Context independent reductions in external processing during self-generated episodic social cognition

Ongoing cognition supports behavioral flexibility by facilitating behavior in the moment, and through the consideration of future actions. These different modes of cognition are hypothesized to vary with the correlation between brain activity and external input, since evoked responses are reduced when cognition switches to topics unrelated to the current task. This study examined whether these reduced evoked responses change as a consequence of the task environment in which the experience emerges. We combined electroencephalography (EEG) recording with multidimensional experience sampling (MDES) to assess the electrophysiological correlates of ongoing thought in task contexts which vary on their need to maintain continuous representations of task information for satisfactory performance. We focused on an event-related potential (ERP) known as the parietal P3 that had a greater amplitude in our tasks relying on greater external attention. A principal component analysis (PCA) of the MDES data revealed four patterns of ongoing thought: off-task episodic social cognition, deliberate on-task thought, imagery, and emotion. Participants reported more off-task episodic social cognition and mental imagery under low external demands and more deliberate on-task thought under high external task demands. Importantly, the occurrence of off-task episodic social cognition was linked to similar reductions in the amplitude of the P3 regardless of external task. These data suggest the amplitude of the P3 may often be a general feature of external task-related content and suggest attentional decoupling from sensory inputs are necessary for certain types of perceptually-decoupled, self-generated thoughts.


Introduction
Cognition supports action by shaping behavior in response to our current or future needs. Responding to immediate environmental demands requires attending to the here and now, while anticipating how to act in the future depends on mental contents from memory (Smallwood & Schooler, 2015). Contemporary theoretical views of ongoing thought suggest these modes vary in the correlations between brain activity and external information: Features of neural processing are argued to be correlated with signals arising externally for efficient action in the moment, while in situations where cognition is self-generated this correlation is reduced (Smallwood, 2013). Accumulating evidence supports this process-perspective of ongoing thought. For example, during sustained attention, off-task thoughts are associated with reductions in electrophysiological indicators of the processing of task-relevant perceptual signals (Barron, Riby, Greer, & Smallwood, 2011;Bozhilova, Kuntsi, Rubia, Michelini, & Asherson, 2021;Kam et al., 2021;Kam, Rahnuma, Park, & Hart, 2022;Smallwood, Beach, Schooler, & Handy, 2008;Villena-Gonz alez, L opez, & Rodríguez, 2016). More generally, off-task states are linked to errors in situations demanding detailed processing of the external environment (McVay & Kane, 2009;Smallwood, McSpadden, & Schooler, 2008). Although accumulating evidence establishes greater coupling with external input during periods of externally focused cognition, it has become important to understand the context dependence of this phenomena. For example, studies using functional Magnetic Resonance Imaging (fMRI) have identified that certain neural correlates of task-related and off-task thought emerge in a similar way across situations, while other may emerge in a more context dependent manner. Recent studies using fMRI Turnbull et al., 2019a, b), for example, established regions of dorsal parietal cortex reduce activity when individuals report off-task thought regardless of whether they performed a nondemanding task or one that required continuous task monitoring. In contrast, activity within the dorsolateral prefrontal cortex exhibited a context dependent pattern: greater activity was associated with off-task thought during simple nondemanding tasks and greater external task focus when tasks were more complex.
Since prior studies have identified both context dependent and independent correlates of off-task thought, our study, therefore, aimed to understand whether the relation between evoked responses and patterns of ongoing thought emerge in a similar manner across different tasks. To this end, we assessed the correlation between off-task thought and brain activity in situations varying in their need for continuous taskrelated attention using the same paradigm as employed before (e.g., Bozhilova et al., 2021;Konishi, Brown, Battaglini, & Smallwood, 2017;Konishi, McLaren, Engen, & Smallwood, 2015;Turnbull et al., 2019b). This paradigm alternates between a situation requiring continuous external attention where task-relevant information must be maintained in working memory (1-back) and a situation with no equivalent requirement (0-back) (Fig. 1). In both tasks, participants made intermittent decision about the location of shapes (Fig. 1). Most electrophysiological studies on self-generated cognition have characterized mind wandering as unitary "off-task" attentional state (however see Villena-Gonz alez et al., 2016). To understand whether the mental contents of both off-task and on-task states affect evoked responses to external stimuli, the decisions about the location of shapes were occasionally replaced by questions that sampled participants' ongoing experience (Smallwood et al., 2016) along several dimensions (Table 1). To ensure that participants were unable to anticipate when the decisions occurred, we varied the time between these target decisions and experience sampling from approximately 5 to 15 s following earlier reports (Turnbull et al., 2019a). Since our design had these features, we also explored how the length of these temporal intervals varied the magnitude of the brain response to task events, to reported patterns of ongoing thoughts and their association.
