Serial bias and response time: Prior stimulus not only biases but also modulates the speed of decision for a new stimulus

Abstract

A decade of research has demonstrated that the reported perception of a new stimulus can be biased by task-irrelevant prior stimuli. However, existing studies have primarily focused on explaining the direction and magnitude of this bias effect, often neglecting other relevant aspects of perceptual behavior that may also be influenced by prior stimuli. In this study, we examined how decision speed for a new stimulus might be influenced by prior stimuli in motion-direction estimation tasks. We found that direction reports exhibited a repulsive serial bias along with a systematic response time (RT) effect, where reports were faster when the prior motion direction was more dissimilar to the current motion direction. Follow-up experiments replicated this RT effect and showed that it occurred only when repulsive serial bias was evident. Subsequent analyses revealed that the RT effect was positively correlated with repulsive serial bias, indicating that both effects are driven by common underlying mechanisms. Together, these results demonstrate that prior stimuli not only bias but also modulate response speed to new stimuli, suggesting that existing theories should incorporate decisional mechanisms that influence response speed to fully account for the serial bias phenomenon.

Introduction

Our perceptual experience is shaped not only by the perception of objects currently in view but also by the memory of recent perceptual experiences. For example, finding well-ripened red tomatoes can be influenced both by the color of the tomatoes that one is currently looking at and by the memory of the tomatoes that the person saw a moment ago. In laboratory experiments, research has consistently found that the reported perception of a new stimulus is systematically biased by the prior stimulus even when the prior stimulus is completely irrelevant to the task at hand (Cicchini, Mikellidou, & Burr, 2024; Fischer & Whitney, 2014; Kiyonaga, Scimeca, Bliss, & Whitney, 2017; Manassi, Murai, & Whitney, 2023; Pascucci et al., 2023). Studies have proposed that serial bias is a general phenomenon that reflects a fundamental goal of visual information processing: The visual system should stabilize perceptual experiences over time in the face of perceptual uncertainties (Cicchini, Mikellidou, & Burr, 2018; Fischer & Whitney, 2014; Manassi & Whitney, 2024). The visual system may achieve this goal by relating new visual inputs with the information obtained in the recent past. 

One key aspect of this serial bias is that the direction of the bias can be either attractive (i.e., serial dependence) (Ceylan, Herzog, & Pascucci, 2021; Cicchini et al., 2018; Fornaciai & Park, 2018) or repulsive (Bae & Luck, 2017; Bae & Luck, 2019; Bae & Luck, 2020; Bansal, et al., 2020; Shan & Postle, 2022). The exact cause of the opposing directions of serial bias is still under active investigation (Pascucci et al., 2023). Studies have found that the direction and the magnitude of the bias depend on multiple factors, including the delay between the stimulus encoding and the report (Bliss, Sun, & D'Esposito, 2017), the task relevance of the prior stimulus (Ceylan & Pascucci, 2023), specific stimulus values being reported (Bae, 2024), and the way the stimulus is maintained in working memory (WM) (Chen & Bae, 2024a). It has also been shown that both directions of serial bias coexist within the same task: The reports are biased toward the prior response but away from the prior stimulus (Sadil, Cowell, & Huber, 2024; Sheehan & Serences, 2023). 

To account for the complex nature of the serial bias, studies have proposed a framework that includes two computational principles that may underlie the attractive and repulsive components of the serial bias effect (Fritsche, Mostert, & de Lange, 2017; Fritsche, Spaak, & de Lange, 2020; Moon & Kwon, 2022; Pascucci et al., 2019). According to the framework, the visual system recruits an adaptation mechanism (Gibson & Radner, 1937) during the encoding of the new stimulus (i.e., efficient encoding), producing a repulsive bias away from the prior stimulus, and it recruits a mechanism for decision making during the post-perceptual stage of the processing (i.e., Bayesian decoding), producing an attractive bias toward the prior stimulus. The framework assumes that the bias in the reports reflects the combined results from the two processes. For example, a stronger weight on the adaptation mechanism should lead to an overall repulsive serial bias. Indeed, the pattern of repulsive serial bias resembles the pattern of perceptual after-effects driven by adaptation (Clifford, Wenderoth, & Spehar, 2000). 

Although most previous studies have focused on explaining the biases in the reports, research has provided evidence that the prior stimulus impacts the speed of reports, as well. In an orientation delayed estimation task, for example, Cicchini et al. (2018) found that the response time was faster when the previous stimulus was similar to the new stimulus and that this response time (RT) effect was associated with attractive serial bias. Another study confirmed that the RT effect occurs for the repulsive serial bias and that the direction of the RT effect can be reversed (e.g., faster response for dissimilar trials) depending on task contexts (Chen & Bae, 2024a). Given that most of the previous research on serial bias has focused on explaining the bias effect, these findings provide important new insights into the mechanisms underlying the serial bias effect, especially on the processing stages where the bias effect might emerge. 

