Computational analysis of 100 K choice dilemmas: Decision attributes, trade-off structures, and model-based prediction

Abstract

We present a dataset of over 100 K textual descriptions of real-life choice dilemmas, obtained from social media posts and large-scale survey data. Using large language models (LLMs), we extract hundreds of choice attributes at play in these dilemmas and map them onto a common representational space. This representation allows us to quantify the broader themes and specific trade-offs inherent in life choices and analyze how they vary across different contexts. We also present our dilemmas to human participants and find that our LLM pipeline, when combined with established decision models, accurately predicts people’s choices, outperforming models based on unstructured textual content, demographics, and personality. In this way, our research provides insights into the attributes, outcomes, and goals that underpin life choices, and shows how large-scale LLM-based structure extraction can be used in combination with existing scientific theory to study complex real-world human behavior.