Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments
Researchers have developed a new methodology called GenAI-Powered Inference (GPI) that leverages large language models (LLMs) to improve causal inference with text data. This approach uses LLMs to generate treatments and their internal representations, which helps to isolate specific features like sentiment or topics from confounding factors. The GPI methodology aims to provide more accurate and efficient estimates by not requiring direct learning of causal representations from the data, and it can be extended to settings involving human perception or text reuse. AI
IMPACT This new methodology could enable more robust analysis of text-based data in fields requiring causal inference, such as social sciences and marketing.