Characterizing serial dependence as an attraction to prior response

Abstract

Serial dependence refers to a common misperception that can occur between subsequently observed stimuli. Observers misreport the current stimulus as being more similar to the previous stimulus than it objectively is. It has been proposed that this bias may reflect an attraction of the current percept to prior percept (Fischer & Whitney, 2014). Alternatively, serial dependence has also been proposed to be the result of an assimilative effect between observer decisions (Fritsche, Mostert, & de Lange, 2017; Pascucci, Mancuso, Santandrea, Libera, Plomp, & Chelazzi, 2019). Lying within this debate is the issue of how we quantify serial dependence. Should this be as a bias induced by prior stimuli or by prior responses? We investigated this by manipulating the orientation of the current stimuli such that they fell between previous stimulus and previous response. We observed an attraction to previous response and a concomitant repulsion from previous stimulus. This suggests that the attractive effect of serial dependence in orientation judgments is best quantified in relation to prior response.

Figure 1.

Typical trial sequence for the adjustment response portion of the noise experiment. The top image shows the screen as seen by a participant; the stimuli (shown below) appear in the dashed circle, and the fixation circle was present in all screens. The orientation stimulus appeared on-screen for 500 ms. Full screen noise was then displayed for 1000 ms to remove any visual aftereffects. The adjustment stimulus was then displayed in the same position previously occupied by the orientation stimulus. A blank screen was then displayed for 2000 ms. The next trial sequence appeared at a point 45° counterclockwise of the previous trial sequence (dashed circle).

Figure 2.

Diagrammatic representation of oppositional trial generation. When an error made in response orientation (red arrow) to a stimulus orientation (black arrow) on trial t was of sufficient magnitude (between 10° and 30°), the stimulus orientation (black arrow) on trial t + 1 was intermediate between the prior stimulus orientation and prior response orientation.

Figure 3.

The commonly used DoG fit to non-oppositional data produced in this study. The left graph shows stimulus-contingent data, and the right graph shows response-contingent data. Note this is only intended as a visual comparison to previous studies. This approach was not appropriate for further analysis.

Figure 4.

Response- and stimulus-contingent biases for oppositional and non-oppositional trials. Error bars represent 95% confidence intervals. Response-contingent analysis was corrected by residualization.

Figure 5.

Biases in shuffled control data. The left graph shows data without oblique bias correction, and the right graph shows the same data with correction applied. S NO, stimulus-contingent, non-oppositional trials; R NO, response-contingent, non-oppositional trials; S O, stimulus-contingent, oppositional trials; R O, response-contingent, oppositional trials.

Figure 6.

Stimulus- and response-contingent oppositional trials, sorted on the basis of whether the two-back stimulus is on the same side of the current stimulus as the one-back stimulus or not. Response bars represent 95% confidence intervals.