I am an applied econometrician at the U.S. Congressional Budget Office. I'm interested in using model selection, machine learning, and structural modeling to improve policy evaluation, particularly in the area of public health.
Email: eric.schulman@cbo.gov
Profiles: google scholar, linkedIn, github, quantecon
"Shapes as Product Differentiation: Neural Network Embedding in the Analysis of Markets for Fonts" with Sukjin Han, Kristen Grauman, and Santhosh Ramakrishnan
Revise & Resubmit RAND Journal of Economics
Many differentiated products have key attributes that are high-dimensional (e.g., design, text). Quantifying these attributes is important for economic analyses. This paper considers one of the simplest design products, fonts, and quantifies their shapes by constructing embeddings using a modern convolutional neural network. The embedding maps a font's shape onto a low-dimensional vector. Importantly, we verify the resulting embedding is economically meaningful by showing that the mutual information is large between the embedding and descriptions assigned to each font by font designers and consumers. This paper then conducts two economic analyses of the font market. We first illustrate the usefulness of the embeddings by a simple trend analysis of font style. We then study the causal effect of a merger on the merging firm's creative product differentiation decisions by using the embeddings in a synthetic control method. We find that the merger causes the merging firm temporarily to increase the visual variety of font design.
github 1 (neural net), github 2 (synthetic control)
Included in the MIT graduate machine learning course.
"Uniformly Valid Model Selection with the Bootstrap"
Under Review
This paper explores how bootstrapping can improve the Vuong test. I suggest a bootstrap test that is easy to implement and similar to bootstrapping the original Vuong test. I establish that the suggested bootstrap has uniformly valid asymptotic size control. For non-overlapping models, I show the test achieves better power than Shi (2015) and an asymptotic refinement. In Monte Carlo simulations, the suggested test controls size equally well and achieves higher power than other Vuong tests. Finally, I provide an empirical example involving benchmarks and enrollment in Medicare Advantage. The new test selects a model at with lower p-values.
github 1 (implementation and monte carlo), github 2 (empirical examples)
"Testable Implications of Timing and Beliefs in Healthcare Bargaining" with David S. Sibley
We explore the testable implications of assumptions about beliefs and timing in a Nash- in-Nash bargaining model of hospital insurer negotiations with one hospital and two insurers. We consider passive beliefs, active beliefs (similar to Ho and Lee (2019)), and sequential bargaining (similar to Horn and Wolinsky (1988)) when modeling a two-to-one merger. We show one assumption may suggest the merger will lower premiums, but another may indicate an increase. To resolve the discrepancy in an empirical setting, we propose a test for beliefs and timing and show it has good size and power in Monte Carlo simulations.
github (numerical examples)
"Tracking the Incidence of Medicare Advantage Payments" with Jared Lane Maeda
"Comparing the Effects of Current Pay and Defined Benefit Pensions on Employee Retention" with Justin Falk
"Assessing the effects of the Medicaid drug coverage expansion with Double Machine Learning" with Christopher P. Adams