Multivariate Transfer Entropy Quantifies Information Transfer Between Dyadic Partners During Real-Time Perceptual Decision-Making

Abstract

1Many social behaviours unfold continuously in time, yet most studies of social influence rely on discrete, trial-based paradigms. To quantify how humans actually integrate sensory and social signals during ongoing perceptual decision-making we here apply an information-theoretic analysis to data obtained with the Continuous Perceptual Report (CPR) paradigm. In the CPR task, pairs of participants tracked a dynamic random-dot motion stimulus, continuously reporting perceived direction and their confidence. Crucially, each participant had access to their partners moment-by-moment direction and confidence reports embedded within the stimulus display. We quantified information flow between the stimulus and behavioural responses using multivariate transfer entropy, allowing us to detect whether participants acquired additional information from their partner once stimulus information had been received and integrated it into their behavioural response. We show that real-time social interactions predominantly occur within matching behavioural dimensions: partners perceptual choices influence participants own choices, and partners confidence influences participants confidence. Using both, more stimulus and social information contributed to improved task performance, and participants selectively acquired more information from better-performing partners. Crucially, information integration was flexibly modulated by stimulus reliability: as sensory evidence became noisier, participants relied more on social information, but only when interacting with a highly reliable (e.g. computer-controlled) partner. Latency analyses further revealed that participants responded faster to changes in their partners perceptual choices of direction than to changes in stimulus direction. They also responded more slowly to changes in their partners confidence than to changes in their partners choice of direction. Together, these findings demonstrate that humans dynamically and adaptively integrate social information in a time-continuous manner, and they establish multivariate information-theoretic analyses as a powerful framework for studying real-time social cognition.