Marvin Maechler¹ (), Alan Stocker¹; ¹University of Pennsylvania

Decision-making often involves categorical choices that can bias subsequent judgments (e.g., confirmation bias). We investigated these choice-induced biases in a psychophysical task where participants first had to categorize the generative mean of a sequence of angular position stimuli (clockwise/counterclockwise relative to a reference) before ultimately estimating this generative mean by adjusting a cursor. In each trial we presented eight samples drawn from a normal distribution around different mean angular positions, displayed as small white dots (1dva diameter) on uniform gray background equidistant (5dva) from fixation. Importantly, a categorical decision was required after seeing either four samples (prospective condition) or all eight samples (retrospective condition). Unlike previous studies, detailed feedback was given in each trial to ensure that potential biases were not caused by miscalibrated inference strategies. To isolate choice-induced biases from effects of sampling noise, we presented identical stimulus sequences multiple times in both conditions. In line with previous results, we find robust estimation biases away from the reference position in both conditions. Interestingly, while estimates showed nearly identical distributions in both conditions, the correlation between categorical choice and subsequent estimate was strongly reduced in the prospective condition. This indicates that participants’ categorical choices had a limited influence on their perceptual estimates in the prospective condition. Rather, we hypothesize that the mere presence of the reference induced a category-level stimulus representation that caused participants to make an implicit categorical judgment before their final estimate even in the prospective condition. Biases persisted despite monetary incentives for accuracy and trial-by-trial feedback. Therefore, these biases may arise from an intentional process that optimizes for multiple objectives beyond pure estimation accuracy. A holistic Bayesian matching model that formulates such multi-level objectives across the representational hierarchy can explain the data and demonstrates the normative nature of the observed biased stimulus estimates.