Apr 2025,

Revise & Resubmit,
Journal of American Statistical Association

Most work on treatment choice and policy learning focuses on utilitarian welfare (i.e., average welfare), which can be sensitive to skewed heterogeneity. We propose a robust policy learning framework that enables the policymaker to act with prudence/negligence and to be influenced by vote shares.


Shapes as Product Differentiation︎︎︎


with Eric Schulman, Kristen Grauman, Santhosh Ramakrishnan


arxiv︎︎︎  slides︎︎︎

Nov 2022,
Revise & Resubmit,
RAND Journal of Economics

Many differentiated products have key attributes that are high-dimensional (e.g., design, text). We consider one of the simplest design products, fonts, and quantify their shapes by constructing neural network embeddings. Using the embeddings and data from the world's largest online market place for fonts, we study the causal effect of a merger on the merging firm's creative decisions of product differentiation.

*This project is featured in a typography magazine ︎︎︎ and included in the MIT graduate machine learning course ︎︎︎.

Jan 2025

EC’25

To understand the role of copyright policy in markets for products with visual attributes, we estimate a structural model of supply (e.g., product positioning) and demand (e.g., tastes for visual attributes) using image data. Visual similarity, calculated using neural network embeddings, serves as a crucial metric for the analysis.

Feb 2025


The classic control function approach requires invertibility or point-identified controls, which limits its applicability (e.g., discrete treatments, controls as intervals). We allow the control function to be set-valued and derive sharp bounds on structural parameters.

Mar 2025

Finding IVs is a heuristic and creative process, and justifying exclusion restrictions is largely rhetorical. We propose using large language models (LLMs) to systematically search for new IVs through narratives and counterfactual reasoning.
© Sukjin Han