Motivated by these studies, the present study sought to find further evidence that the RT effect is driven by the mechanisms underlying serial bias. To that end, we noticed that the RT effect was observed in tasks where the perceived stimulus had to be maintained in WM (Chen & Bae, 2024a; Chen & Bae, 2024b; Cicchini et al., 2018). To confirm that the RT effect is not merely driven by the involvement of the memory demand, Experiment 1 attempted to replicate the RT effect and the serial bias using a motion-direction estimation task where explicit demand for the stimulus maintenance was not imposed (Experiment 1, Figure 1a). In the experiment, participants reported the perceived motion direction right after the offset of the motion stimulus. And, to confirm that the RT effect and the serial bias are driven by the shared underlying mechanisms, the following experiments (Experiment 2 and a reanalysis of Bae & Luck, 2020) investigated whether the RT effect occurs only when the serial bias occurs. 

Figure 1.

 

(a) Motion direction estimation with random-dot-kinematograms. After 1000 ms of RDK (100% coherence), only a white fixation dot on the black disk remained on the screen, indicating that participants should report the motion direction. As soon as participants moved the mouse, a response probe (a white line started from the central dot and aligned with the current mouse pointer) appeared, and participants adjusted the response probe to report the perceived motion direction. Response time was measured by the time difference between the onset of the response screen and the mouse click. (b) Mean bias as a function of absolute direction differences between the current and prior trial. Positive errors indicate repulsive serial bias. The inset bar graph represents the average bias for small (0° < Δ < 90°) and large (90° < Δ < 180°) direction differences. (c) Mean response time as a function of absolute direction difference between the current and prior trial. The inset bar graph represents the average response time for small and large direction differences. Error bars indicate ±1 SE.

To foreshadow the main findings, Experiment 1 found that the direction reports exhibited repulsive serial bias and that the response time was faster when the current motion direction was more dissimilar to the prior motion direction. The following experiments replicated these effects and further confirmed that the RT effect occurred only when repulsive serial bias was present. Our analyses with a large sample size (N = 96) also showed that the RT effect was positively correlated with the serial bias, confirming that the response time effect and the serial bias were driven by the same underlying processes. Together, these results provide converging evidence that the prior stimulus not only biases but also modulates the speed of response for a new stimulus. These findings strongly suggest that the proposed computational principles are not sufficient and a mechanism that modulates the speed of response is necessary to fully account for the serial bias phenomenon. 

Experiment 1: Response time effect in repulsive serial bias

Participants perceived motion directions in random dot kinematograms (RDKs) and reported the perceived motion direction in a continuous response scale (Figure 1a). We examined whether the serial bias in the direction reports would be associated with systematic changes in response time. 

Methods

Participants

A relevant previous study reported large effect sizes for the stimulus history effect on response time in visual working memory tasks (Cohen's d > 1) (Chen & Bae, 2024a). A priori power analysis indicated that at least 10 participants would be necessary to detect such an effect with a power of 0.8. Because the present study used a different task paradigm and the task may lead to a smaller effect size, we recruited a total of 20 college students (17 female). The study was approved by the Arizona State University Institutional Review Board and was conducted in accordance with the tenets of the Declaration of Helsinki. Participants provided informed consent. 

