Statistical learning of spatiotemporal regularities dynamically guides visual attention across space

Xu, Zhenzhen; Theeuwes, Jan; Los, Sander A.

doi:10.3758/s13414-022-02573-5

Statistical learning of spatiotemporal regularities dynamically guides visual attention across space

Open access
Published: 07 October 2022

Volume 85, pages 1054–1072, (2023)
Cite this article

Download PDF

You have full access to this open access article

Attention, Perception, & Psychophysics Aims and scope Submit manuscript

Statistical learning of spatiotemporal regularities dynamically guides visual attention across space

Download PDF

3248 Accesses
3 Citations
Explore all metrics

Abstract

In dynamic environments, statistical learning of spatial and temporal regularities guides visual attention in space and time. In the current study, we explored whether and how combined spatiotemporal regularities regarding target events guide visual attention. In three experiments, participants performed the additional singleton task. They were asked to search for a target stimulus with a unique shape among five non-target distractors and respond to the orientation of a line inside the target. Unbeknownst to the participants, the moment in time that the search display was presented was predictive of the target location. Specifically, the target was more likely to be presented at one high-probability location after a short interval and at another high-probability location after a long interval. The results showed that participants’ performance was better for high-probability locations than for low-probability locations. Moreover, visual search efficiency was greater when the target appeared at the high-probability location after its associated interval than when it occurred there after its nonassociated interval, regardless of whether the distribution of intervals was uniform (Experiment 1), exponential (Experiment 2), or anti-exponential (Experiment 3). Taken together, the results indicate that implicitly learned spatiotemporal regularities dynamically guide visual attention towards the probable target location.

Visual statistical learning of temporal structures at different hierarchical levels

Article 11 April 2016

Statistical learning of target location and distractor location rely on different mechanisms during visual search

Article 13 December 2022

Visual statistical learning is not reliably modulated by selective attention to isolated events

Article 30 August 2014

Use our pre-submission checklist

Avoid common mistakes on your manuscript.

In a dynamically changing world, it can be quite challenging to extract information that is relevant for the ongoing task. To cope with this problem, it has been shown that individuals implicitly use spatial and temporal regularities that the environment provides (Fiser & Aslin, 2001; Saffran et al., 1996). In particular, it has been revealed that statistical learning (SL) of spatial and temporal distributions of events biases visual attention to relevant locations and relevant time points (Olson & Chun, 2001; Turk-Browne et al., 2005; Wagener & Hoffmann, 2010; Wang & Theeuwes, 2018a), which is consistent with the selection history component in the taxonomy of attentional control proposed by Awh et al. (2012; see also Failing & Theeuwes, 2018; Theeuwes, 2019).

The idea that SL of spatial regularities biases visual attention is not new. In early studies of contextual cueing, it was found that participants are able to implicitly learn the associations between spatial configurations and target locations. Target localization and discrimination are facilitated when targets consistently appear at a specific location within an earlier presented context than when they appear at a location within a new context even if participants have no explicit knowledge about the contexts (Chun & Jiang, 1998). This benefit suggests that the visual context serves as a cue which implicitly guides spatial attention to the potential target location (Chun & Jiang, 1999). In addition, it has also been shown that visual search efficiency for targets is higher when the target appears at probable locations than when it appears at improbable locations while participants are unaware of the distribution regularities (Geng & Behrmann, 2002, 2005; Jiang et al., 2013), suggesting that spatial attention is oriented implicitly towards probable target locations. Similarly, visual search efficiency for targets has been found to be higher when distractors appear at high-probability distractor locations than when they appear at low-probability distractor locations, indicating that SL of distractor regularities guides spatial attention away from probable distractor locations (Wang et al., 2019; Wang & Theeuwes, 2018a, b, c).

Alongside the study of attentional orienting to locations in space, there has been growing interest in the study of attentional orienting to moments in time (Coull & Nobre, 1998; Miniussi et al., 1999; Nobre & van Ede, 2018). Within this field, it has been shown that temporal regularities of target presentation may come to expression in performance measures, suggesting that, due to SL of temporal regularities, moments in which the target presentation is expected are prioritized over other moments in time. In an early study, Olson and Chun (2001) manipulated the sequential structure of visual events, in which the timing of the next element could be predicted from the temporal position within a sequence. They found that this sequential temporal context could be implicitly learned and come to guide visual attention to a point in time. Subsequent studies confirmed that SL of temporal order regularities biases visual attention in an automatic and implicit way (Li & Theeuwes, 2020; Turk-Browne et al., 2005; Wang et al., 2021; Zhao et al., 2013).

