ESRA 2025 Program
Get the ESRA 2025 App on iOS or Android:
All time references are in CEST
Papers
Using Large Language Models to Predict Subjective Life Expectancy in the Context of Cross-Cultural Surveys: Insights into Cognitive Biases and Response Quality
Mr Mao Li (University of Michigan)
Miss Xinyi Chen (University of Michigan) - Presenting Author
Mr Zeyu Lou (University of Michigan)
Mr Kaidar Nurumov (University of Michigan)
Miss Stephanie Morales (University of Michigan)
Professor Sunghee Lee (University of Michigan)
The subjective life expectancy (SLE) question, which asks respondents to rate their probability of living beyond certain ages on a 0 to 100 scale, provides valuable data for predicting mortality-related behaviors. However, interpreting why respondents provide specific answers on SLE remains challenging. Probing questions, such as "Why do you choose ABC response?", are commonly used to capture respondents' reasoning processes. To deepen our understanding of these responses, this study leverages Large Language Models (LLMs) to explore two key objectives: first, to evaluate answer quality by categorizing probing responses as supportive, contradictory, or unrelated to the SLE rating, aiming to uncover potential cognitive biases that may lead respondents to over- or underestimate their lifespan; and second, to predict SLE ratings based solely on probing responses and assess the accuracy of these predictions.
This study contributes to the broader discourse on the utility of LLM-generated responses (synthetic respondents) in social science research. Specifically, we evaluate the potential of LLMs to approximate human reasoning by comparing their performance in predicting SLE ratings with actual human responses from a web panel survey (N=1,793) and the Survey of Consumer Attitude (N=506) across three languages (English, Spanish, German). The model’s predictions differed from respondents’ ratings by an average of 15%, suggesting reasonable alignment. However, misalignment between LLM-synthetic responses and SLE ratings illuminated heuristic or emotional influences on lifespan estimation, offering insights into the limits of synthetic responses in replicating complex human cognition. Further, socio-demographic patterns within these misaligned cases provide a window into how different groups approach probing questions, illustrating the potential and limitations of synthetic respondents in reflecting human diversity. This research underscores the importance of methodological rigor in designing and interpreting LLM-human comparative studies and highlights.
