Neural dynamics of context-sensitive adjustments in cognitive flexibility

Abstract

To adaptively interact with the uncertainties of daily life, we must match our level of cognitive flexibility to situations that place different demands on our ability to focus on the current task while remaining sensitive to cues that signal other, more urgent tasks. Such cognitive-flexibility adjustments in response to changing contextual demands (meta-flexibility) have been observed in cued task-switching paradigms, where the performance cost incurred by switching versus repeating tasks (switch cost) scales inversely with the proportion of switches (PS) within a block of trials. However, the neural underpinnings of these adjustments in cognitive flexibility are not well understood. Here, we recorded 64-channel EEG measures of electrical brain activity as participants switched between letter and digit categorization tasks in varying PS contexts, from which we extracted event-related potentials (ERPs) elicited by the task cue and EEG alpha-power differences during the cue-to-target interval and the resting pre-cue period. The temporal resolution of EEG/ERPs allowed us to test whether contextual adjustments in cognitive flexibility are mediated by tonic changes in processing mode, or by changes in phasic, task-cue-triggered processes. We observed reliable modulation of behavioral switch cost by PS context that were mirrored in both cue-evoked ERP and time-frequency effects, but not in blockwide pre-cue EEG changes. These results indicate that different levels of cognitive flexibility are instantiated in response to the presentation of task cues, rather than by being maintained as a tonic state throughout low- or high-switch contexts.

Introduction

Daily life confronts us with ever-changing situations that pose different demands on our ability to focus strongly on an ongoing task, on the one hand, or to be able to shift our focus away from the task at hand due to other urgent matters arising (Dreisbach & Fröber, 2019; Braem & Egner, 2018; Goschke, 2013). We here refer to the ease with which someone switches from one task to another as switch-readiness or cognitive flexibility. Importantly, different contexts require different levels of cognitive flexibility. For example, while cooking a meal, frequent switching between different tasks is desirable, such as moving from chopping parsley to adding pasta to the boiling water. Conversely, when writing a paper, a low level of cognitive flexibility is helpful to avoid being side-tracked by alternative activities. Accordingly, the brain needs to find, and to continually adjust, a contextually optimal level of flexibility, an ability that has been referred to as meta-flexibility (Dreisbach & Fröber, 2019; Braem & Egner, 2018; Goschke, 2013; Hommel, 2015). The present study investigates the mechanisms underlying this ability to change one’s level of switch readiness to suit varying contexts.

We investigate meta-flexibility through the lens of cued task-switching (for reviews, see Kiesel et al., 2010; Koch et al., 2018; Monsell, 2003; Vandierendonck et al., 2010). In such a paradigm participants are cued on each trial which one out of (typically) two tasks to perform, and the canonical finding is that switching to a different task, compared to repeating the same task, produces slower and more error-prone responses, referred to as the switch cost (Kiesel et al., 2010; Meiran, 2000; Monsell, 2003; Vandierendonck et al., 2010). A smaller switch cost indicates that switch processes can be carried out with greater ease, or – in other words – with greater cognitive flexibility (switch-readiness); vice versa, when switch costs are larger, switch-readiness is considered low (and stability high).

Previous behavioral studies have probed whether people strategically adapt their level of flexibility to changing contexts by varying the frequency of switch trials between blocks of trials. Here, blocks with a high proportion of switches (high-PS condition) have been found to be associated with smaller mean switch costs than blocks with a low proportion of switches (low-PS condition) (Bonnin et al., 2011; Dreisbach et al., 2002; Dreisbach & Haider, 2006; Duthoo et al., 2012; Mayr, 2006; Monsell & Mizon, 2006; Siqi-Liu & Egner, 2020). Participants thus demonstrate meta-flexibility by adopting more flexible behavior in contexts that require more frequent task switching. This inverse relationship between switch cost and PS in a given temporal context is referred to as the list-wide proportion switch (LWPS) effect. Though the LWPS effect has been reliably demonstrated behaviorally, relatively little is understood about the control processes that enable meta-flexibility. Theoretically, there are two rival accounts of how the behavioral LWPS effect could come about:

First, being exposed to an extended period of frequent task switches in a high-PS block could make people adopt a sustained or tonic mode of greater flexibility, which is maintained throughout the entire block. Such list-wide adjustments of control state would be consistent with behavioral evidence from conflict-control experiments, which utilize conceptually similar manipulations of the block-wise proportion of congruent versus incongruent trials in conflict tasks (e.g. Stroop task). For example, Bugg & Chanani (2011) found that conflict-control adjustments were implemented at a list-wide level and were maintained throughout the temporal context of a block, rather than varied on a trial-by-trial, stimulus-specific manner. Thus, one plausible possibility is that PS-dependent switch-readiness adjustments are implemented in a similar manner, as a change in sustained processing mode over a block of trials.

Alternatively, the switch-readiness adjustments observed in the LWPS effect could arise from phasic or transient instantiation of different levels of switch-readiness in response to specific task cues or task sets that are associated with low or high switch likelihood in low versus high-PS blocks, respectively. Our group recently reported behavioral evidence of such task-specific adjustments in cognitive flexibility (Siqi-Liu & Egner (2020, Exp 3a-b). Specifically, we employed a three-tasks cued-switching paradigm wherein two “biased” tasks drove overall list-wide switch proportions (being presented more frequently as switch trials in the high-PS blocks and as repeat trials in the low-PS blocks), while one unbiased “transfer” task was equally often presented as switch and as repeat trials in both block types (cf. Bugg & Chanani, 2011). While switch cost reductions in high-PS blocks compared to low-PS blocks were identified in the biased tasks, they were not found in the unbiased task. This observation runs counter to the idea of tonic state adjustments, which would predict the transfer of switch-readiness adjustments to the unbiased task set. Rather, these data suggest that the cognitive processes mediating the LWPS effect were instantiated only following the presentation of the relevant cues for task sets that were more often associated with switches or repetitions in a given context. Studies using electroencephalographic (EEG) measures have shown that cue-evoked event-related potentials (ERPs) (Jost et al., 2008; Karayanidis et al., 2011; Lavric et al., 2008; Nicholson et al., 2006; Wong et al., 2018) and time-frequency effects in theta, alpha, and beta bands (Cooper et al., 2015, 2017, 2019; McKewen et al., 2020) differentiate between switch- and repeat- trials, constituting a “neural switch cost.” More specifically, switch cues have been associated with more positive centroparietal ERPs occurring between 200-900 ms compared to repeat cues, and effect referred to as a switch-related positivity (Barceló & Cooper, 2018; Han et al., 2018; Jost et al., 2008; Karayanidis et al., 2010, 2011; Nicholson et al., 2006; Wong et al., 2018). These neural signatures, particularly ERP correlates of switch-related activity, have been shown to be modulated by advance task-preparation during the cue-to-target interval (CTI) following exposure to an informative cue (for reviews, see De Baene & Brass, 2013; Karayanidis et al., 2010). Since ERP signatures of switch-related activity are sensitive to task preparation, it is also likely that they can be used to identify when and how strategic adaptations to context are implemented.