To study how attention comes to prioritize certain time points over others, researchers have varied the foreperiod, which refers to the interval between a neutral cue and the target stimulus (Niemi & Näätänen, 1981). During this interval, a postulated process of temporal preparation or expectation develops, which facilitates responses to the target stimulus (e.g., Los et al., 2014; Salet et al., 2022). The time course of general temporal preparation has been shown to be tuned by the context of intervals. When the interval randomly varies from trial to trial according to a uniform distribution, participants respond faster and more accurately to the target stimulus as the interval lengthens. When the proportion of short intervals increases, the function relating reaction time (RT) to interval gradually lies down, and it becomes approximately flat in the case of an exponential interval distribution. By contrast, when the proportion of long intervals increases, the RT-interval function gradually becomes steeper (Los et al., 2017; Los et al., 2021; Niemi & Näätänen, 1981; Trillenberg et al., 2000; Vangkilde et al., 2013). There is ongoing controversy as to whether general temporal preparation is driven by the hazard function (Nobre et al., 2007; Nobre & van Ede, 2018; Visalli et al., 2019; Visalli et al., 2021) or by learned memory representations of time (Los et al., 2017; Los et al., 2021; Salet et al., 2022). It is furthermore unclear which process(es) are facilitated by temporal preparation. Some studies favor a motor or premotor locus for the effect of temporal preparation, while others claim that temporal preparation affects spatial attention (Rolke & Ulrich, 2010; Vangkilde et al., 2013). In addition, the role of arousal and alertness are also mentioned in relation to temporal preparation (Hackley & Valle-Inclán, 1998; Steinborn & Langner, 2012). Regardless of how temporal preparation is explained, there is consensus that it reflects a gradual learning of the possible moments in time for either target occurrence or response execution.

Recently, researchers have started to combine paradigms of spatial and temporal learning. This work has shown that task performance is influenced by SL of spatiotemporal regularities when the interval provides information about the location or identity of upcoming events (Thomaschke & Dreisbach, 2013; Thomaschke et al., 2011). For instance, Wagener and Hoffmann (2010; Experiment 2) used a choice RT task, in which a target stimulus was presented at one of two possible locations after one of two possible intervals. Critically, the interval indicated the location at which the target was most likely to appear. The target was presented more frequently at one location after a short interval and at another location after a long interval. They found that participants responded faster and more accurately when the target appeared at a temporally valid location than appeared at a temporally invalid location, revealing the behavioral expression of spatiotemporal regularities about target events. However, on the basis of this study, it is difficult to determine what caused this behavioral benefit. On one hand, it could reflect visual attentional orienting, resulting in faster processing of a stimulus that appeared at the likely location. On the other hand, it could reflect motor preparation, resulting in a faster motor response to a stimulus appearing at the likely location (Thomaschke et al., 2011; Wagener & Hoffmann, 2010).

To disentangle these two accounts, Thomaschke et al. (2011) used four target stimuli which differed along the dimension of orientation (vertical, horizontal) and shape (diamond, oval) in a speeded binary choice task. One of these dimensions determined the response, while the other dimension was response-irrelevant. In addition, one of the dimensions was correlated with the preceding interval, while the other dimension was not. The critical finding was that time-event regularities only affected performance when the response-relevant dimension was correlated with the interval, not when the response-irrelevant dimension was correlated with the interval. Thus, they concluded that the effect of the time-event regularities was response specific. Subsequent studies confirmed that motor responses play an important role in the learning of time-based event regularities (Thomaschke & Dreisbach, 2013; Volberg & Thomaschke, 2017).

At the same time, other studies provided evidence in favor of the attention account. In a Posner spatial cueing task, Rieth and Huber (2013) implicitly manipulated the validity of the exogenous cue based on the interval between the cue and the target stimulus. For one group, the target appeared on the cued side after a short interval and on the uncued side after a long interval, while for another group this contingency was reversed. As participants’ task was to press a key after the appearance of the asterisk target stimulus, the response was independent of the spatiotemporal regularities of the target. The two groups showed differential attention orienting effects. The cueing effect was significantly larger for the short invalid/long valid group, while inhibition of return (IOR; Klein, 2000) was significantly larger for the short valid/long invalid group. These findings remained in place during a test phase in which the regularities were removed, indicating that the interval-based spatiotemporal regularities about the targets were implicitly learned. Based on these findings, they argued that spatial attention implicitly adapts to spatiotemporal regularities.

