High-dimensional Causal Inference Using Probabilistic Models

• In Event: Causal Inference

Abstract

Over the last two decades, the advancement of statistical and machine learning methods led to new probabilistic modeling techniques for high-dimensional and unstructured data including texts, images, audio, and video. And yet, the application of causal inference methods has been largely limited to structured data that often are of low dimension. In this paper, we propose a general methodological framework that integrates probabilistic models into high-dimensional causal inference. To demonstrate the effectiveness of the proposed methodology, we apply it to three causal inference problems in social sciences, in which textual data play the roles of treatment, outcome, and/or confounder. We show that our approach is able to relax the essential but unrealistic identification assumption required by the existing methods. Finally, we briefly discuss potential applications of the proposed methodology to other types of high-dimensional causal inference problems.

Authors

• Kosuke Imai, Harvard University

• Connor T Jerzak, University of Texas at Austin