Research

Constrained Learning for Causal Inference (job market paper)

Tiffany Cai*, Yuri Fonseca*, Kaiwen Hou, Hongseok Namkoong (* denotes co-first authorship)

Paper

Summary In challenging settings with limited overlap between treatment and control, causal estimators with desirable asymptotic properties require ad hoc adjustments in order to produce stable estimates. In contrast, simple plug-in estimators produce stable estimates but lack important asymptotic properties. We propose a new estimation framework based on constrained optimization, that combines the best of both worlds, and we demonstrate its superior empirical performance across several settings. Our framework is compatible with modern machine learning, and we include settings with text covariates.

Forthcoming; poster at CODE@MIT 2024 and ACIC 2024

Active Exploration via Autoregressive Generation of Missing Data

Tiffany Cai*, Hongseok Namkoong, Daniel Russo, Kelly W Zhang* (* denotes co-leading)

Paper

Summary We propose a scalable solution to the problem of decision-making under uncertainty in a meta-bandit setting by using a calibrated generative model to impute a sequence of missing (e.g. future) rewards. Our proposed method is a principled implementation of Thompson (a.k.a. posterior) sampling. We prove decision-making performance is controlled by the log loss of the generative model, and we demonstrate on a news recommendation setting with text covariates.

Submitted; poster at Neurips 2024 Workshop: Bayesian Decision-Making and Uncertainty and talk at 2024 Economics and AI+ML Meeting

Diagnosing Model Performance Under Distribution Shift

Tiffany Cai, Steve Yadlowsky, Hongseok Namkoong

Paper, GitHub, Slides

Summary When your model performs worse out of distribution, should you use a domain adaptation method, or do you need to collect more data? If the latter, from where should you collect more data? We propose a new diagnostic using causal inference methods to attribute changes in performance to X shifts and Y|X shifts. We demonstrate its utility in settings with tabular and image data.

Minor revision at Operations Research; presented at FORC 2023, INFORMS 2023

Tutorial: Modeling and Exploiting Data Heterogeneity under Distribution Shifts

Jiashuo Liu, Tiffany Cai, Peng Cui, Hongseok Namkoong

Tutorial

Presented at NeurIPS 2023