In a recent eye-movement study, Pfeuffer et al. (2020) associated two intervals with two target locations. The interval predicted the upcoming target location with 100% validity. The results showed that participants first moved their eyes to the location associated with the short interval. If there was no stimulus presented after the short interval had passed (trials with the long interval), then they looked towards the location associated with the long interval. While revealing the adaptation of overt attention to implicit spatiotemporal target regularities, this design is somewhat impoverished because only one stimulus appears at either one of two locations.

Boettcher et al. (2022) explored the adaption to the spatiotemporal regularities of a target in a noisy visual search task. Participants were asked to search for eight instances of a target (vertical line) that faded in and out of a display containing similar transient distractors (tilted lines). If the target was found, they had to click on it. On each trial, the presentation of four of the eight targets was spatiotemporally predictable, as each target appeared in a specific quadrant at a specific time point within the trial. The other four targets appeared randomly in time and space within each trial. Participants’ performance was significantly better for spatiotemporally predictable targets than for unpredictable targets. The corresponding eye-movement data also showed a higher probability to fixate the target quadrant for predictable compared with unpredictable targets. Based on these findings, the authors concluded that spatiotemporal regularities guide attention in dynamic visual search. This study is important because it extends the findings of SL of spatiotemporal regularities to noisy dynamic contexts. However, the findings do not necessarily imply that participants have learned to orient attention to a particular location and moment in time. Instead, participants may have learned that they needed to make a motor response towards a particular quadrant at certain time points. Also, since the detection of the blurry target among noisy distractors required eye movements, participants may have learned that they needed to make a saccadic eye movement to detect the target that was presented at particular quadrants at particular moments in time. Therefore, this study does not necessarily show learning of spatiotemporal attentional orienting, as it could reflect spatiotemporal (saccadic) motor learning instead.

As time-based motor preparation plays an important role in explaining the behavioral expression of the spatiotemporal regularities, studies in which the motor response is independent of the spatiotemporal regularities would provide more convincing evidence for the notion that attentional orienting can be guided by spatiotemporal regularities. A recent study of Xu et al. (2021) did exactly this by using the additional singleton paradigm (Theeuwes, 1991, 1992) to explore the effect of spatiotemporal regularities regarding the salient distractor. Participants searched for a unique target shape (e.g., a diamond) presented among five differently shaped nontargets (e.g., circles) and responded to the orientation of a line within the target. On the majority of the trials, one of the nontargets had a unique color (the color singleton distractor). Critically, the distractor appeared relatively frequently at one high-probability distractor location after a short interval and at another high-probability distractor location after the long interval. The results showed higher search efficiency for targets when the distractor appeared at a high-probability location after its associated interval than when it appeared at that location after the other interval. Since the motor response was fully separated from the spatiotemporal regularities in this study, these findings indicated that attention was dynamically guided away from the probable distractor location at the specific moments in time through SL of the spatiotemporal regularities.

The current study investigated whether spatiotemporal regularities of target events dynamically guide attention towards the probable target location. To that end, we adopted the paradigm used in Xu et al. (2021) but without a colored singleton distractor. Participants were instructed to search for a unique shape singleton presented among five nontargets and they responded to the orientation of the line inside it. To minimize eye movement learning, participants were asked to keep their gaze at the center of the screen all the time during this covert attention task. Following the fixation dot, the search display was presented either after a short or long interval. Unbeknownst to participants, the target appeared more frequently at one high-probability target location after a short interval and at another high-probability target location after a long interval. The probability that the target appeared at any one of the other four low-probability locations after either interval was equal. If spatiotemporal regularities dynamically guide attentional selection, performance should be better if the target is presented at the temporally congruent high-probability location than when it is presented at either the temporally incongruent high-probability location or at any one of the low-probability locations. We tested these predictions across three experiments, in which we used different distributions of intervals (uniform, exponential and anti-exponential distribution) to ensure that the spatiotemporal guidance of attention was independent of general temporal preparation.