Brain activity was measured using an event-related potential (ERP) extracted from electroencephalography (EEG) signal, which provide an unambiguous index of the momentary correlation between task events and brain activity. We focused on a visual evoked response, occurring at a latency of 300e500 ms and known as a visual P3. This response broadly reflects attentional allocation (Polich, 2012) and activations in temporal/parietal cortex for visual stimuli (Linden, 2005;Polich, 2012). We chose this evoked component because prior work has shown that P3 is routinely suppressed during periods of off-task thought, an observation made by multiple different research teams (reviewed in Kam et al., 2022). We hypothesized that if the correlation between P3 amplitude and off-task thought varied as a function of condition (i.e., between the 0-back and 1-back tasks) this would suggest that P3 is sensitive to fluctuations in ongoing thought when external tasks require continual maintenance of task relevant material. In contrast, if P3 amplitude is reduced in a broadly similar manner in association with off-task thinking across task contexts, this would support the view that this pattern is a more general feature of reduced external task focus.

Participants
We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/ exclusion criteria were established prior to data analysis, all manipulations, and all measures in the study. No part of the study procedures or analyses were pre-registered prior to the research being conducted. Twenty-eight volunteers (22e36 years, mean age: 26 years; 14 females, 14 males; 3 left-handed) participated in the study. This sample size corresponds to previous studies (Barron et al., 2011;Smallwood, Beach, et al., 2008) reporting within participants mind wandering effects on ERPs. All participants were free of psychiatric or neurological history, current use of medication, and had normal or corrected to normal vision. They gave a written informed consent prior to the experiment, and the experiment was performed in accordance with the Declaration of Helsinki and with permission by the Ethical Review Board in the Humanities and Social and Behavioural Sciences, University of Helsinki.

Experimental design
Participants saw six possible pairs of shapes including a square, a triangle, and a circle (two different left/right configurations for each, and the pairs never had shapes of the same kind, Fig. 1). The pairs were black, shown on grey background divided by a vertical line and comprised the nontarget (NT) stimuli. Each NT-stimulus was presented for .5 sec followed by a central fixation cross presented for a random duration between 1.8 and 2.2 sec. A sequence of NTs consisted of 2e6 NTs (approximately 5e15 sec) and was followed by a target stimulus requiring participants to make a manual response. The target was presented until response or for a Fig. 1 e Schematics of the tasks. Participants performed alternating blocks of a) 0-back and b) 1-back tasks. In both conditions, Non-Targets (NTs, duration .5 sec, sequence varying between 2 and 6 NTs) were presented before the target trials. In the 0-back task, during target trials (panel a, bottom), participants had to respond based on which side the larger shape matched the small target shape in the middle. In the 1-back task (panel b, bottom), the target shape was flanked between question marks, and the response was based on which side the target shape was in the preceding trial. The conditions were indicated by target stimulus colors (either blue or red) that were counterbalanced across participants. The NTs were irrelevant for the task in the 0-back, allowing periods when attention is unconstrained by the task. In the 1-back task, the responses are based on the previously attended NTs and participants must maintain attention on the NTs to perform accurately in the task. In both conditions here (bottom of the panels), the correct answer is 'Left'. Target trials were presented until response or for the maximum duration of 4 sec. Between each NT and target stimulus, a central fixation cross was presented for random duration varying between 1.8 and 2.2 sec. Note that all fixation crosses are not illustrated here. Ongoing thought was measured by randomly replacing a proportion (20%) of the target trials with the MDES questions that were presented until response or for the maximum duration of 10 sec. My thoughts were intrusive c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 maximum duration of 4 sec. The task consisted of two conditions. In the 0-back condition, the target was a small shape flanked by a pair of shapes and participants had to respond with their right hand using left or right buttons of a response pad depending on which shape matched the target shape.
Participants were instructed to respond as fast and accurately as possible to the targets. In the 1-back condition, the target was flanked by two question marks and participants had to respond depending on which side the target was on the previous trial. After each target, either "STAY" or "SWITCH" was presented for 1 sec, signaling a continuum or a switch of the condition respectively. One session included 3 blocks for each condition and each block lasted approximately 3 min. The number of targets per block varied randomly between 8 and 10 before the switch to another condition. In total, each session included 33 target and 129 non-target (NT) stimulus presentations per condition. Please note that since the experience sampling probes occurred instead of when targets occurred (see also Smallwood et al., 2011;Smallwood, Ruby, & Singer, 2013;Turnbull et al., 2019a, b), and that there are no behavioral responses to non-targets, this design makes it impossible to link momentary variation in performance (i.e., RT or accuracy) to thought patterns. The target stimulus colors (blue and red) indicating the conditions were matched for luminance. The starting order of the 0-back and 1-back conditions as well as the color/condition pairing was counterbalanced across participants. The experiment was divided into four recording sessions. First, resting state activity was collected with eyes open for 10 min. These results are not reported here. The three task sessions lasted approximately 18 min each and consisted of alternating blocks of the 0-back and 1-back tasks.