To the best of our knowledge, only two prior studies assessed neural signatures related to block-wise PS manipulations, one employing ERPs (Nessler, 2012), and one using functional magnetic resonance imaging (fMRI, De Baene & Brass, 2013). The fMRI study addressed the question of which brain regions’ neural switch cost activity profile may be modulated by PS context and found that a subset of regions of the frontoparietal network displayed such sensitivity (De Baene & Brass, 2013). The authors argued that the often-inconsistent findings from studies that attempt to identify brain regions underlying preparatory control may be due to the variability in switch probability across different studies. They demonstrated that only a subset of brain regions showed switch-related preparatory activity in low (30%) PS blocks were also activated in high (50%) PS blocks. However, this data provides only limited insight into how adaptations in switch-readiness are achieved in terms of how context influences control recruitment during the task-preparation process, and the time-resolution limits of fMRI meant that the study could not access whether control adjustments were accomplished in a tonic or phasic manner.

Calculating ERPs from midline electrodes, Nessler et al (2012) compared early and late switch-related parietal positivities in blocks with frequent (50% PS) versus infrequent (25% PS) switches and compared to single task blocks, utilizing either informative or uninformative cues. The uninformative cues condition was included because previous studies have demonstrated that PS had little effect on performance when participants had ample time to prepare for the upcoming target stimulus (i.e. when CTI is long [>790 ms], as it was in Nessler et al, 2012). The authors focused on comparing “general switch costs,” or the behavioral and neural differences between performance on repeat trials in mixed-task versus single-task blocks, and “specific switch costs” which refer to the difference between switch and repeat trials within mixed blocks. The investigation of specific switch costs is more aligned with the goals of the current study. While Nessler et al. (2012) made important theoretical observations in regards to general switch costs, their paradigm was not optimized to elicit the LWPS effect for specific switch costs in mixed task blocks. They only found an LWPS effect when they used uninformative cues, but, even in the uninformative cue condition, no main or interaction ERP effects of PS context, or any switch-related effects during the CTI, were identified.

The current study differs from Nessler et al.’s (2012) design in several ways. First, we created experimental conditions that incentivized participants to utilize information from both the task cue and the block-wide switch proportion context. We did so by jittering target-onset relative to an informative task cue, creating a variable CTI (190 – 500 ms). Since participants did not know how much time they had to prepare for the upcoming task on the target stimulus, it would have been of value to prepare for behavioral adjustments according to block-wide switch context for optimal performance. Utilizing both an informative task cue and a CTI that should elicit behavioral LWPS effects (based on Monsell and Mizon, 2006 and Siqi-Liu and Egner, 2020), our design was set up to investigate the novel question of how PS context modulates switch-cue versus repeat-cue processing and thus allowed us to identify the timing and characteristics of cognitive processes that contribute to meta-flexibility. Second, we employed a more data-driven approach compared to Nessler et al. (2012), conducting nonparametric analyses on data from 64 channels across the scalp rather than focusing only on midline electrodes, in order to identify potentially new neural activation patterns that underly context-sensitive flexibility adaptations. These new design features make the current study more suitable for investigating the LWPS effect.

We utilized the temporal resolution of ERPs and time frequency analyses to examine whether a tonic or a phasic account of switch-readiness adjustments more accurately describes the cascade of cognitive and neural processes that lead to the behavioral LWPS effect. We collected EEG data during a letter-digit task switching paradigm with a block-wise PS manipulation which has been shown to produce highly reliable LWPS effects (Bejjani et al., 2021; Siqi-Liu & Egner, 2020), with several minor modifications for EEG recording.

We conducted two types of analyses. First, we carried out cue-locked ERP and oscillatory EEG analyses, focused on the time between the onset of the task cue specifying the task to be performed on the upcoming task stimulus, and the onset of the task stimulus (i.e., the CTI). This allowed us to cleanly isolate activity relevant to cue-triggered preparation of the upcoming task. The tonic flexibility account would predict that PS context modulation of cue-evoked ERP amplitude differences between switch- and repeat-trials would already be evident at cue onset. By contrast, the phasic flexibility account would predict that contextual modulation of switch-related ERPs emerge only after cue onset, since the task cue itself serves as the trigger for the cognitive processes implementing the modulation.

Second, we also investigated whether sustained control state differences between switch contexts could be identified in non-phaselocked time-frequency analyses. We analyzed oscillatory EEG power during the CTI and during the fixation period before cue-onset, primarily focusing on activity in the alpha band (8-14 Hz), given that alpha power decreases (“alpha suppression”) are commonly associated with the recruitment of attention resources to concentrate on a task (Foster et al., 2017). In a task switching context, it is possible that entering into a more stable processing mode could elicit greater alpha suppression, reflecting greater on-task concentration. Context-dependent differences in alpha power prior to the onset of the task cue would provide evidence for strategic adjustments in tonic (i.e., blockwise) attentional states to match environmental demands. In contrast, a lack of lasting differences in alpha power between high and low PS contexts would support the hypothesis that switch-readiness was not being implemented tonically.