Experiment 1

In Experiment 1, we tested whether participants could learn spatiotemporal regularities regarding the target with two equally likely intervals (i.e., under a uniform distribution, 1:1). From the onset of the fixation dot, the search display was presented with equal probability after either a 500-ms interval or a 1,500-ms interval. Critically, the target appeared more frequently at one high-probability target location after a short interval, while it appeared more frequently at the opposite high-probability target location after a long interval. We expected that attentional selection would be dynamically guided by spatiotemporal regularities, so the temporally congruent high-probability target location should facilitate search relative to either the temporally incongruent high-probability location or any one of the neutral locations.

Method

Participants

The sample size was pre-determined based on the effect size observed by Wagener and Hoffmann (2010). To observe the reported effect of η_p² = .35 with .80 probability (α = .05) in a 2 × 2 (interval × target location) repeated-measures analysis of variance (RM-ANOVA), MorePower 6.0.4 (Campbell & Thompson, 2012) suggests a sample of 18 participants. However, we aimed for a much larger sample size of 40 participants because the experiment was conducted online. To this end, we measured 47 participants, of whom six were excluded because of low accuracy (less than 70%). Participants were compensated by credits or payments. The final sample thus consisted of 41 participants (age 20.0 ± 3.37 years, 30 females), including 31 participants recruited from the Vrije Universiteit Amsterdam and 10 participants recruited from Prolific. Participants all gave written informed consent at the beginning of the experiment, and this study was approved by the Ethical Review Committee of the Faculty of Behavioral and Movement Sciences of the Vrije Universiteit Amsterdam.

Apparatus

The experiment was built and run on OpenSesame/OSWeb (Version 3.3.9b; Mathôt et al., 2012). The programmed experiment was hosted by JATOS. Participants were directed to the experiment on JATOS by the participant recruitment platforms Prolific and SONA (VU). Participants were instructed to perform the experiment in a quiet environment on a laptop or PC, and to turn off all possible sources of distraction.

Procedure and design

As shown in Fig. 1, each trial started with a central white fixation dot against a black background. The fixation dot was a white filled circle (8-pixel radius) with a central hole (2-pixel radius). After an interval of either 500 ms or 1,500 ms, a search display consisting of six shapes arranged on an imaginary circle (210-pixel radius) around the central fixation dot was presented for 2,000 ms or until a response was made. These shapes were either a diamond (112 × 112 pixel) among five circles (51-pixel radius) or vice versa. All shapes had either a red or green outline (randomly varied across trials) with a gray line oriented horizontally or vertically inside it. Participants were asked to search for the unique shape and respond to the orientation of the line inside it by pressing the “z” or “m” key. They were instructed to fixate at the central fixation dot throughout a block of trials and to respond as quickly and accurately as possible. After the response or search display timeout, a feedback display (57 × 57 pixel) with either a happy face (correct response) or a sad face (incorrect response) was presented for 300 ms. The intertrial interval was from 500 ms to 750 ms.

For each participant, two opposite locations were designated as high-probability target locations and denoted as the “high-short” and “high-long” target locations. The three possible pairs of high-probability target locations as well as the location assigned to the high-short and the high-long condition within each pair were randomized across participants. In half of the trials, the interval between the onset of the fixation dot and the search display was 500 ms. Across all trials with the 500-ms interval, the target was more frequently presented at the high-short target location (64.3%) than at either the high-long target location or any of the four low-probability target locations (7.1% each). In the other half of the trials, the interval was 1,500 ms. Across all trials with the 1,500-ms interval, the target was more frequently presented at the high-long location (64.3%) than at either the high-short location or any of the low-probability target locations (7.1% each). All conditions were randomized in each experimental block (shown in Table 1). The experiment consisted of a practice block of 25 trials and 6 experimental blocks of 112 trials each.

Table 1 Number of trials in each condition of each experimental block in Experiments 1, 2, and 3

Full size table

At the end of the experiment, we tested participants’ awareness about the regularities. First, participants were asked to indicate at which two locations they thought the target appeared most often to estimate their awareness about the spatial distribution of targets. Then, participants were asked to indicate the frequent target location after a short interval and the frequent target location after a long interval to estimate their awareness about the spatiotemporal regularities.