Stimuli and tasks

Stimuli were generated in MATLAB (MathWorks, Natick, MA) using Psychtoolbox (Brainard, 1997; Pelli, 1997) and were presented at 60 Hz on an LCD monitor (Dell U2412M; Dell Technologies, Round Rock, TX) with a white background (87.6 cd/m²) at a viewing distance of 70 cm. A white fixation dot (87.6 cd/m²) was continuously visible on a black disk (5° diameter, < 0.1 cd/m²) at the center of the screen except during the intertrial interval (see Figure 1a). We used a standard RDK algorithm to generate the motion stimulus (Roitman & Shadlen, 2002). Motion coherence was set to 100% to avoid the opposite-direction motion perception that can be observed in lower coherence RDKs (Bae & Luck, 2022) and to maximize the adaptation effect. Each trial began with the fixation dot for 500 ms on a black disk. The RDK was then presented for 1000 ms (60 frames). The direction of motion on a given trial was one of 16 discrete values (from 11.25° to 348.75°, in steps of 22.5°). Cardinal directions were excluded to prevent potential noise created by systematic interaction between the serial bias and the cardinal bias (Bae, 2024). At the end of the RDK, a mouse cursor appeared at the center of the disk (without a response probe). As soon as participants moved the mouse cursor, a white line connecting the center of the disk and a point on the circumference of the disk appeared. The orientation of the white line was continuously updated depending on the orientation created by the cursor and the center of the screen, allowing observers to adjust the orientation of the white line to the perceived motion direction. The observer finalized the direction report by clicking a mouse button. Note that many prior studies in serial bias presented a random probe at the start of the report and had participant adjust the probe for a report (Ceylan et al., 2021; Cicchini, Benedetto, & Burr, 2021; Fischer & Whitney, 2014; Fritsche et al., 2020; Hajonides, van Ede, Stokes, Nobre, & Myers, 2023; Manassi, Liberman, Kosovicheva, Zhang, & Whitney, 2018; Pascucci et al., 2019; Samaha, Switzky, & Postle, 2019; Teng, Fulvio, Jiang, & Postle, 2022; van Bergen & Jehee, 2019). Such a method is not ideal for investigating response times because response times can be impacted by the initial probe value. In the present study, the probe line appeared only after participants moved the mouse to make a report. After 16 practice trials, each participant completed 240 experimental trials, with exactly 15 trials for each of the 16 directions of motion. Participants took a short break after each block of 48 trials. 

Data analysis

For each trial, we computed the direction difference between a given trial and its preceding trial (i.e., previous trial direction minus current trial direction). The direction difference was wrapped within a range of −180° to 157.5° because the motion direction was in circular space (e.g., −180° difference is identical to 180° difference). The first trial in each block was necessarily excluded (a total of five trials per participant). Because we used 16 discrete motion directions, there were 16 discrete direction differences. To increase the signal-to-noise ratio in bias estimation, we used nine absolute direction differences (0°, 22.5°, 45.0°, 67.5°, 90.0°, 112.5°, 135.0°, 157.5°, and 180.0°) for analyses. 

Response error and response time were our main dependent variables. Response errors were computed for a given trial by taking the difference between the reported direction and the true motion direction for the current trial. The grand average of response errors was subtracted for each participant to compensate for the general bias in the reports (Samaha et al., 2019). We assigned a positive sign to the response errors if the direction report was biased away from the previous trial direction, and vice versa. The signed response errors were then averaged separately for each of the nine absolute direction differences. Serial bias exhibits greater biases for small stimulus differences than for large stimulus differences (Fischer & Whitney, 2014). We therefore assessed the presence of a serial bias by averaging the signed response error for the small direction difference trials (i.e., 22.5°, 45.0°, and 67.5°) and comparing this to the average of the signed response error for the large direction difference trials (i.e., 112.5°, 135.0°, and 157.5°) using a two-tailed paired t-test. Trials with 0° and 180° differences were excluded from the statistical comparisons because the direction of the bias (i.e., attraction versus repulsion) was undefined for them. Also, the trials with a 90° difference were excluded to balance the number of data points for small and large direction differences. 

To confirm that our analysis accurately reflects the direction of serial bias, we conducted a simulation where we simultaneously presented a sign for prior motion direction (red dot) and a sign for current motion direction (white dot) on the perimeter of the black aperture during the response period. We intentionally generated repulsive serial bias by clicking on the opposite side of the red dot (i.e., the previous) from the white dot (i.e., the current). This was repeated for 240 trials. Our analyses from this simulation showed repulsive bias as intended, confirming that our analyses accurately reflect the direction of serial bias. 

In addition to the bias measure, we also estimated the precision of direction reports by fitting a von Mises distribution to the unsigned response error distributions. The concentration parameter (κ) reflects the variability of the direction reports, with smaller values indicating more variability (and thus less precise). To relate the precision to the bias, we compared the average κ between the small (0° < Δ < 90°) and large (90° < Δ < 180°) direction differences using a two-tailed paired t-test. 

The response time was measured by the time difference between the onset of the response screen and the mouse click. The effect of the prior motion direction on the response time for the current motion direction was assessed by comparing the average response time for small direction difference trials (i.e., 22.5°, 45.0°, and 67.5°) and large direction difference trials (i.e., 112.5°, 135.0°, and 157.5°) using a two-tailed paired t-test. This was intended to match the statistical comparison with the serial bias analysis (i.e., the bias for small direction differences vs. the bias for large direction differences). We conducted additional analyses where we included 0° and 180° direction differences in the statistical test for the response time (Supplementary Materials). The results from this analysis were consistent with the results reported in the main text. Because the mean may not accurately reflect the central tendency of response time, we conducted additional response time analysis using median response time and found consistent results (Supplementary Materials). 