Experimental Methods

Participants

Thirty-five undergraduate students who reported no underlying medical conditions were recruited from the Duke University Department of Psychology and Neuroscience Subject Pool. Five participants were excluded from analysis due to lower than 70% accurate task performance (1), technical or recording difficulties (3), or noisy data (1), leaving a final sample size of 30 (14 male, 16 female; mean age of 22, ranging from 18-29 years, with a standard deviation of ~3 years). All participants gave informed consent and received course credit in accordance to a protocol approved by the Duke University Institutional Review Board. All had normal or corrected-to-normal vision.

Stimuli

Task stimuli consisted of a letter and a digit displayed simultaneously in 12 pt font on opposite sides of a fixation point at the center of the screen for each trial. The letter was randomly selected from ‘A’, ‘E’, ‘I’, ‘U’, ‘G’, ‘K’, ‘M’, or ‘R’ and the digit was randomly selected from ‘2’, ‘3’, ‘4’, ‘5’, ‘6’, ‘7’, ‘8’, or ‘9’. Whether the letter or the digit was presented on the left or right side of fixation was randomized across trials.

Procedure

The behavioral paradigm was programmed in Presentation from Neurobehavioral Systems (https://www.neurobs.com/). Each trial consisted of an inter-trial interval (ITI) of 1100-1410 ms (jittered by 10 ms bins drawn from a uniform distribution), followed by a cue display lasting 150 ms, followed by a delay of 40-350 ms (jittered by 10 ms bins matched to the ITI, such that the ITI and post-cue delay summed to 1450 ms on each trial), followed by a target stimulus presented for 1200 ms, which was followed by a feedback screen lasting 300 ms. The CTI thus varied from 190-500 ms. The fixation dot remained at the center of the screen throughout the run (Figure 1).

Figure 1.

Experimental protocol. The trial structure involved an ITI, cue display, a delay, a target stimulus, and response feedback. In each block, 30% of trials were catch trials where the cue was not followed by a stimulus display and no response was required. CTI is marked as the sum of the cue-display interval and the jittered interval between cue-offset and target onset.

We utilized a relatively short CTI (190-500 ms), although at the risk of significant overlap between cue- and target-evoked waveforms, because prior research (Siqi-Liu & Egner, 2020; Monsell & Mizon, 2006) did not find context-sensitive switch-cost adjustments at longer CTIs. This is likely because participants relied less on list-wide context when they have ample time to reconfigure for the upcoming task following the cue. Two measures were implemented to minimize the effect of potential cue-target overlap. First, we introduced a jitter in the delay period between the cue and target stimulin in order to decorrelate cue and target onsets. The length of the ITI was also jittered to complement the delay period, thus ensuring that the response-stimulus interval, or the summed length of the ITI, cue, and delay display, was always 1600 ms. In other words, a trial with a longer delay period (and CTI) would include a correspondingly shorter ITI. The cue display period was also kept constant so that participants did not receive longer exposure to the cue on longer CTI trials. Second, to ensure that we could observe cue-evoked activity later than 190 ms (the shortest possible CTI) without overlapping target-related activity, 30% of the trials were selected to be catch trials where no target display followed the cue, and no response was required (c.f., Grent-’t-Jong & Woldorff, 2007).

Participants were required to perform a letter classification task (“Is the letter a vowel or consonant?”) if they saw the cues “letter” or “abcd”, and to perform a digit classification task (“Is the digit odd or even?”) if they saw the cues “digit” or “number.” They were informed that, on some trials, the cue will not be followed by a target stimulus (catch trials), in which case no response was required. The 2:1 cue-to-task mapping allowed us to change the cue on every trial, regardless of whether the task was switched or repeated, thus ensuring that any observed neural and behavioral differences between task switch and repeat trials would not be attributable to cue repetition effects (Logan & Bundesen, 2003; Mayr & Kliegl, 2003). Participants had to press the ‘d’ or ‘k’ key to categorize the stimuli as vowel/consonant or odd/even. Participants were randomly assigned to different response mappings for each task. The words “correct”, “incorrect”, and “no target” were displayed during the feedback interval on correct, incorrect, and catch trials, respectively. Responses made while the task stimulus was no longer onscreen were considered incorrect.

Each participant completed 16 blocks of 61 trials each. The PS context of each block was 30% or 70%, and there were eight blocks of each switch proportion condition. The trial sequence for each block was generated pseudo-randomly according to an algorithm that ensured each task was presented an approximately equal number of times. In every block, 30% of the trials were randomly selected as catch trials (i.e., cue only with no target following). Accordingly, ~30% of the repeat and switch trials respectively in PS30 and PS70 blocks did not include target stimulus displays and did not require responses. Blocks of the same PS level were presented consecutively in groups of four blocks each to increase the saliency of the switch/repeat context (e.g., “AAAA BBBB AAAA BBBB”). Participants were randomly assigned to block sequences with PS30 or PS70 blocks presented first. Post-hoc analyses found no significant two- or three- way interactions between block order (PS30 first versus PS70 first), task sequence, and PS context, suggesting that it is unlikely that the order of PS context presentation affected switch costs or the LWPS effects in the current dataset. Before beginning the main experiment, participants completed one practice block with 50% switch trials and with catch trials occurring 30% of the time. Participants were instructed to maintain fixation on the central fixation dot throughout the experiment. Trials were coded as being either task switch trials (following a different task on the previous trial) or task repeat trials (following the same task).

Behavioral Analyses

We employed a 2 (task sequence: switch v. repeat) x 2 (PS: 30 v. 70) repeated-measures analysis of variance (rmANOVA) to assess effects on response times (RTs) and error rates. Significant interaction effects were followed up using dependent-samples t-tests. The first trial of each block and catch trials were excluded for both analyses (given that there was no target and no behavioral response on such trials), and trials with incorrect or no responses were excluded from RT analysis.

EEG Acquisition and Preprocessing

Participants were seated approximately 60 cm from a 24-inch monitor in a dimly lit, electrically shielded, and sound-insulated room. EEG data were recorded from a 64-channel custom, extended-coverage cap (Woldorff et al., 2002) using active electrodes (actiCAP, Brain Products GmbH, Gilching, Germany). Electrodes were kept at impedences of <15 kΩ and referenced to the right mastoid during recording. The recording was obtained at a sampling rate of 500 Hz using a three-staged cascaded integrator-comb anti-aliasing filter with a corner frequency of 130 Hz (actiCHamp, Brain Vision LLC, Cary, NC, USA).

