Large language model (LLM) assisted causal knowledge graph generation framework for survey data

The U.S. Census Bureau is the largest statistical organisation in the nation; it collects data on almost every aspect of the economy, demography, and geography to provide a snapshot of the country's population size and growth along with the characteristics of the economy. All this information is collected through numerous surveys such as the American Community Survey (ACS), the Annual Integrated Economic Survey (AEIS), and the Decennial Census of Population and Housing (decennial). These datasets provide a robust foundation for investigating complex socioeconomic phenomena. At the same time, extracting meaningful causal insights from them remains a major challenge due to their high dimensionality, observational nature, and structural complexity. Here we present a workflow for generating Causal Knowledge Graphs (CKGs) from large scale survey data by utilising Large Language Models (LLMs). We utilise LLMs to augment traditional causal discovery methods through a robust pipeline that integrates natural language understanding with statistical validation. We apply a combination of prompt tuning methods to generate preliminary causal hypotheses based on survey metadata, variable descriptions, and domain context. By integrating LLM derived domain knowledge with structural equation modelling and counterfactual analysis, we produce robust, interpretable causal graphs that bridge the gap between unstructured expert knowledge and structured causal inference. We demonstrate the effectiveness of our method on subsets of the ACS, highlighting its potential to uncover latent causal structures and support downstream tasks such as policy analysis, simulation, and decision support.