Pretesting Family Support Policy Effects on Fertility Intentions through Survey-Grounded LLM Agent Simulation

Background and Purpose: Persistently low fertility rates have led countries such as China to introduce family-support policies intended to encourage childbearing. Yet before these policies are implemented at scale, it is often difficult to understand how people may respond. Traditional policy evaluation approaches are often time-consuming, costly, and difficult to scale, especially when new post-policy data must be collected. More efficient approaches may help researchers assess likely response patterns and refine policy design at an earlier stage. This study developed and tested a survey-grounded large language model framework for simulating changes in fertility intentions in response to a family-support policy scenario in China.

Methods: We analyzed data from a survey experiment with 813 respondents in China. Participants reported fertility intentions before and after exposure to a family-support policy scenario, along with related factors such as financial pressure, work-family conflict, family orientation, policy responsiveness, and family stage. We then used these survey data to build synthetic respondent profiles and assign them to large language model (LLM) agents that simulated changes in fertility intentions after policy exposure. Simulated results were compared with observed survey responses and with two simpler GPT-5 baseline approaches: direct prompting and persona-only prompting. We also examined a supplementary calibrated version using a small amount of labeled post-scenario data.

Results: The survey-grounded agentic framework performed better than the direct and persona-only GPT-5 baselines in reproducing observed response distributions. Average mismatch was 10.9% for baseline fertility intention and 8.4% for post-policy fertility intention, compared with 46.3% and 67.6% for the direct baseline and 42.7% and 41.2% for the persona-only baseline. The supplementary calibrated version best matched the observed magnitude of change.

Conclusions and Implications: Survey-grounded LLM agents show promise as a scalable and cost-efficient tool for pretesting family-support policies. For social work and family policy research, this approach may be useful when researchers need an early, lower-cost way to examine likely response patterns and sharpen future study design before launching larger, resource-intensive data collection.