Data were pre-processed in MATLAB using the EEGLab toolbox (Delorme & Makeig, 2004). Offline EEG data were low-pass filtered at 30 Hz using EEGLAB’s Hamming-windowed sinc finite infinite response (FIR) filter (pop_eegfiltnew), resampled to 250 Hz, high-pass filtered at .05 Hz also using an FIR filter, and subsequently re-referenced to the algebraic average of the left and right mastoids. Upon visual inspection of the data, noisy channels were interpolated using EEGLab’s spherical spline interpolation (Perrin et al., 1989). A copy of the EEG data that was filtered more strongly with a high pass filter of 1 Hz was submitted to independent component analysis (ICA), and the resulting weights were then applied to the original 0.1 Hz dataset. High pass filtering of activity below 1-2 Hz to reduce low-frequency noise before the ICA calculations has been found to generally improve ICA performance and produce more clearly isolated components (Winkler et al., 2015; Klug & Gramann, 2020). Components reflecting eye movements and heartbeat were identified and removed via ICA.

In both the ERP and time frequeny analyses, epochs were tagged for artifact rejection if activity between −500 to 1500 ms was greater than a participant-specific absolute threshold, which depended on variations in overall noise in the dataset. The smallest threshold for the artifact rejection was +− 75 μV and the largest was +− 95 μV.

ERP Analyses

For the ERP analyses, data were epoched from −500 to 1500 ms time-locked to cue presentation and baseline corrected from −300 to 0 ms. Epochs were binned into four conditions reflecting the cells of the 2 x 2 (task sequence x switch proportion) design, and averaged by condition to produce cue-locked ERPs. Only catch trials were included for the first round of analyses so that both early and late cue-related effects could be captured without potential contamination by target processing. Any early effects found among catch trials were then replicated in follow-up analyses that included both catch trials and non-catch trials that had targets occurring after the identified ERP differences in order to confirm the early effects with increased power. These additional late target trials were included in order to maximize trial count and increase power of any observed early effects.

Using the Fieldtrip toolbox (Oostenveld et al., 2010), we conducted dependent-samples two-tailed t-tests at each 4-ms time point between 0 ms to 700 ms (at the 250 Hz resample rate) and across all channels, corrected for multiple comparisons with nonparametric cluster-based Monte Carlo permutations (10,000 repetitions). Clusters were defined via the triangulation method, and samples with significant t values with less than 2 significant neighboring channels were discarded. To explore switch-related ERP activity, we tested the contrast between switch – repeat trials. To explore the potential interactions between task sequence and PS context, we calculated separate switch – repeat difference waves for PS30 and PS70 contexts, and compared the difference of differences (i.e., PS30_{switch-repeat} – PS70_{switch-repeat}). To further explore drivers of interaction effects, we conducted separate analyses of the effect of PS on switch trials (i.e. PS30_switch – PS70_switch) and the effect of PS on repeat trials (i.e. PS30_repeat – PS70_repeat).

Note that we did not constrain the interaction analysis based on channel locations and timepoints that were significant in the switch – repeat contrast. Rather, all timepoints from 0 ms to 700 ms across all channels were also tested for the interaction effect of task sequence and PS context. This approach allows us to uncover potential clusters where switch-related activity was moderated by PS context in the absence of a main effect of switching.

Time Frequency Analyses

For time frequency analysis, the data were epoched from −1500 to 1500 ms time-locked to cue presentation, without baseline correction. Frequency decomposition was performed using Fieldtrip’s multitaper method, where power was estimated using discrete prolate Slepian sequences in logarithmically spaced frequencies from 2 to 30 Hz. The window widths for the tapers were 2 cycles for 2-4 Hz, 3 cycles for 4-7 Hz, 5 cycles for 8-14 Hz (alpha band), 7 cycles for 15-20 Hz, and 10 cycles for 21-30 Hz. Multitaper smoothing was specified as 5 x log₁₀ of each frequency. The event-related power spectra, time-locked to the cue, for each participant were then binned and averaged according to the four conditions resulting from the 2 x 2 interaction of the task sequence and PS factors. Condition-averaged cue-locked ERPs were subsequently subtracted from the power spectra in order to focus on the effects on the cue-induced oscillatory activity, or non-phase-locked power.

We conducted dependent-samples two-tailed t-tests at each 4-ms time point between 0 ms to 1000 ms and across all channels, averaging across 8-14 Hz (alpha-band power), correcting for multiple comparisons with nonparametric cluster-based Monte Carlo permutations (10,000 repetitions) implemented via the Fieldtrip toolbox. Clusters were defined using the triangulation method, and samples with significant t values with less than 2 significant neighboring channels were discarded.

We conducted these analyses timelocked to the trial-onsets as wells as timelocked to the cues. For the trial-onset-locked analyses, we examined differences in alpha power between the PS30 and PS70 block-wide conditions. Trial onset-locked analyses were tailored to identify sustained modulations of attentional state (i.e., blockwise). Thus, we focused on the fixation period after trial-onset but before cue-onset in order to isolate potential baseline attentional differences between the block types unrelated to cue or stimulus processing. In other words, in the trial-onset-locked analyses, the fixation point functioned as a neutral cue (uninformative of the upcoming task identity) that simply signaled the start of a trial. For the cue-locked analyses, which were aimed at assessing oscillatory EEG reflections of phasic modulations of switch readiness triggered by the cues, we investigated the same contrasts as in the ERP analysis, namely switch-related effects (switch – repeat) and the interaction between task sequence and PS context (PS30_{switch-repeat} – PS70_{switch-repeat}). All cue-locked analyses were done on catch trials only. As with ERPs, the time-frequency interaction analyses were not constrained to timepoints and channels that were significant in the main effect of task sequence.

We also performed nonparametric cluster-based permutation analyses on theta (4-8 Hz) and beta (15-30 Hz) bands, but no significant main or interaction effects were observed, and thus the results are not reported in this manuscript.

