research
My research focuses on causal inference in political economy, AI evaluation, and LLM-based measurement. Below are my current projects. Please reach me at ml4967@columbia.edu if you have any questions or are interested in collaborating.
Political Economy
- Campaign Giving Under Employment Threat: Federal Employees and the DOGE Workforce Reduction
Does the threat of economic harm to government workers change their political giving? Using Federal Election Commission records and a difference-in-differences design that compares federal and non-federal donors in 2025 with the same weeks of 2017, I find that federal employees’ Democratic giving fell most sharply in the early weeks of 2025, before any separations occurred, while their Republican giving showed no comparable decline. The pattern suggests that employment threats can reduce publicly disclosed opposition giving.
- Buying American Abroad: Political Identity and Consumer Choice
Does America’s standing in the world shape what consumers abroad buy? Using household-level purchase panels, we study how shifts in the United States’ global reputation affect demand for American products.
- Claiming Corporate Success: Economic Exposure and Support for the Incumbent Party
Does a president’s public claim of credit for a company’s success change political giving in the places where that company employs people? I combine LLM-coded presidential credit claims with local employment and stock-compensation exposure to publicly listed firms and with campaign-contribution records.
AI Evaluation and Measurement
- Better Questions for the Better Trained: LLM Memory and Unequal Research Advice
AI assistants now remember what users tell them, including how much training they have. Varying only what two LLMs remember about a Ph.D. student’s mathematical preparation, this paper finds that blind LLM judges rate the research questions and plans suggested to more highly trained users as stronger, not merely harder, and argues that evaluations of personalized assistants should judge the strength of advice separately from its difficulty and feasibility.
- The Unequal Collapse of Disciplinary Barriers: Generative AI and the Reallocation of Scientific Talent
Does generative AI lower the barriers to doing research outside one’s own discipline, and for whom? We develop a formal model in which AI relaxes some capability constraints, such as codifiable domain knowledge, more than others, such as substantive judgment, and use it to predict who enters new fields and who produces high-quality research proposals. We are designing an experiment to test these predictions.
- Borrowed Battlegrounds: The Formation and Reorganization of Foreign-Language Event-Publics around U.S. Political Violence
How do foreign publics discuss political violence against American politicians? Studying 29 U.S. political-violence events from 2010 to 2025 in five foreign-language communities on X, with LLM-labeled samples and a 1.3-million-post collection around the attacks on Donald Trump and Charlie Kirk, I find that even assassinations draw only short-lived attention abroad: most participants are first-time entrants whose activity fades within days, and views of the targeted politician’s party do not change.
Electoral Behavior
- When Group Issues and Core Issues Collide: Evidence from Tailored Conjoint Experiments on Latino Voters
How do Latino voters respond when candidates take different positions on group-related issues and on broader ideological or policy issues? Using tailored conjoint experiments, we study how voters weigh group attachments against other political considerations.
- Democracy Out of Sync: Circadian Misalignment and the Cost of Voting
How does time-zone-induced circadian misalignment affect the cost of voting? We examine whether biological and temporal constraints shape electoral participation.