Richard Xu

Traditional analyses of heterogeneous treatment effects often rely on pre-specified, structured numerical baseline indicators to test for significant differences in treatment effects across groups. This paper introduces a novel framework that utilizes word embeddings generated from ChatGPT’s textual outputs to identify factors driving significant heterogeneous treatment effects. Applying this approach within a regression discontinuity in time design, we examine the key determinants influencing the impact of Pfizer’s announcement of 90% efficacy for its first COVID-19 vaccine on weekly consumption patterns across U.S. cities during the pandemic’s first year. The announcement boosted public confidence in future economic conditions, thereby influencing current spending behavior. Our findings indicate that cities with high population density, elevated income levels, and weaker healthcare infrastructure experienced significantly stronger positive impacts. This study demonstrates the potential of combining innovative AI tools, such as ChatGPT, with textual information to advance causal inference methods in applied econometrics.

This paper investigates how diverse stakeholders perceived the IMF’s 2018 Exceptional Access Program (EAP) in Argentina, with the aim of understanding how such perceptions influence policy outcomes and program effectiveness. Leveraging recent advances in natural language processing, the study applies large language models (LLMs) to perform multilingual sentiment analysis across a range of sources, capturing views from domestic publics, national authorities, and international investors. To uncover the drivers of these sentiments, the paper employs topic modeling to identify key themes shaping narratives around the IMF and the broader economic context. By addressing three core questions—how stakeholders perceive the economy and the IMF, what incentives shape these perceptions, and how these perceptions influence policy outcomes—the study offers a novel framework for evaluating IMF engagement in crisis contexts. Through rigorous data filtering and methodological controls for bias and multiple comparisons, the paper provides one of the first large-scale, LLM-based analyses of multi-stakeholder sentiment toward an IMF program. Findings will contribute to a deeper understanding of the political economy dynamics that underpin IMF-supported reforms, with broader implications for global financial governance and stakeholder-informed policy design.