Results

Behavioral Results

Descriptive and inferential statistics on the behavioral data are presented in Tables 1 and 2, and summary data are displayed in Figure 2. The average RT across subjects on the task was 719 ms, and average error rate was 14%. Switch costs are calculated as switch – repeat trial RTs and error rates. The main effect of task sequence was significant for both RTs (p < .001) and error rates (p < .001) due to slower RTs and higher error rates on switch trials (M_RT = 736 ms, M_error = ..16) compared to repeat trials (M_RT = 711 ms, M_error = .10), reflecting the classic switch cost. RTs were faster in general on PS30 trials (M = 717 ms) compared to PS70 trials (M = 729 ms), reflecting a main effect of PS (p = .024). Most crucially, the interaction effect of task sequence x PS was significant for both RTs (p = .002) and error rates (p = .006), due to greater switch costs (M_RT = 34 ms, M_error = .064) in the PS30 condition than in the PS70 condition (M_RT = 15 ms, M_error = .039), replicating the LWPS effect. Of note, this interaction was driven primarily by switch context affecting performance on task-repeat trials, which were reliably slower and less accurate in the high-PS than in the low-PS context (RT: F(1,29) = 17.22, p < .001, η_p² = .37; error rate: F(1,29) = 7.91, p = .009, η_p² = .21), whereas switch-trial performance did not differ significantly between the two contexts (RT: F(1,29) = .15, p = .70; error rate: F(1,29) = .21, p = .65). In sum, these behavioral results replicate context-sensitive adjustments in switch processes as has been observed in the prior literature (e.g., Siqi-Liu & Egner, 2020), thus setting the stage for assessing the neural processes underpinning these adjustments in the EEG data.

Table 1.

Mean Response Times (ms) and Error (percentage) as a function of the proportion of switch trials (PS) and task sequence

	PS30	PS70
Switch	734.3 / 15.8	736.7 / 15.4
Repeat	699.8 / 9.4	721.7 /11.5

Note. Data refers to group mean RTs on the left (excluding error trials and catch trials) and percentage error on the right (excluding catch trials).

Table 2.

Inferential Statistics

RT Effects	df	F	ηp2	p-value
Task Sequence (TS)	1, 29	27.30***	.49	<.001
Proportion Switch (PS)	1, 29	5.66*	.16	.024
TS x PS	1, 29	12.04**	.29	.002
Accuracy Effects	df	F	ηp2	p-value
Task Sequence (TS)	1, 29	39.01***	.57	<.001
Proportion Switch (PS)	1, 29	1.16	.39	.290
TS x PS	1, 29	8.86**	.23	.006

Figure 2.

Behavioral results. Upper panels depict RT (left) and error rate (right) group means for repeat (teal) and switch (orange) trials with within-subject error bars indicating 95% confidence intervals (1.96 x standard error). Lower panels depict switch costs for RTs (left) and error rates (right), which are calculated as switch – repeat. Box plots are overlaid above dots representing individual mean switch costs. Switch costs for both RT and error rates are higher in the PS30 condition compared to the PS70 condition.

ERPs

Task Sequence Effects

Cluster-corrected dependent-sample t-tests on the switch versus repeat cue-evoked ERPs on the catch trials revealed one cluster with a significant positive-polarity ERP deflection (p = 0.015), occurring between approximately 112-172 ms, and one cluster with a significant negative-polarity deflection (p < .001), occurring between 380-596 ms.

Visual examination of the positive cluster’s topography revealed that this early switch-related effects was centered around frontal-central channels (Figure 3A), reflecting a larger positive-polarity waveform elicited on switch compared to repeat trials. To corroborate this early switch-related positivity with an analysis with higher trial counts, we conducted the same analysis after including trials with targets occurring later than 280 ms (i.e. well after significant activity was resolved). A similar positive cluster was observed in this larger trial-count analysis (p < .001), again exhibiting frontal-central topography and with significant differences arising between approximately 104-196 ms (Figure 3B). These data document that differential neural processing on task switch compared to repeat trials is already apparent starting at ~100 ms after cue onset, and independently of physical cue changes (given that cue words changed on every trial), thus likely reflecting the rapid, cue-triggered initiation of task-set reconfiguration processes on switch trials.

Figure 3.

Cue-Locked Early Effects of Task Sequence. (Top Left) Switch, repeat, and switch minus repeat ERP activity over frontal-central electrodes (highlighted in red) for catch trials only, following cue onset.(Bottom Left) topography of a significant positive cluster (p = 0.015) in frontal central electrodes revealed by a nonparametric cluster-based Monte Carlo permutation test of with 10,000 repetitions, unconstrained in time and electrode locations. Only electrodes that remain significant throughout the entirety of each time bin (e.g. 120 to 160 ms) are highlighted as an additional control for noise. Significant activity began ~ 112 ms and lasted until ~172 ms. (Top Right) Switch, repeat, and switch minus repeat ERP activity over frontal-central electrodes (highlighted in red) for catch trials and late target trials (post 240 ms). To maximize trial counts and increase power in this analysis, we also included non-catch trials in which the targets occurred after the significant difference was identified. (Bottom Right) The significant positive cluster (p < .001) on this higher-trial-count analysis again exhibited similar topography and significant activity, lasting from ~104 to ~196 ms. Again, only channels that remained significant throughout the entirety of each time bin are highlighted in white on the topography plots.

The longer-latency, negative cluster was characterized by lower amplitudes of ERP waveforms in switch compared to repeat trials, occurring at central-parietal electrodes (Figure 4). In line with the above interpretation for the early positive effect, this later effect might index late-stage task set reconfiguration processes in anticipation of the imminent target onset. However, a clean interpretation of this effect, observed in catch trials, is complicated by the fact that it reaches significance around or even after the time window that the participant would expect the target to appear in non-catch trials. It may therefore be contaminated by participants terminating the process of task-set reconfiguration as they realized that no target was going to occur on that trial. No follow-up analyses that include non-catch trials with later targets were run since this negative cluster would overlap with the presentation of even later targets on those trials.

Figure 4.

Cue-Locked Longer-latency Effects of Task Sequence. Switch-minus-repeat activity in catch trials only, following cue onset. Top: ERPs over central parietal sites, corresponding to the negative cluster. Bottom: topography of negative cluster. Cluster was significant between ~380 and ~596 ms after correction for multiple comparisons (p < .001) and generally located at central parietal channels. Only electrodes that remain significant throughout the entirety of each time bin are highlighted.