To sample participants' ongoing thoughts, in the last two sessions, task performance was interrupted occasionally, at moments when otherwise a target would have occurred, by the presentation of the MDES questions. The task was designed so that there was a 20% chance of a thought probe being presented in place of a target stimulus (Fig. 1). Thought probes were presented until response or for the maximum duration of 10 sec. The experience sampling protocol consisted of a series of ten questions ( Table 1). The first question always prompted participants to rate their task focus ("Just before this question was presented, were you focused on the task or were you thinking of something unrelated to the task?") using four buttons of the response pad with their left hand from (1) not at all to (4) completely. To the other nine questions, participants responded using the same scale as in the first question, except for the question regarding the valence of their thoughts for which the scale went from (1) negative to (4) positive. The same questions have been used in previous studies (e.g., Konishi et al., 2017). MDES seeks to use data driven methods to establish patterns of thoughts based on the covariance amongst how participants answer the range of items sampled. For example, it is common to identify patterns of covariance in an MDES study in which participants describe their experiences as unrelated to the task, but instead are focused on information related to themselves or other people and directed to other times (e.g., past or future). Importantly, this pattern broadly corresponds to one widely accepted definition of mindwandering (Smallwood & Schooler, 2015). To measure the potential impact of experience sampling on behavior and brain activity, in the first task session, the thought probe questions were presented only once at the end of the session, allowing for uninterrupted task performance.
Prior to the experiment, participants completed a practice trial of the tasks, including full instructions. We also presented them the ten MDES questions, to make sure they understood the questions and in particular the definition of being on-task (e.g., being fully focused on the task and having Fig. 2 e Accuracy, RT and task focus reports as a function of task condition and elapsed time. The line graphs show the behavioral accuracy, RT, and task focus results averaged over participants (N ¼ 28) and the sessions for the task conditions as a factor of elapsed time, i.e., number of trials from the last target trial. a) Accuracy was higher in the 0-back task (grey) but did not differ significantly with elapsed time. b) RTs were longer in the 1-back task (black) and decreased with elapsed time, specifically in the 1-back condition. c) Participants engaged in more on-task focus in the 1-back task, but the task focus reports did not differ significantly as a function of elapsed time. Note that the line graphs visualize the results but do not directly describe the slopes reported in the LMM analyses. Error bars show the standard error of the mean. c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 thoughts related to the task) versus being off-task (e.g., thinking about something unrelated to the task, such as planning a grocery list for dinner). All instructions and questions were given in English. Stimulus material and Presentation scripts for running this task are accessible online at https://github.com/jsimola/MW-EEG.

2.3.
Data acquisition and pre-processing Stimuli were presented on a 24.5-in display from a viewing distance of 1.2 m and the stimulus timing was controlled by Presentation™ software (Neurobehavioral Systems, Inc., Albany, CA, USA). Manual responses were collected with a Cedrus RB-840 response pad (Cedrus Corporation, San Pedro, CA, USA). Participants were comfortably seated during the recordings and asked to relax and avoid excessive movements and muscle tension. EEG data were collected using an ActiCap 32-channel active electrode system with a BrainAmp DC amplifier (Brain Products, Gilching, Germany) with a sampling rate of 500 Hz. An electrode was also placed below the left eye to detect vertical eye movements. During recordings, EEG data were referenced to the vertex (FCz) electrode.
The EEG data were processed in Matlab using the EEGLAB toolbox (Delorme & Makeig, 2004). The data were filtered with .5e45 Hz band pass, downsampled to 256 Hz and rereferenced to the average of mastoids (electrodes TP9 and TP10). For one participant two electrodes with excessively noisy signals were interpolated from neighboring electrodes. Blink, eye movement and muscle related artifacts were removed from the data using independent component analysis (ICA) and ICLabel classifier with the threshold of .9. Epochs from 100 ms prior to the onset of the NT stimuli to 1000 ms post-stimulus were extracted and epochs containing artifacts exceeding ±100 mV were removed. On average, 2e5% of the NT-trials in the 0-back and 2e4% of the NT-trials in the 1-back condition were removed from the further analyses across sessions. Before plotting and statistical analysis, the epochs were baseline corrected relative to a 100 ms period before stimulus onset.
