Authors: Emily Wall, Arup Arcalgud, KUHU GUPTA, Andrew Jo
Abstract: Recently, Wall et al. proposed a set of computational metrics for quantifying cognitive bias based on user interaction sequences. The metrics rely on a Markov model to predict a user's next interaction based on the current interaction. The metrics characterize how a user's actual interactive behavior deviates from a theoretical baseline, where "unbiased behavior" was previously defined to be equal probabilities of all interactions. In this paper, we analyze the assumptions made of these metrics. We conduct a study in which participants, subject to anchoring bias, interact with a scatterplot to complete a categorization task. Our results indicate that, rather than equal probabilities of all interactions, unbiased behavior across both bias conditions can be better approximated by proximity of data points.