In sum, the ERP data provide evidence for both early and late differences in cue-driven processing between task switches and repetitions. We next sought to determine whether differential processing between these conditions was affected by the switch proportion context, akin to the interactions observed in the behavioral effects.

Interaction Effect of Task Sequence and PS

Cluster-corrected dependent-sample t-tests on the switch-minus-repeat catch-trial difference waves in PS30 versus PS70 blocks (PS30_{switch-repeat} – PS70_{switch-repeat}) revealed two significant negative clusters with similar parietal-occipital topography. These clusters were significant between approximately 176-280 ms (p = .005) and 296-376 ms (p < .001), respectively. This interaction effect was driven by a switch-related negativity in the PS30 condition versus a switch-related positivity in the PS70 condition (Figure 5). In PS70 blocks, the switch-related waveform was greater in amplitude than the repeat-related waveform; on the other hand, in PS30 blocks, the switch-related waveform was smaller in amplitude than the repeat-related waveform. In other words, the more frequent trial type in each context elicted a larger positive response.

Figure 5.

Cue-Locked Interaction Effect of Block-wise PS and Task Sequence. (Top Left) ERP traces of all four conditions reflecting the task sequence x PS interaction for catch trials only. Activity over parietal occipital electrodes (highlighted in red) plotted. The switch – repeat difference was positive in the PS70 condition, while it was negative in the PS30 condition. This difference is driven by block-wise frequency modulation on repeat trials (PS30 – PS70 contrast in repeat trials revealed positive cluster between 148 to 380 ms, p < .001) but not on switch trials (no significant clusters found). (Bottom Left) Cluster-based analysis found two significant clusters both centered around parietal-occipital electrodes, with the first one being significant between 176 ms and 280 ms (p = .005), and the second between 296 and 376 ms (p = .012). (Top Right) ERP traces of all four conditions reflecting the task sequence x PS interaction for catch trials and the non-catch trials with targets appearing post-380 ms. (Bottom Right) Cluster-based analysis again found two significant clusters centered around parietal-occipital electrodes. The first one was significant between 176 ms and 288 ms (p = .004), and the second between 300 and 384 ms (p = .007).

We conducted follow-up analyses to examine whether switch or repeat catch trials (or both) were driving this interaction. While the PS30 – PS70 contrast produced a significant negative cluster in repeat trials between 148-380 ms (p < .001), the same contrast yielded no significant clusters in switch trials. This indicates that the interaction of task sequence and PS context was mostly driven by modulation of repeat-trial rather than switch-trial activity.

To corroborate these effects with higher trial counts, we conducted the same analysis after including trials with targets occurring later than 380 ms. Again, two negative clusters were found over parietal-occipital channels, displaying significant effects between 176-288 ms (p = .004) and 300-384 ms (p = .007), respectively (Figure 5).

These ERP results indicate that the switch-likelihood (PS context) reliably modulated cue-evoked processing starting from ~170 ms after cue onset, and that this contextual modulation affected primarily the manner in which task repetitions were processed, thus paralleling the behavioral effects. Also note that while contextual effects only emerged ~170 ms post cue, task sequence effects were evident as early as ~100 ms post cue, suggesting that context-specific effects on processing only occurred after participants distinguished between switch task cues versus repeat task cues.

Time Frequency Decomposition: Alpha (8-14 Hz) Power

We first analyzed alpha power time-locked to trial onset, examining activity during the pre-cue interval, in order to gauge sustained, across-block effects of PS context. The analyses of this trial onset-locked alpha power between the PS30 and PS70 conditions yielded no significant effects. Thus, similar to the ERP results, we obtained no evidence for sustained tonic attentional state differences between the blocks where switches were frequent versus blocks where they were rare.

Next, we examined cue-locked alpha power. We observed no significant differences in alpha power between switch- and repeat- catch trials. However, the interaction effect between task sequence and PS (PS30_{switch-repeat} – PS70_{switch-repeat}) yielded a significant cluster of decreased alpha-band power (p = .003) over occipital-parietal channels, which was significant throughout the trial time course (Figure 6A). Follow-up specific contrasts found that this was due to significantly lower alpha power evoked in the repeat- compared to switch- trials in the PS70 condition (cluster p < .001), but no differences between switch and repeat trials in the PS30 condition (Figure 6B). More specifically, this figure shows that there was a marked decrease in alpha for repeat trials in the PS70 blocks compared to the other three conditions, demonstrating the PS context by task-sequence interaction appeared to be driven by alpha suppression for the rare repeat trials in PS70 blocks. These alpha power differences thus mirror the ERP and behavioral data patterns, in that the PS context affected primarily activity on repeat trials rather than on switch trials.

Figure 6.

Cue-locked alpha effects. Nonparametric analysis of the interaction effect between PS and Task Sequence revealed a positive cluster in the 8-14 Hz frequency band over parietal-occipital channels (p = .003) which persisted throughout the trial time course. (A) Top: power spectrum of the difference between task switch and repeat trials in the PS30 condition (left) and in the PS70 condition (right). Bottom: topographic map of the power in the 8-14 Hz frequency band from 300-500 ms post-cue. (B) Average log₁₀ power between 8-14 Hz frequency, over occipital-parietal electrodes (highlighted in red) from 100-500 ms post-cue in all four conditions. The error bars represent the within-participant standard error. The interaction effect was driven by a greater difference in alpha power between switch and repeat trials in the PS70 condition. Follow up cluster-based analyses found significant positive switch related activity (**cluster p < .001) in the PS70 condition but no significant clusters in the PS30 condition.

