LLM-Generated Survey Data Tested for Population Synthesis

A research investigation assesses the feasibility of utilizing health survey data produced by zero-shot large language models (LLMs) as a substitute for actual survey data in the context of geographically explicit population synthesis. Employing GPT-4.1 and Gemini-2.5-Pro, scientists generated synthetic datasets for Colorado and Mississippi, referencing the 2023 Behavioral Risk Factor Surveillance System (BRFSS). These datasets were incorporated into an iterative proportional fitting (IPF) framework to develop synthetic populations at the census tract level. Findings revealed that both LLMs effectively captured significant state-level differences, demonstrating that zero-shot generation can yield geographically distinct data. Nonetheless, the study underscores the varying performance and the potential as well as the limitations of LLMs in generating synthetic demographic data.