Extracting Life-course Trajectories from Longitudinal Survey Data Using Large Language Models

• In Event: Computational Methods and Models for Social Analysis

Abstract

Social events and their sequencing are fundamental to how humans understand the life-course. This study proposes a pipeline to extract individuals’ life-course trajectories from large-scale survey data. I reconceptualize survey participation as an act of self-annotation, which generates question text–response label pairs representing one’s latent life. Using this natural language data and large language models (LLMs), I demonstrate that LLMs can identify multiple trajectories of chained social events for each survey respondent. Data were drawn from Waves 1–5 of the Future of Families and Child Wellbeing Study (FFCWS), which oversampled low-income “fragile families” in the U.S. Gemini-3.1-Flash and Pro were used to process an average of 3,229 variables per family. These models successfully captured critical social events and trajectories that are unique to this sample, such as the receipt of public assistance (e.g., TANF), multipartner fertility, and residential instability. The extracted trajectories demonstrated that these families experienced both hardships, such as employment precarity, and protective mechanisms like job training participation. I conclude that sociologists’ focus on events and trajectories can deepen our understanding of people’s life hardship, resilience, and agency, and that LLMs can reliably extract this information from longitudinal survey data.

Author

• Youngjin YJ Chae, Rutgers University-New Brunswick