In all analyses, we excluded trials in which the current trial response error was larger than 60° (0.54% of the total trials), which likely reflected lapses of attention. We also excluded trials with response times faster than 300 ms or slower than 5 seconds (0.25% of the total trials) as they can bias the mean estimation. These trial exclusion criteria were set following our previous studies (Bae & Luck, 2020; Chen & Bae, 2024a). 

Data availability

The data are publicly available at https://osf.io/tsaef/. 

Results and discussion

Direction reports exhibited a repulsive serial bias (i.e., positive error) for the trials with small direction differences (Figure 1b). We tested this effect by comparing the average response errors for small (0° < Δ < 90°) versus large direction differences (0° < Δ < 90°). A paired t-test confirmed that the effect was significant, M = 3.60°; 95% confidence interval (CI), 2.23°–4.97°; t(19) = 5.49; p < 0.001; Cohen's d = 1.23. 

The reports were faster for large direction differences (Figure 1c). This effect was tested by comparing the average response time for small (0° < Δ < 90°) versus large (90° < Δ < 180°) direction differences. A paired t-test confirmed that the response time difference was significant, M = 113 ms; 95% CI, 84 ms–142 ms; t(19) = 8.12; p < 0.001; Cohen's d = 1.82. In a follow-up analysis, we confirmed that this RT effect was not a mere consequence of speed-precision trade-off, as the precision (κ) was higher for the trials with large direction differences than for the trials with small direction differences (small difference, κ = 28.99; large difference, κ = 37.67), t(19) = −3.95; p < 0.001; 95% CI, −13.28 to −4.08; Cohen's d = −0.80, paired t-test. 

The RT effect might be driven by the discrete stimulus sampling (i.e., 16 discrete directions) or by the evidence accumulation processes for the RDK task. However, we found the same RT effect even when the stimulus was randomly sampled from the 360° space and when the task did not require evidence accumulation processes (i.e., location estimation) (Supplementary Materials) (Chen & Bae, 2024a). Together, these results demonstrate that a prior stimulus modulates the speed of decision for a new stimulus and that the WM maintenance process is not necessary for the RT effect to occur. 

Experiment 2: No response time effect when no serial bias occurred

Experiment 1 demonstrated that repulsive serial bias in motion direction estimation is associated with systematic changes in response time, indicating that both the serial bias and the RT effect are based on common underlying processes. To further confirm this, Experiment 2 examined whether the RT effect would occur only when the serial bias was present. We used a dual-feature version of the motion task where participants reported the color instead of the direction of the dots in a random half of the trials (Figure 2a). The feature to be reported was indicated via a 100% valid pre-cue. In a post-cue version of this task (i.e., the feature to be reported was determined after the RDK) (Figure 3a), it was found that otherwise significant repulsive serial bias for the direction estimation was either eliminated or significantly reduced when the color was reported in the previous trial, even when the direction was encoded into working memory in the trial (Bae & Luck, 2020). Because WM encoding of the direction information was not necessary in the pre-cue task, we predicted that the serial bias would occur only after prior direction reports. More importantly, if the RT effect is dependent on the serial bias, then it should not occur when the serial bias is absent. 

Figure 2.

 

(a) A dual-feature motion task with a pre-cue. At the start of each trial, a fixation dot on the black disk was shown for 500 ms (not shown in the figure) and then a pre-cue (letter D or C) was presented (equal probability for each), indicating which feature should be reported on a given trial. The RDK was then displayed with either pink or green dots. For direction reports, the response cue was presented on the screen until mouse movement (as in Experiment 1). As soon as participants moved the mouse, a response probe appeared, and they adjusted the orientation of the response probe for reporting. For color reports, a pink circle and a green circle were presented on each side of the black disk (the side was counterbalanced), and participants pressed either the left or right arrow key (2AFC) on the keyboard to report the color of the dots in the RDK. (b) Serial bias following direction report (black) and color report (red). A positive error indicates repulsion from the prior motion direction. The inset represents the bias means for small (0° < Δ < 90°) and large (90° < Δ < 180°) direction differences for prior motion report (grays) and for prior color report (reds). (c) Mean response times for direction reports after previous direction report (black) and previous color report (red). The inset represents the response time means for small (0° < Δ < 90°) and large (90° < Δ < 180°) direction differences for prior motion report (grays) and for prior color report (reds). Error bars indicate ±1 SE.

Figure 3.