Data analysis

Trials with RT below 200 ms or above 2,000 ms (1.68% of all trials) were excluded. For the analysis of RT, incorrect trials (10.79 % of all trials) and trials with RTs outside ± 2.5 standard deviations of the condition mean for each participant (2.27 % of all trials) were also excluded.

We analyzed the filtered RT data with linear mixed models (LMMs) and the accuracy (ACC) data with generalized linear mixed models (GLMMs) using the lme4 package (Version 1.1.29; Bates et al., 2015) in RStudio (RStudio Team, 2021). Compared with repeated measures ANOVAs (RM-ANOVAs), mixed-effects models can deal with unbalanced designs (Jaeger, 2008), such as the current design, with different number of trials in the various conditions.^{Footnote 1}

We ran LMMs with RT as the dependent variable and GLMMs with ACC as the dependent variable. More specifically, the (G)LMMs included the fixed effects of interval (500 ms, 1,500 ms), target location (high-short, high-long, low) and their interaction. In addition, as participants’ awareness of the regularities, intertrial location priming, and the physical location of the target might explain some variance of participants’ performance in the visual search task (Huang et al., 2021), we also included participants’ awareness of spatial regularities (both aware, one aware, and unaware), awareness of the high-probability location after the short interval (aware, unaware), awareness of the high-probability location after the long interval (aware, unaware), intertrial target location priming (yes, no), and physical location of the target on the screen (0–5) as additional fixed effects in our models. All the fixed effects were dummy coded. The random effect structure of the models was determined by running the maximal random effect structure justified by the design which allowed model convergence (Barr et al., 2013). In Experiment 1, for the LMMs, the random effect structure included a by-participant random intercept and a by-participant random slope for target location and interval. For the GLMMs, the random effect structure included a by-participant random intercept and a by-participant random slope for target location. The p values were obtained by the likelihood ratio test for all model comparisons in which the model with the fixed effect of interest was compared with the model without it (α = .05). Pairwise tests were investigated using the emmeans package with the Tukey correction for multiple comparisons.

Results

RTs

Figure 2A shows the mean RT as a function of interval (500 ms, 1,500 ms) and target location (high-short, high-long, low). As shown in Table 2, the LMM analyses of RTs revealed a significant fixed effect of interval, showing a reduction of RT from the short to the long interval condition. This reflects the effect of general temporal preparation under the uniform interval distribution. There was also a significant fixed effect of target location. Compared with the mean RT for the low-probability target locations, the mean RTs were significantly shorter for the high-short location, β = −77.900, SE = 8.360, t = −9.319, p < .001, and for the high-long location, β = −26.200, SE = 9.050, t = −2.894, p = .017. The mean RT for the high-long location was significantly longer than that for the high-short location, β = 51.700, SE = 13.440, t = 3.848, p = .001.

Table 2 Tests of the fixed effects in Experiment 1

Full size table

More importantly, the interaction between interval and target location was also significant. Across trials with the 500-ms interval, mean RT was significantly shorter when the target appeared at the high-short location than when it appeared at any one of the low-probability target locations, β = −81.000, SE = 8.690, t = −9.313, p < .001, implying that the temporally congruent location was prioritized compared with the temporally neutral locations at the short interval. In contrast, at the 500-ms interval, there was no significant difference between the temporally incongruent high-long location and the low-probability locations, β = −15.400, SE = 11.840, t = −1.299, p = .399. Also, the mean RT was significantly longer when the target appeared at the temporally incongruent high-long location than when it appeared at the temporally congruent high-short location, β = 65.600, SE = 14.980, t = 4.378, p < .001. In short, there was a response benefit for the temporally congruent high-short location compared with all other locations in the short interval condition.

Across trials with the 1,500-ms interval, participants responded faster when the target appeared at the high-long location relative to the neutral low-probability locations, β = −26.300, SE = 9.360, t = −2.811, p = .020, suggesting a behavioral advantage for the temporally congruent location than the neutral locations at the long interval. Interestingly, the mean RT for the target appearing at the high-short location was also significantly shorter than that for the target appearing at any one of the neutral low-probability locations, β = −60.300, SE = 11.280, t = −5.347, p < .001. The mean RT for the high-long location was not significantly different from that for the high-short location at the 1,500-ms interval, β = 34.000, SE = 14.930, t = 2.277, p = .067. That is, after the long interval, there was a response benefit for both high-probability locations compared with the neutral locations.

