LLMs Predict Missing Public Opinion in Historical Surveys

Researchers have developed a new method that uses large language models (LLMs) to predict missing responses in repeated cross-sectional surveys. This framework can estimate opinions from years where data is absent and even forecast views that weren’t asked about at all. By combining embeddings for questions, respondents, and survey years, they tested it on the General Social Surveys (GSS) spanning 1972 to 2021. The LLM models effectively uncovered hidden opinions and estimated public sentiments from other organizations for years when the GSS didn’t ask certain questions. This innovation helps recover overlooked trends and identify changes in public attitudes, like growing support for specific issues, despite the limitations of annual survey questions.