Discussion

This study set out to identify ERP and oscillatory-EEG signatures underlying meta-flexibility. EEG was recorded during a cued task-switching paradigm with a block-wide manipulation of the proportion of task switches. Our behavioral results replicated previous findings of the LWPS effect (e.g., Dreisbach & Haider, 2006; Monsell & Mizon, 2006; Siqi-Liu & Egner, 2020) – of lower switch costs in high-PS compared to low-PS blocks, indicating adaptation of cognitive flexibiility to situational demands. The ERP results showed significant differences in neural processing between switch and repeat trials that emerged as early as ~100 ms post-cue, and significant modulation of switch-related processes by context by ~170 ms after cue onset. In addition, a modulaton of task-cue processing by switch frequency was also identified in time-frequency decomposition of cue-locked alpha power, but not by any sustained, blockwise, oscillatory effects in place prior to cue onset. Both ERP and time-frequency results paralleled the patterns of RT effects, suggesting that they indexed neural processes that ramified to observed behavioral responses. In combination, these EEG signatures of the LWPS effect provide evidence that block-wise switch context phasically modulates task-set reconfiguration early on during the cue to stimulus interval, triggered by the task cue, rather than producing tonic cognitive state changes that are in place prior to cue onset.

In addition to the inverse relationship between switch costs and switch proportion, the behavioral results also replicate some previous observations that switch-cost adjustments were primarily driven by RT slowing on repeat trials, rather than RT acceleration on switch trials, and specifically in high-PS blocks (Siqi-Liu & Egner, 2020). This seemingly runs counter to the intuitive expectation for adjustments of cognitive flexibility to manifest as performance benefits on switch trials in frequent switch contexts, rather than performance costs on repeat trials. However, as suggested in Siqi-Liu & Egner (2020), “backward inhibition”, the finding that it is more difficult to return to a task that was switched away from on the prior (N-2) trial (Mayr & Keele, 2003), may increase average switch trial RTs specifically in high switch frequency blocks, which contain more sequences with consecutive task alternations (e.g. ‘ABABA’) compared to low switch frequency blocks, which contain longer sequences of consecutive task repeats (e.g. “AAABB”). In line with this explanation, Bonnin et al. (2011) demonstrated that backward inhibition could limit switch-trial RT reductions in high switch frequency contexts, particularly in short response-stimulus interval (RSI) conditions. Though we cannot test for backward inhibition effects in the current dataset because participants only switched between two tasks (i.e. we cannot compare ABA versus ABC task sequences), the similar switch trial RTs in high and low PS contexts suggest that backward inhibition may have masked, or cancelled out existing context-sensitive switch-trial RT reductions in the high-PS context. Conversely, if no context-sensitive switch-trial RT reductions existed in the PS70 blocks, one would expect to observe slower switch-trial RTs compared to the PS30 blocks due to backward inhibition.

On the other hand, expectancy violations may have contributed to comparatively slower RTs on repeat trials in frequent switch contexts (King et al., 2012). Nonetheless, building up expectations for switch- versus repeat trials when they occur more frequently is an inherent and important part of making behavioral adaptations to different PS contexts.

ERP analyses of switching-related neural processing identified an early cue-locked switch-related positivity in frontal-central channels starting at ~100 ms following cue onset. This early difference would seem to indicate that task-set reconfiguration processes occurred quite rapidly after cue onset. Note that this observation was not confounded by cue repetitions in task repeat trials, since we included two cues for each task and used alternate cues even on repeat trials. This analysis further identified a slightly later switch-related negativity in posterior electrodes between ~400 ms to 600 ms. This switch related-negativity would seem to be inconsistent with previous findings of posterior switch-related positivity in the general P3b time window (Barceló & Cooper, 2018; Han et al., 2018; Jost et al., 2008; Karayanidis et al., 2010, 2011; Lavric et al., 2008; Nicholson et al., 2006; Wong et al., 2018). One potential explanation for this is our inclusion of different PS contexts, which was not manipulated in any of these prior studies.

Although there was no strict overlap between the significant timepoints of the main and interaction effects, the time window (~180 to 280 ms; ~300 ms to ~380 ms) and posterior topography of our interaction effects aligned more closely with the later switch-related negativity over posterior channels compared to the earlier activity over frontal channels in the switch-repeat contrast. The interaction effects also fell within the time window of the previously reported P3b-like switch-related positivity over centroparietal electrodes (see Karayanidis et al., 2010 for review). Importantly however, a closer examination of the interaction effect revealed that it was driven by a switch-related negativity in PS30 blocks in contrast to a switch-related positivity in PS70 blocks (see Figure 5). The switch-related positivity that emerged over posterior channels when task switches were frequent is thus largely consistent with previous findings of P3b-like switch-related positivity. Thus, the current results suggest that the switch-related positivity reported in previous studies may in fact be context-dependent, emerging only when switches are relatively common. This conjecture would require confirmation by future studies. Notably, these context-specific interaction effects occurred after basic switch-related processing (which occurred ~100 ms post-cue) and there was no main effect of PS context in our ERP analyses, contrary to the idea of a tonic change in flexibility between PS contexts. Notably, there was no main effect of PS context in our ERP analyses, and these context-specific interaction effects occured after basic switch-related processing (occurring ~100 ms post-cue), thus speaking against the idea of a tonic change in flexibility between PS contexts.

Similar to the behavioral RT results, this interaction effect was driven primarily by repeat-trial amplitude differences across different PS contexts, with higher amplitudes associated with the PS70 condition. Since previous studies have shown that backward inhibition (relative to the task on the previous trial) in cued-task switching paradigms can affect ERP waveform amplitudes on the current trial (Sinai et al., 2007; Zhang et al., 2016), context-sensitive adjustments in the EEG data may also have been harder to observe in switch than in repeat trials. In other words, repeat-trial ERPs may serve as cleaner representations of context effects because they are not affected by disproportionate backward inhibition in high switch proportion blocks. Relatedly, the lack of difference in switch-trial activity between the PS30 and PS70 conditions may also reflect the delayed or dampened initialization of task set reconfiguration processes in the PS70 condition due to backward inhibition. While a main effect of PS context would have provided evidence for “tonic” flexibility adjustments based on list-wide context, the lack of even numerical trends in this direction in our data set suggests that context-sensitive adjustments are expressed primarily after task-cue processing. The finding that contextual interaction effects only occurred after the ERPs differentiated between basic switch versus repeat trial types further supports this conclusion. The idea that the LWPS effect does not represent tonic changes of flexibility state is consistent with Siqi-Liu & Egner’s (2020) finding that behavioral adjustments occurred in a task-specific manner, rather than a block-wide one, such that switch costs were lower for tasks that were more frequently presented as switch compared to repeat trials (switch-biased), and vice versa, while no switch cost adjustments were observed in “unbiased” tasks (equally associated with switches and repeats) presented in high or low PS lists. Here, in addition, the high-temporal resolution of the ERPs supplies neural evidence for the time course and sequence of the influence of these cognitive factors on the underlying neural processes.