Accuracy

Figure 2B shows the mean ACC as a function of interval (500 ms, 1,500 ms) and target location (high-short, high-long, low). The GLMM analyses of ACCs only revealed a significant main effect of target location (Table 2). Compared with the mean ACC for the low-probability target locations, the performance for the high-short location was significantly more accurate, β = 0.394, SE = 0.080, z = 4.947, p < .001. There was no significant difference between the high-long location and the low-probability locations, β = 0.108, SE = 0.063, z = 1.712, p = .201. Participants responded less accurately when the target appeared at the high-long location than when it appeared at the high-short location, β = −0.286, SE = 0.088, z = −3.251, p = .003.

Awareness assessment

In Experiment 1, a certain number of participants showed awareness for the spatial distribution of targets. Among the 41 participants, 25 participants reported both high-probability target locations correctly. Fourteen participants reported only one high-probability location and two participants reported none of the high-probability locations. There were few participants who showed awareness regarding the spatiotemporal regularities, as only 11 participants reported the associations between the intervals and high-probability locations correctly. Sixteen participants reported none of the spatiotemporal associations. Fourteen participants reported the association between either the short interval (11 participants) or long interval (three participants) with the high-probability target location.

We investigated whether participants’ awareness about the regularities had a significant effect on their performance in the (G)LMM analyses. As shown in Table 2, LMM revealed only a significant effect on RT of the awareness of the high-probability location after the long interval. Participants who were aware of the high-probability target location after the long interval responded faster than those who were unaware of it, β = −86.100, SE = 38.700, t = −2.225, p = .032. In the GLMMs, there was no significant effect on ACC regarding the awareness of regularities.

Furthermore, to determine whether participants’ levels of awareness affected the learning of spatiotemporal regularities, we rated participants’ overall awareness by combing their awareness for the spatial and spatiotemporal regularities. We categorized those participants who correctly reported either the spatial or spatiotemporal regularities as “aware”, while others as “unaware”. Apart from the control factors of target location priming and physical target position reported above, the (G)LMMs included the fixed effects of interval (500 ms, 1,500 ms), target location (high-short, high-long, low), awareness (aware, unaware), and the three-way interaction. For the LMMs, the random effect structure included a by-participant random intercept and a by-participant random slope for target location and interval. For the GLMMs, the random effect structure included a by-participant random intercept and a by-participant random slope for target location and awareness. The results showed that there was no significant interaction between interval, target location and awareness on RT, χ²(2) = 1.032, p = .597, or on ACC, χ²(2) = 4.994, p = .082. Therefore, there is no relationship between the level of awareness of the regularities and the effect of those regularities on behavior.

Intertrial location priming

As the current findings might also be attributable to short-lasting intertrial location priming (Maljkovic & Nakayama, 1994), we also included “target-location priming” (intertrial target repetition versus alternation) as a fixed effect in the (G)LMMs. The influence of target location priming was significant (Table 2). Participants responded slower, β = 35.800, SE = 3.420, t = 10.486, p < .001, and less accurately, β = −0.097, SE = 0.048, t = −2.016, p = .044, when the target location was not repeated than when it was repeated from the previous trial. However, after considering the effect of target location priming, the behavioral expression of spatiotemporal regularities remained. Thus, the observed effects cannot be explained by intertrial location priming.

In sum, the results of RT revealed an asymmetric time-specific behavioral benefit between two high-probability locations at two intervals. Specifically, there was a response benefit for the temporally congruent high-short location as compared with all other locations at the short interval, while there was a response benefit for both high-probability locations compared with the neutral locations at the long interval. The effects cannot be explained by the participants’ awareness of the regularities or intertrial location priming.

