Fall Seminar - Florian Gunsilius

Abstract: Optimal transport has become an increasingly prominent framework in many facets of causal inference. Its applications range from identifying structural causal models and estimating effects through matching on observables to constructing robust causal estimators, analyzing classical reduced-form effects, and addressing questions of external validity. This talk focuses on the most basic application of optimal transport in causal inference, namely, to couple counterfactual distributions. A main theme here is the distinction between imposing restrictions through a structural model with latent heterogeneity and working directly with the set of admissible couplings of counterfactual marginals. This perspective allows one to uncover a common thread behind many sharp identification bounds and clarifies the connection to closely related ideas in distributionally robust formulations of external validity and in instrumental-variables models. I conclude by discussing how entropy-regularized transport connects these identification problems to statistical physics and large-deviation theory. The talk draws on material from my ongoing, unpublished book manuscript on optimal transport in causal inference and joint work with Bruno Nunes Costa (University of Michigan) Bio: Florian Gunsilius is an Associate Professor of Economics at Emory University. His research spans statistical optimal transport theory and causal inference, both as distinct areas and at their intersection. His research has appeared in Biometrika, Econometrica, Econometric Theory, the Journal of the American Statistical Association, the Journal of Econometrics, and the Journal of Machine Learning Research. Before joining Emory, he was an Assistant Professor at the University of Michigan and held visiting positions at Princeton and MIT. He received his Ph.D. in Economics from Brown University. Refreshments will be offered prior to the seminar (at 3:30) in the 129 suite.