 

(a) The post-cue motion–color paradigm used in Experiment 1 of Bae and Luck (2020). The feature to be reported for a given trial was indicated by a post-cue (letter D or C) presented at the end of the motion stimulus. The study used three different versions of this task. (b) Serial bias, collapsed across the data from the three experiments of Bae and Luck (2020) (N = 72) plotted as in Experiment 2. Bae and Luck (2020) reported serial bias analyses for each of the three experiments separately. Here, we show the results collapsed across the three experiments. (c) Mean response times in Bae and Luck (2020), plotted as in Experiment 2. Error bars indicate ±1 SE.

Methods

Participants

A previous study (Bae & Luck, 2020) using a post-cue design reported large effect sizes (Cohen's d > 0.9). A priori power analysis indicated that at least 12 subjects were necessary to detect such an effect with the power of 0.8. The present experiment recruited a total of 24 college students between the ages of 18 and 30 (12 female). The study was approved by the Arizona State University Institutional Review Board and conducted in accordance with the tenets of the Declaration of Helsinki. Participants provided informed consent. 

Stimuli, tasks, and analysis

The RDK was presented with either green (RGB = [68, 190, 170]) or pink (RGB = [236, 143, 173]) dots. Each trial started with a presentation of the central fixation dot on a black disk (500 ms). A 100% valid pre-cue (letter “D” for direction or “C” for color; 87.6 cd/m²; 0.8° × 1.0° width × height) was then provided at the center of the disk for 500 ms. The pre-cue randomly indicated the feature to be reported across the trials. At the end of the RDK, participants reported either the direction of the dots or the color of the dots according to the pre-cue. The letter cue was presented on the black disk to remind the participants which feature should be reported. The method for the direction report was identical to Experiment 1. For the color report, two colored circles—red and green (diameter 1.5°)—were presented 4° to the left and to the right of the fixation dot. Participants pressed either the left or right arrow key on a computer keyboard to report which of the two colors was identical to the color of the dots. The correct color was equally likely to appear on the left and right sides. We used this simple two-alternative forced choice (2AFC) color task to avoid any possible interference from the color task to the motion task. The next trial began after a 1000-ms intertrial interval during which only the black disk was visible. As in Experiment 1, a discrete set of 16 motion directions was tested. Participants completed a total of 448 trials (224 trials for color report and 224 trials for direction reports) after completing 16 practice trials. 

The analysis focused on the presence of serial bias in the direction reports separately for the trials following color reports and direction reports. The general procedure for quantifying serial bias was identical to Experiment 1. A statistical test comparing the serial bias between the two types of previous trial was conducted using a 2 (small vs. large direction differences) × 2 (prior direction reports vs. prior color reports) within-subject analysis of variance (ANOVA). 

The general procedure for quantifying the RT effect was identical to Experiment 1; however, the response time was analyzed separately for the trials preceded by color reports and for the trials preceded by direction reports. As in the serial bias analysis, a 2 × 2 within-subject ANOVA was conducted to test statistical significance of the RT effect. 

We applied the same trial exclusion criteria used in Experiment 1. We excluded trials in which the current trial response error was larger than 60° (2.36% of the total direction-report trials) and trials with response times faster than 300 ms or slower than 5 seconds (1.55% of the total direction-report trials), as they could bias the mean estimation. 

Data availability

The data are publicly available at https://osf.io/tsaef/. 

Results and discussion

The mean accuracy of the color reports was 0.97. The direction report exhibited a typical pattern of repulsive serial bias following previous direction reports, whereas there was only a hint of this effect following previous color reports (Figure 2b). This was tested by a 2 × 2 ANOVA with direction differences (small vs. large) and previous trial type (direction vs. color) as the within-subject variables. The main effect of direction difference was significant, F(1, 23) = 8.91, p = 0.007, η_p² = 0.28, but the main effect of the previous trial type was not, F(1, 23) = 2.98, p = 0.098, η_p² = 0.11. Critically, the two-way interaction was significant, F(1, 23) = 5.03, p = 0.035, η_p² = 0.18. Planned pairwise comparisons with Bonferroni correction confirmed a significant repulsive serial bias following direction reports, M = 2.08°; 95% CI, 1.1°–3.06°; t(23) = 4.39; p < 0.001; Cohen's d = 0.9, paired t-test. However, the serial bias was not found following color reports, M = 0.63°; 95% CI, −0.63° to 1.97°; t(23) = 1.07; p = 0.30, paired t-test. 

