LLMs as Emotion Analyzers for Causal Models: Partial Identification with Fuzzy Interval Data

Information systems (IS) researchers utilize machine learning algorithms to identify emotions in vast online content and investigate their impact on business outcomes. Large language models (LLMs) serve as automatic emotion analyzers, presenting new avenues for advancing the causal understanding of emotions in IS research. However, directly using LLM-generated emotional variables in econometric models can introduce bias due to LLMs’ imperfect emotion recognition, which introduces noise into causal models. To address this, we propose a novel algorithm that considers predictions from multiple LLMs to create fuzzy interval data, enabling partial identification of causal parameters. The algorithm satisfies three key design requirements: it operates in an unsupervised manner, eliminating the need for additional labeled data for causal estimation; it is flexible by integrating predictions from different LLMs, harnessing collective AI insights; and it ensures theoretically consistent causal estimators. This research has significant implications for causal inference using LLMs, the study of emotions in causal contexts, and prescriptive analytics for handling fuzzy interval data.