Special Colloquium
Speaker: Promit Ghosal, University of Chicago
Host: Parthanil Roy
Title: Riesz-Calibrated Optimal Transport for Generalizing Treatment Effects
Time, day and date: 5:30:00 PM – 6:30:00 PM, Wednesday, September 02
Venue: Ramanujan Hall
Abstract: Randomized trials identify causal treatment effects within the trial population, but transporting these effects to a different target population requires accounting for differences in covariate distributions. We develop a new approach based on entropically regularized, semi-unbalanced optimal transport. A key observation is that standard regularized optimal transport can produce a non-vanishing bias in the induced importance weights, even as the sample size grows. We remove this bias by introducing calibration constraints into the transport problem and show that, with a growing calibration sieve, the resulting weights consistently estimate the target-to-trial density ratio, or equivalently the Riesz representer of the target expectation functional, while the transport regularization parameters remain fixed. Combined with outcome regression, this yields a doubly robust estimator that attains the semiparametric efficiency bound under suitable rate conditions, with inference obtained directly from its influence function. I will discuss the underlying statistical theory, the role of optimal transport as a regularized weighting mechanism, and numerical results illustrating robustness under weak overlap and model misspecification.