Twin-2K-500: A dataset for building digital twins of over 2,000 people based on their answers to over 500 questions

LLM-based digital twin simulation, where large language models are used to emulate individual human behavior, holds great promise for research in AI, social science, and digital experimentation. However, progress in this area has been hindered by the scarcity of real, individual-level datasets that are both large and publicly available. This lack of high-quality ground truth limits both the development and validation of digital twin methodologies. To address this gap, we introduce a large-scale, public dataset designed to capture a rich and holistic view of individual human behavior. We survey a representative sample of $N = 2,058$ participants (average 2.42 hours per person) in the US across four waves with 500 questions in total, covering a comprehensive battery of demographic, psychological, economic, personality, and cognitive measures, as well as replications of behavioral economics experiments and a pricing survey. The final wave repeats tasks from earlier waves to establish a test-retest accuracy baseline. Initial analyses suggest the data are of high quality and show promise for constructing digital twins that predict human behavior well at the individual and aggregate levels. By making the full dataset publicly available, we aim to establish a valuable testbed for the development and benchmarking of LLM-based persona simulations. Beyond LLM applications, due to its unique breadth and scale the dataset also enables broad social science research, including studies of cross-construct correlations and heterogeneous treatment effects.

The Challenge of Using LLMs to Simulate Human Behavior: A Causal Inference Perspective

Large Language Models (LLMs) have demonstrated impressive potential to simulate human behavior. Using a causal inference framework, we empirically and theoretically analyze the challenges of conducting LLM-simulated experiments, and explore potential solutions. In the context of demand estimation, we show that variations in the treatment included in the prompt (e.g., price of focal product) can cause variations in unspecified confounding factors (e.g., price of competitors, historical prices, outside temperature), introducing endogeneity and yielding implausibly flat demand curves. We propose a theoretical framework suggesting this endogeneity issue generalizes to other contexts and won't be fully resolved by merely improving the training data. Unlike real experiments where researchers assign pre-existing units across conditions, LLMs simulate units based on the entire prompt, which includes the description of the treatment. Therefore, due to associations in the training data, the characteristics of individuals and environments simulated by the LLM can be affected by the treatment assignment. We explore two potential solutions. The first specifies all contextual variables that affect both treatment and outcome, which we demonstrate to be challenging for a general-purpose LLM. The second explicitly specifies the source of treatment variation in the prompt given to the LLM (e.g., by informing the LLM that the store is running an experiment). While this approach only allows the estimation of a conditional average treatment effect that depends on the specific experimental design, it provides valuable directional results for exploratory analysis.