The response time was faster for large direction differences following direction reports; however, the RT effect was not evident following color reports (Figure 2c). This was tested by a 2 × 2 ANOVA with direction differences (small vs. large) and previous trial type (direction vs. color) as the within-subject variables. The main effect of direction difference was significant, F(1, 23) = 10.03, p = 0.004, η_p² = 0.30. However, the main effect of the previous trial type was not, F(1, 23) = 0.47, p = 0.50, η_p² = 0.02. As in the bias effect, the two-way interaction was significant, F(1, 23) = 15.76, p < 0.001, η_p² = 0.41. Planned pairwise comparisons with Bonferroni correction confirmed a significant response time effect following direction reports, M = 147 ms; 95% CI, 85 ms–209 ms; t(23) = 4.95, p < 0.001, Cohen's d = 1.01, paired t-test). However, there was not a significant response time effect following color reports, M = −1 ms; 95% CI, −62 ms to 60 ms; t(23) = −0.04; p = 0.97, paired t-test. These results demonstrate that the RT effect was not due to the estimation procedure used in the study. In additional analyses, we confirmed that the RT effect following direction reports was not merely driven by a speed–precision trade-off, as the precision was greater for large direction differences; for small differences, 24.06; large differences, 30.90; t(23) = −2.57; p = 0.017; 95% CI, −12.34 to −1.33; Cohen's d = −0.52, paired t-test. These results demonstrate that the RT effect occurred only when the repulsive serial bias was present, providing converging evidence that the RT effect and the repulsive serial bias are not independent but share the same underlying mechanism. 

Response time analysis for Bae and Luck (2020)

Bae and Luck (2020) conducted post-cue versions of the motion–color task (Figure 3a) and found that the repulsive serial bias for direction reports was eliminated or significantly reduced following color reports as in Experiment 2. Here, we investigated whether the RT effect in the study occurred only when the repulsive serial bias was present. 

Methods

Participants, stimuli, tasks, and analysis

Bae and Luck (2020) recruited a total of 72 college students between the ages of 18 and 30 years (46 female). The study was approved by the Arizona State University Institutional Review Board and was conducted in accordance with the tenets of the Declaration of Helsinki. The procedure of the motion–color tasks used in the study was similar to the procedure used in Experiment 2 of the present study. The main difference was that a post-cue was used to indicate which feature should be reported at the end of the RDK (Figure 3a). Bae and Luck (2020) used three versions of post-cue motion–color experiments (N = 24 each). The differences among the three experiments were in the color sampling (within category vs. between categories) and the method for the color report (AFC vs. estimation). However, the motion direction task was identical among all three experiments. 

The present study focused on the aggregated data from the three experiments. This increased the signal-to-noise ratio by 73% compared with Experiment 2 of the present study (because the sample size was three times larger than the sample size in Experiment 2). Our main goal in this study was investigating the response time effect for direction reports (which has not been reported previously) as a function of the feature reported in the previous trial. As in Experiments 1 and 2 of the present study, we excluded trials with response error larger than 60° (1.36% of the total direction-report trials) and trials with response times faster than 300 ms or slower than 5 seconds (0.71% of the total direction-report trials). The results of response time analysis for each individual experiment in Bae and Luck (2020) are reported in the Supplementary Materials. 

Data availability

The data from this study are available at https://osf.io/g35jp. 

Results and discussion

Both the repulsive serial bias and the RT effect were evident after prior direction reports, but not after prior color reports (Figures 3b and 3c). These effects were tested using the 2 × 2 ANOVAs as in Experiment 2. For the repulsive serial bias, the main effect of direction difference was significant, F(1, 71) = 42.14, p < 0.001, η_p² = 0.37, and the main effect of the previous trial type was significant, F(1, 71) = 38.7, p < 0.001, η_p² = 0.35. The two-way interaction was also significant, F(1, 71) = 23.42, p < 0.001, η_p² = 0.25. Planned comparisons with Bonferroni correction confirmed a significant repulsive serial bias following direction reports, M = 2.88°; 95% CI, 2.18°–3.58°; t(71) = 8.23; p < 0.001; Cohen's d = 0.97, paired t-test. Direction reports following color reports exhibited a hint of serial bias, mostly driven by the attractive bias for large direction differences (Figure 3b). However, this was not significant after Bonferroni correction, M = 0.78°; 95% CI, −0.06° to 1.58°; t(71) = 2.17; p = 0.034 (>0.025), paired t-test. 

For the RT effect, the main effect of direction difference was significant, F(1, 71) = 38.33, p < 0.001, η_p² = 0.35, and the main effect of the previous trial type was significant, F(1, 71) = 10.9, p = 0.002, η_p² = 0.13. The two-way interaction was also significant, F(1, 71) = 25.3, p < 0.001, η_p² = 0.26. Planned comparisons with Bonferroni correction confirmed a significant RT effect following direction reports, M = 133 ms; 95% CI, 97 ms–168 ms; t(71) = 7.47; p < 0.001; Cohen's d = 0.88, paired t-test. However, no significant RT effect following color reports was found, M = 18 ms; 95% CI, −13 ms to 49 ms; t(71) = 1.16; p = 0.25, paired t-test. We confirmed that the RT effect following direction reports was not merely driven by a speed–precision trade-off, as the precision was greater for large direction differences (small difference, 26.37; large difference, 30.70), with t(71) = −3.29; p = 0.002; 95% CI, −6.96 to −1.71; Cohen's d = −0.39, paired t-test). 

