Bayesian average or truncation at boundaries? The mechanisms underlying categorical bias in spatial memory

Spatial memory is often biased by various factors, such as the region a target belongs to, which can be defined based on physical, perceptual, or implicit boundaries. In the typical dot-localization task first introduced by Huttenlocher, Hedges, and Duncan (Psychological Review 98: 352-376, 1991), individuals normally divide the task space into four quadrants delineated at the Cartesian axes (forming "default categories") and show systematic bias in target localization toward the center of the category. At least two mechanisms have been proposed to account for these categorical biases, namely (a) weighted-average of a metric representation and the category prototype representation and (b) truncation of an un-biased metric representation at the category boundary. Both models can account for these findings and cannot be differentiated by existing research methods. Using a new distribution analysis, the current study sought to differentiate between these two models. Participants viewed a dot inside a circle and recalled its location after a delay either with the same blank circle (i.e., the standard dot-in-circle paradigm) or when an alternative V-shaped category boundary was visually presented at retrieval. The data from three experiments showed symmetrical distribution of the errors that shifted toward the category center when people primarily used the default category, supporting the weighted-average model. In contrast, when people primarily used the alternative category, the errors showed a highly skewed distribution, more consistent with the truncation model. Overall, these results provided the first experimental evidence for both mechanisms separately.