Science

First Advisor

Sanford, Luke

Abstract

Causal inference provides statistical tools for generating empirical evidence about causal relationships. In conventional settings, randomized experiments can enable estimation of causal effects with simple methods and under weak assumptions, but many pressing causal questions concern complex systems where standard approaches break down. This dissertation presents advances and insights for such cases, first in the context of carbon markets and conservation programs, and then in settings where text serves as treatment. In the first chapter, we develop a generalized causal inference framework for analyzing forest carbon protocols. Recent work has questioned the credibility of forest carbon offsets as an environmental intervention and nature-based solution for mitigating climate change. Debates often focus on effect quantification methods, of which there are many, but rarely engage with underlying causal assumptions. This chapter reviews the wide array of methodologies used across protocols, presents a typology of the approaches, and clarifies the implicit causal assumptions each requires. In the second chapter, we consider challenges that interference poses for forest conservation policy evaluation. A pre-analysis plan for a randomized experiment in the Brazilian Amazon where spillover effects are expected serves as a guiding case study. We take a simulation-based approach to illustrate - in the presence of interference - power assessment strategies, the power-bias tradeoff of different estimation methods, and the consequences of using finite-population variance estimators for super-population estimands. In the third chapter, we consider the setting of causal inference using text data. We build upon previous experimental methods for estimating the impacts of text on human evaluation and propose an interpretable machine learning method for identifying influential text treatments from a corpus of unstructured text. We apply the method to two datasets. One allows validation of the model's ability to detect phrases known to cause the outcome, and the other demonstrates its ability to flexibly discover text treatments with varying textual structures. We provide an R package for general use in deploying the method.

Download

Share

COinS