Time-frequency analyses of alpha power were also conducted as a way to observe potential adjustments in attentional states. Consistent with the time course of our ERP effects, no alpha power differences between block-wide switch contexts were identified in the period preceding cue-onset, again further supporting the idea that flexibility adjustments did not occur in a block-wide, tonic manner, at least as reflected by ongoing neural activity. Power differences in other frequency bands such as theta and beta (Cooper et al., 2015, 2017, 2019; McKewen et al., 2020) have previously been shown to underly proactive control recruitment, though no significant effects at those frequency bands were identified in the current study. It is important to note however, that neither lack of pre-cue alpha effects nor the lack of effects at theta and beta frequency bands in the current study constitutes evidence for a strong conclusion that lasting differences in attentional or control states across blocks as a function of PS context do not exist.

While there was no main effect of either switch/repeat trial type or list-wide PS context on the cue-locked alpha activity, there was a significant interaction effect between trial type and context on the alpha, again centered over parietal-occipital electrodes. This effect was driven by relative alpha-suppression in task-repeat trials compared to task-switch trials in the PS70 condition, with no differences in alpha power being observed between switch and repeat trials in the PS30 condition. Again, paralleling the behavioral and ERP results, the context-sensitive adaptations were driven by changes in repeat-trial rather than switch-trial processing.

Since alpha power decrease (or suppression) is generally thought to correspond with increased recruitment of attentional resources (e.g., Foster et al., 2017), our time-frequency results could reflect the increased difficulty of repeat-trials in contexts where they are rare compared to when they are common. That these adjustments only occur in repeat trials may indicate that participants employ a default strategy of expecting and preparing for task repetitions when they occur most of the time (Duthoo et al., 2012), and changing this default repetition expectation may be effortful. That is, when the repetition expectation is violated in PS70 blocks, participants may have to abandon this default strategy, which leads to less efficient processing of task repetitions in high-PS compared to low-PS contexts. Alternatively, as discussed above, it is possible that parallel effects on switch trials are masked by backward inhibition effects.

In sum, the lack of any main effects of list-wide context in either ERP or trial-onset-locked time-frequency measures in the current study suggests that contextual modulations occur during cue-processing, wherein task cues can rapidly trigger different processing strategies depending on whether they are associated with frequent task switches or repetitions (Braem & Egner, 2018). An intriguing possibility is that block-wide differences in flexibility state may be “activity-silent” in terms of ERP and time-frequency power, analogous to the concept of activity-silent working memory representations (reviewed in Stokes, 2015) that are only revealed only in response to external inputs (Rose et al., 2016; Wolff et al., 2017). That is, there may be sustained state differences in the brain as a function of PS context, but such state differences may not be observable in EEG and ERP activity patterns until they are triggered by the task cue. Indeed, our finding that the PS context influenced switch-vs-repeat cue processing as early as 176 ms following cue onset would seem to suggest that PS context may induce some latent preparedness for task repetitions or switches, such that the optimal neural response is rapidly triggered when the task cues occur. Future studies will be needed to corroborate this conjecture.

Siqi-Liu & Egner (2020) proposed that such post-cue adjustments of cognitive flexibility could potentially be initiated via decreased task-set activation of frequently switched-to tasks. That is, task-sets frequently associated with switches may be maintained at lower levels of activation, which promotes greater ease of switching at the cost of reduced repetition benefits. The idea that greater flexibility is mediated by weaker task-set activation is supported by neuroimaging evidence from Qiao et al., (2017), who found that neural task-set representations were less stably encoded in frontal-parietal cortex activity patterns on switch compared to repeat trials. Reduced repetition advantage in high-PS contexts is evident in the current study’s finding that behavioral, ERP, and time frequency effects were all largely driven by repeat-trial adaptations, rather than switch-trial adaptations. Furthermore, the observed alpha-suppression in response to rare repeat trials in high-PS blocks supports the idea that repeat trials were perceived as more surprising or effortful (despite behavioral switch costs) when participants have learned to expect frequent task switches.

Importantly, the lack of sustained neural activity evidence for tonic flexibility adjustments to list-wide PS context in the current study does not necessarily contradict findings of list-wide adjustments of conflict-control in the congruency literature, which are often interpreted as evidence for proactive control in contrast to stimulus-driven, reactive control (Gonthier, Braver, & Bugg, 2016). For example, using a picture-word Stroop task, Bugg & Chanani (2011) found list-wide adjustments of conflict-control after controlling for stimulus-specific associations. Importantly, in contrast to task-switching paradigms, conflict paradigms like the Stroop task never involve task-set shifts (e.g., participants are always instructed to name the picture and ignore the word) and so may not be as suited for identifying tonic control strategies that function across different task-sets. The current study’s findings suggest that list-wide context may only influence switch-related neural activity after task-cue onset, highlighting the important role that task sets play in determining information processing strategies. People may learn to link appropriate control settings to task sets and rapidly retrieve those control settings when cued in a bottom-up manner in a process that represents an integration of proactive and reactive control mechanisms (see also Siqi-Liu & Egner, 2020).

Conclusions

The present study identifies ERP and time frequency signatures of the LWPS effect, and provides evidence that neural processing is modulated by block-wide switch proportion context during the cue-target interval. Behavior, ERP, and time-frequency measures consistently indicated that modulations of repeat-trial processing drove context-sensitive flexibility adaptations. The lack of pre-cue effects or main effects of switch context suggest that “tonic” adjustments in attentional or flexibility states across the temporal context of the block are not what drives the LWPS effect. Rather, contextual modulation seems to occur during cue-processing and is potentially underpinned by weaker activation of tasks when participants have to switch between tasks more frequently compared to contexts where task repeats are more common.