Distinguishing response from stimulus driven history biases

Abstract

Perception is shaped by past experience, both cumulative and contextual. Serial dependence reflects a contextual attractive bias to perceive or report the world as more stable than it truly is. As serial dependence has often been examined in continuous report or change detection tasks, it unclear whether attraction is towards the identity of the previous stimulus feature, or rather to the response made to indicate the perceived stimulus value on the previous trial. The physical and reported identities can be highly correlated depending on properties of the stimulus and task design. However, they are distinct values and dissociating them is important because it can reveal information about the role of sensory and non-sensory contributions to attractive biases. These alternative possibilities can be challenging to disentangle because 1) stimulus values and responses are typically strongly correlated and 2) measuring response biases using standard techniques can be confounded by context-independent biases such as cardinal bias for orientation (i.e., higher precision, but repelled, responses from vertical and horizontal orientations). Here we explore the issues and confounds related to measuring response biases using simulations. Under a range of conditions, we find that response-induced biases can be reliably distinguished from stimulus-induced biases and from confounds introduced by context-independent biases. We then applied these approaches to a delayed report dataset (N=18) and found evidence for response over a stimulus driven history bias. This work demonstrates that stimulus and response driven history biases can be reliably dissociated and provides code to implement these analysis procedures.