Abstract

Quantifying the effects of textual interventions in social systems, such asreducing anger in social media posts to see its impact on engagement, ischallenging. Real-world interventions are often infeasible, necessitatingreliance on observational data. Traditional causal inference methods, typicallydesigned for binary or discrete treatments, are inadequate for handling thecomplex, high-dimensional textual data. This paper addresses these challengesby proposing CausalDANN, a novel approach to estimate causal effects using texttransformations facilitated by large language models (LLMs). Unlike existingmethods, our approach accommodates arbitrary textual interventions andleverages text-level classifiers with domain adaptation ability to producerobust effect estimates against domain shifts, even when only the control groupis observed. This flexibility in handling various text interventions is a keyadvancement in causal estimation for textual data, offering opportunities tobetter understand human behaviors and develop effective interventions withinsocial systems.