The pure effect of spatiotemporal regularities

According to the hypothesis of spatiotemporal modulation of attention, we expected a symmetric temporal benefit at both intervals. That is, we would expect an increase of RT from the 500-ms interval condition to the 1,500-ms interval condition when the target appeared at the high-short location. As we used a uniform distribution of intervals in this experiment, this expectation fails to take into account the counteracting effect of general temporal preparation. In fact, the decrease of RT from the 500-ms interval to the 1,500-ms interval is clearly visible for neutral low-probability locations, β = 20.117, SE = 7.360, t = 2.734, p = .008 (Fig. 2A). When the target appeared at the high-long location, there was also a reduction of RT from the nonassociated 500-ms interval condition to the associated 1,500-ms interval condition, β = 31.044, SE = 9.490, t = 3.271, p = .001. By contrast, when the target appeared at the high-short location, there was no significant difference between the associated 500-ms interval condition and the nonassociated 1,500-ms interval condition, β = −0.547, SE = 9.380, t = −0.058, p = .954. Therefore, the expected increase in RT as a function of interval for the high-short location because of the spatiotemporal regularities might be cancelled out by a concurrent influence of temporal preparation working in the opposite direction.

To assess the pure effect of the spatiotemporal regularities independent of the general temporal preparation, we expressed the effect of interval for the high-short and high-long locations relative to the baseline of this effect for the low-probability locations. In particular, we subtracted out the mean effect of interval (500 ms minus 1,500 ms) obtained at the low-probability locations from the corresponding effect obtained at the high-short location and the high-long location for each participant. Then the effect of interval at these two locations was compared against zero in one-tailed one-sample t tests (Fig. 3).

As shown in Fig. 3A, relative to the baseline, the effect of interval on RT was negative for the high-short condition in Experiment 1, t(40) = −1.915, p = .031, d = −0.299. That is, after baseline correction, participants responded faster when the target appeared at the high-short location after the associated short interval than when it appeared at this location after the nonassociated long interval. In comparison, the baseline-corrected effect of interval for the high-long condition was not significant in Experiment 1, but it tended to be positive numerically, t(40) = 1.310, p = .099, d = 0.205. These results indicate that spatiotemporal dynamics influenced performance in addition to general temporal preparation.

Discussion

Experiment 1 showed that participants learned the spatial distribution of targets at two high-probability locations. Participants responded faster for both the high-short and the high-long location relative to low-probability locations. In addition, there was also evidence for a modest temporal modulation of this effect. Specifically, after the short interval, there was a larger response benefit when the target appeared at the temporally congruent high-short location than when it appeared at any one of the low-probability locations. After the long interval, there was a similar advantage for the temporally congruent high-long location. This finding is consistent with the idea that attention dynamically shifts toward the high-probability location where the target is most likely to appear.

However, this interpretation is complicated by the further finding that, after the long interval, the high-short location continues to have an advantage relative to low-probability locations. This finding seems to suggest that the high-short location continues to be prioritized at the long interval, whereas the high-long location was not yet prioritized at the short interval. Nevertheless, after taking the obscuring influence of general temporal preparation into account, a more symmetrical picture arises. As shown in Fig. 3, in Experiment 1, relative to the baseline of the low-priority location, a long temporal interval leads to a decreased focus of attention at the high-short location and a somewhat increased focus of attention at the high-long location. In Experiments 2 and 3, we further explored the stability of this data pattern under different distributions of intervals. In the general discussion section, we will more deeply consider the implications for the dynamics of attention across space.

In sum, by using the singleton task, we aimed to examine the dynamics of spatial attention separated from motor response learning, which may have contributed to the dynamics in earlier studies (e.g., Boettcher et al., 2022; Wagener & Hoffmann, 2010). The results of Experiment 1 showed that participants are able to implicitly extract the spatiotemporal regularities in the stimulus displays and let those guide attentional search accordingly. This dynamic effect was small, however, and may have been distorted by general temporal preparation, which motivated Experiments 2 and Experiments 3.

Experiment 2

As revealed in Experiment 1, the effect of spatiotemporal modulation of attention by spatiotemporal regularities might be obscured by a concurrent influence of temporal preparation under a uniform interval distribution. To investigate the pure effect of spatiotemporal regularities of target on visual selection, we adopted the same spatiotemporal regularities in Experiment 2 while employing an exponential distribution of two interval (2:1) to control the effect of general temporal preparation. That is, we made the trials with the 500-ms interval twice as frequent as the trials with the 1,500-ms interval. We expected that the exponential distribution of intervals would result in an approximately flat RT–interval function for neutral locations (cf. Los et al., 2017; Los et al., 2021; Näätänen, 1971; Vangkilde et al., 2013). Furthermore, we expected that, relative to this baseline, the effect of spatiotemporal orienting of visual attention would show a symmetric benefit pattern.