Controlling for Text in Causal Inference with Double Machine Learning

June 6, 2022 at ICWSM 2022

Colab Notebook, Proposal, Slides

Abstract: Text plays an increasingly important role in the study of causal relationships. In this tutorial, we consider the specific case of using text as a control to eliminate bias from confounders operating through the text. We formalize the problem of controlling for text using causal graphs and the potential outcomes framework, describe a principled estimation and inference procedure to realize this goal using double/debiased machine learning, and compare this procedure (hands-on) against several alternatives such as controlling for low-dimensional representations of the text obtained via topic modeling, principal component analysis, or other techniques. We conclude with a case study on using text as a control to quantify the causal impact of status on persuasion online.

Organizers

  • Emaad Manzoor (emanzoor@wisc.edu), U. Wisconsin Madison