Abstract
Perception is biased by stimulus history. History biases have been analyzed on different timescales: "Serial dependence" (SD) describes the effects of immediately preceding stimuli, whereas "central-tendency bias" (CTB) describes effects that depend on all stimuli encountered during an experiment. Studies often focus on either short- or long-term effects, but as these two possible sources of bias are correlated in stimulus statistics, one type of effect can disguise as the other if both are not taken into account in data analysis. We sought to disentangle these two effects. In a delayed-matching experiment, observers judged either the hue of a circular patch or the length of a line segment. A reference stimulus was presented for 500 ms, followed by an inter-stimulus interval (500 ms or 1500 ms) and a comparison stimulus for 500 ms, after which a response was given (comparison bluer/greener or shorter/longer than reference). To estimate the temporal properties of SD, the presentation schedule was fixed and did not pause for responses. We split the data based on whether the current reference stimulus was bluer or greener (or shorter/longer) than the running mean, the 1-back stimulus, the 2-back stimulus, and so forth, and fit psychometric functions to estimate the bias in each case. We also fit a generalized linear model with a probit link function to the data to estimate the independent contributions of these history features on judgments of the current stimulus. We found (1) significant attractive serial and central tendency effects, that (2) 1-back SD is stronger than CTB but (3) decreases quickly and disappears after about 4-5 trials, and that (4) the strength of SD depends more on the number of trials than on absolute time passed. Both short- and long-term effects bias visual perception and both need to be taken into account when analyzing psychophysical data.