In a post hoc analysis, we tested whether the RT effect was driven by expedited decisions for large direction differences or delayed decisions for small direction differences by comparing the response time between the two types of preceding trials separately for small and large direction differences with Bonferroni corrections. For large direction differences, the response time following direction reports was significantly faster than the response time following color reports, t(71) = −6.12; p < 0.001; 95% CI, −145 ms to −74 ms; Cohen's d = −0.72, paired t-test. However, for small direction differences, the response time was comparable between the two types of preceding trials, t(71) = 0.268; p = 0.79; 95% CI, −35 ms to 47 ms; paired t-test. This result suggests that, at least in the context of this experiment, the RT effect was mainly driven by expedited decision for large direction differences (see Supplementary Materials for corresponding analysis for Experiment 2). 

Finally, to more directly examine the relationship between the RT effect and the serial bias, we conducted correlation analyses using the data from Experiment 2 of the present study and the data from Bae and Luck (2020) (N = 96). The correlation was not significant following color reports, t(94) = 0.88; p = 0.380; 95% CI, −0.012 to 0.29; r = 0.09 (Figure 4a). However, the RT effect was positively correlated with the magnitude of the serial bias following direction reports, t(94) = 2.48; p = 0.015; 95% CI, 0.05–0.43; r = 0.25. The RT effect was stronger when the serial bias was stronger. These results provide additional evidence that the RT effect and the serial bias were driven by the shared mechanism. 

Figure 4.

 

(a, b) Correlations between the RT effect and the serial bias for direction report following previous color report (a) and previous direction report (b). The magnitude of the RT effect (i.e., RT for small direction difference minus RT for large direction difference) was positively correlated with the magnitude of the repulsive serial bias following direction reports. Solid lines represent regression fits.

General discussion

Reported perception of a new stimulus is systematically biased by a task-irrelevant prior stimulus. Understanding the mechanisms underlying this serial bias has been a central issue in visual cognition because the bias may reflect how the visual system achieves perceptual stability over time under uncertainty while maintaining efficiency in visual information processing (Cicchini et al., 2024; Kiyonaga et al., 2017; Manassi & Whitney, 2024; Pascucci et al., 2023). Yet, serial bias is a complex phenomenon that exhibits large variability across different task contexts (Ceylan & Pascucci, 2023) and groups of individuals (Bansal et al., 2023; Stein et al., 2020). Despite the complexity of the phenomenon, however, studies have focused on only one aspect of perceptual reports (i.e., biases), neglecting other relevant aspects of perceptual behavior that may also reflect the underlying mechanism for the bias. Here, we investigated how the response time may vary with the prior stimulus under the context of the repulsive serial bias and found converging evidence that the prior stimulus systematically modulates the speed of response for a new stimulus. Perceptual reports were faster when the current stimulus was more dissimilar to the prior stimulus. This RT effect occurred only when the repulsive serial bias was present, and the magnitude of the RT effect was positively correlated with the magnitude of the repulsive serial bias. These results suggest that the RT effect and the repulsive serial bias were driven by the common mechanisms. 

Current theories posit that the repulsive serial bias is driven by low-level adaptation mechanism (Fritsche et al., 2017; Fritsche et al., 2020; Moon & Kwon, 2022; Pascucci et al., 2019) or by adaptation-like mechanisms driven by active removal of irrelevant working memory item (Shan & Postle, 2022). Although they do not make explicit predictions on how adaptation might impact the speed of response for a new stimulus, they may possibly explain the RT effect observed in the present study. For example, RTs might have been slower for the trials with small stimulus differences if the task was more difficult due to the additional noise introduced by the changes in neural tuning functions driven by adaptation (Schwartz Hsu, & Dayan, 2007). In addition, RTs might have been faster for the trials with large stimulus differences if adaptation expedited the evidence accumulation process for the direction opposite to the previous trial motion direction (i.e., motion aftereffect) (Anstis, Verstraten, & Mather, 1998; Mather, Pavan, Campana, & Casco, 2008). However, there are reasons to believe that adaptation alone is not sufficient to explain the repulsive serial bias and the RT effect. Specifically, in our motion–color tasks (Experiment 2 and Bae & Luck, 2020), neither the repulsive serial bias nor the RT effect occurred even when participants perceived and remembered the previous trial motion stimulus but they did not report the motion direction in the previous trial. In addition, a recent study found that observers systematically adjusted the mouse report in relation to the prior stimulus during the report period of orientation estimation tasks (Chen & Bae, 2024a), which cannot be easily explained by the adaptation mechanism, as the mechanism should impact the perceptual encoding rather than post-perceptual decision making. 

