Adam Malitek¹ (), Minsuk Chang², Cindy Xiong Bearfield², Keisuke Fukuda^1,3; ¹University of Toronto Mississauga, ²Georgia Institute of Technology, ³University of Toronto

To behave intelligently in today’s data-rich society, we need to accurately perceive, interpret, and remember the data presented to us. To this end, Xiong et al. (2022) showed that the perceived correlation of a scatterplot was biased toward participants’ beliefs about the relationship between predictor and dependent variables. However, it is unclear whether the bias persists beyond the time of perception, and if so, whether the impact reflects a bias in the memory representation of the scatterplot or the interpretation of an unbiased memory of the scatterplot. To test this, participants first performed a correlation perception task. In each trial, participants saw a general statement depicting a relationship between two variables (e.g., A worker with a longer commute tends to be more stressed, X: commute time; Y: stress level) and reported their belief and predicted correlation between them. Participants then saw a scatterplot of the two variables and reported the perceived correlation. After the perception task, participants performed a memory task in which they first saw a pair of variables (e.g., X: commute time; Y: stress level) and indicated how well they remembered the corresponding scatterplot. Subsequently, they drew the remembered scatterplot from their memory and reported the correlation in a counterbalanced order. Our results (N = 70) replicated the belief-driven bias in perceived correlation. Furthermore, the belief-driven bias persisted in the scatterplots drawn from memory and the remembered correlations. These findings highlight the pervasiveness of prior beliefs in correlation estimation and suggest that belief-driven bias influences memory representation. This emphasizes the need for analysts to account for belief-driven biases in visual data processing and retention of data patterns in memory.

Acknowledgements: This research was supported by the Natural Sciences and Engineering Research Council (5009170).