We therefore suggest that mechanisms that can modulate the speed of decision for a new stimulus are also necessary to account for the RT effect. One possible mechanism would be a choice response selection mechanism for a continuous response scale proposed in the studies of perceptual decision making. The traditional choice response selection mechanism was developed in a simpler scenario where a perceptual decision is made between a small set of discrete alternatives (Ratcliff & McKoon, 2008; Smith, 1968). Recent studies, however, have shown that a similar mechanism can be applied to cases where a decision is made on a continuous response scale, such as the typical estimation paradigm used in many serial bias studies (Kvam, Marley, & Heathcote, 2023; Smith, 2019). According to this model, choice response time is dependent on the number of response alternatives and the similarity among them, with response time being slower when alternative responses are more similar to the target response. If we assume that the serial bias is driven by a response selection process that takes the prior stimulus into account, then it makes sense that the response selection should be easier (i.e., faster RT) when the prior stimulus is more dissimilar to the current stimulus because one has to avoid selecting the prior stimulus to accurately report the newly perceived stimulus. Likewise, if the response selection process does not take the prior stimulus into account (i.e., no serial bias), then the RT should be independent of the prior stimulus, as observed in the present study. 

If both low-level adaptation and post-perceptual decision processes contribute to the repulsive serial bias and the RT effect, then it would be important to understand how the low- and high-level components of the biases may eventually interact to produce the observed repulsive effects. One way would be dissociating the two components by changing the locations for the stimulus and response in a given task (Bae, 2024; Bae & Luck, 2019; Fritsche et al., 2017; Pascucci & Plomp, 2021). Because adaptation is a spatially specific process such that the contribution of adaptation to the serial bias should be reduced (if not absent) when the stimuli are presented in separate locations (Bae, 2024; Bae & Luck, 2019), it should be possible to investigate purer contributions of post-perceptual processes on the repulsive serial bias and the RT effect. Alternatively, one could employ gratings with different spatial frequencies, as adaptation is spatial frequency specific (Ceylan et al., 2021). Future studies are necessary to test these possibilities. 

Because the present study focused on repulsive serial bias, we do not make strong conclusions on how the RT might vary under the context of attractive serial dependence. In contrast to the repulsive serial bias, however, it has been proposed that post-perceptual decision processes play a critical role in attractive serial bias (Fritsche et al., 2017; Fritsche et al., 2020; Moon & Kwon, 2022; Pascucci et al., 2019; for an alternative account, see also Cicchini, Mikellidou, & Burr, 2017; Fischer & Whitney, 2014; Murai & Whitney, 2021). Also, a specific decision mechanism, such as decisional inertia (Pascucci et al., 2019), may lead to faster decisions when a new stimulus is more similar to the previous stimulus (Cicchini et al., 2018). Interestingly, a similar pattern of RT effect can be seen in the present study, as well. Some observers exhibited an attractive serial bias, and their response time tended to be faster for the trials with small stimulus differences (negative RT effect) (Figure 4). These results suggest that the attractive and repulsive serial biases may reflect decisional processes with differential decisional criteria, such that attractive serial bias occurs when observers exaggerate the similarity between the two successive stimuli and repulsive serial bias occurs when they exaggerate the difference between them (Chen & Bae, 2024a). More generally, the present study found that serial bias ranged from 5° attraction to 10° repulsion across individuals (Figure 4). This individual difference in serial bias suggests that different individuals dynamically adjusted their decision criteria over time to optimize their decision outcomes, as proposed by the Thurstonian framework (Thurstone, 1927). Investigating this possibility would be an important direction for future research to better understand how the visual system integrates past experiences into the decision for a new stimulus. 

Conclusions

Serial bias is a complex perceptual phenomenon that exhibits significant variability depending on specific task contexts. The present study found converging evidence that task-irrelevant prior stimuli not only bias but also affect the speed of response to a new stimulus. These findings strongly suggest that existing computational principles are insufficient and that mechanisms systematically modulating the speed of decision for new stimuli are necessary to fully account for the serial bias phenomenon. 

Acknowledgments

Commercial relationships: none. 

Corresponding author: Gi-Yeul Bae 

Email: [email protected]. 

Address: Department of Psychology, Arizona State University, Tempe, AZ, USA.