Mean amplitudes were calculated from 300 to 500 ms after NT-stimulus onsets for the parietal electrode channels (P3, Pz, and P4). To investigate the latency for this response, we extracted the time point of the maximum positive voltage in the same channels. The evoked potential was thus identified within selected time window and electrode sites (Keil et al., 2014) and for which the effect was expected to be largest, based on prior research (Barron et al., 2011;Bozhilova et al., 2021;Kam et al., 2021;Smallwood, Beach, et al., 2008) and which correspond to a visual P3 event-related potential (ERP) component, as identified before broadly in association with attention allocation (Linden, 2005;Polich, 2012). The parameters were also verified against the grand average waveforms and scalp topographies (Fig. 4). The interpretation of the P3 component depends on task context, and it should be noted that our task differed from the typical tasks used to elicit the P3 component, i.e., detection of infrequent targets among frequent standard stimuli (Polich, 2012). However, the visual P3 seems to reflect multitude of attentional allocation processes that are disrupted due to internally directed attention (Kam et al., 2022).

Statistical analysis
The multidimensional experience sampling (MDES) data collected with the thought probe questions were analyzed with principal component analysis (PCA) to determine patterns of covariance in the participants' responses to the thought probes. PCA was carried out in SPSS (Version 27.0, 2020) by entering the trial level responses to the 10 experience sampling questions (Table 1). Decomposing of data like these using varimax rotation is standard in our approach (e.g., Konishi et al., 2017;Sormaz et al., 2018;Turnbull et al., 2019a, b). We then plotted the rotated component loadings across dimensions of thought as wordclouds (Fig. 3). The PCA component scores were analyzed by fitting separate models for each component. We also provide an analysis of these data using an alternative rotation strategy of oblique rotation method which assumes that the components are correlated (Supplementary material S1, Table S1) and using an alternative number of components (3) (Supplementary material S2). These minor alterations to our analytical pipeline made little or no difference to the pattern of results so for consistency with prior studies we focus on the varimax 4 factor solution given the similarity with prior approaches (e.g., Konishi et al., 2017;Sormaz et al., 2018;Turnbull et al., 2019a, b). Accuracy data were analyzed with a generalized linear mixed model (GLMM) using binomial distribution and participants' self-reports on task focus were analyzed with a GLMM using a Poisson distribution. All other measures were analyzed with linear mixed models (LMMs). These analyses were run using the lme4 package (Bates, Mӓchler, Bolker, & Walker, 2015) for R (R Core Team, 2018). Separate models were fitted for each dependent variable with the restricted maximum likelihood criterion. When implementing LMMs, model selection problems are unavoidable. We used a stepwise forward selection heuristic for comparing LMMs with different fixed-effects structures. For model comparisons, we used maximum likelihood criteria. The models included either task condition and/or session and/or elapsed time (from the last target trial) and their interaction term as fixed effects. Elapsed time, i.e., number of Non-Target (NT) trials from the last target trial, was included to investigate how task performance, ongoing thoughts and electrophysiological activity evolved over time. Our study examines within participant variation in how ongoing thought patterns (as described in MDES) are related to changes in brain activity. To this end, random intercepts for participants and random slopes for condition at the level of participants were included. We used a Likelihood Ratio Test (LRT), to test if adding factors would improve model fit. If adding factors improved model fit, we compared the model that included interaction to the model having the individual predictors but no interaction. In the Results section, we report the likelihood ratio test as c2(d.f.), where c2 is the likelihood ratio test statistic and d.f. is the degree of freedom (Bates et al., 2015;Harrison et al., 2017). For the mean ERP amplitudes of all NT- Fig. 3 e Components of thought revealed by the experience sampling probes. Principal Component Analysis (PCA) was applied to the multidimensional experience sampling (MDES) data collected from participants (N ¼ 28) with the 10 experience sampling questions and resulted in four principal components of thought. The wordclouds on the right describe the loadings of each item for these components, with the font size representing the magnitude of the loading and the color the direction of the loading (blue ¼ negative, red ¼ positive). Components were labelled as a) Off-task episodic social c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 Fig. 4 e Grand-averaged P3 event-related potentials (ERPs), brain topographies and P3 dynamics over time for Non-Target (NT) stimuli. Mean ERP amplitudes in microvolts (mV) across all participants (N ¼ 28) for all NT stimuli at the parietal Pz electrode as a function of task condition (0-back ¼ grey, 1-back ¼ black) separately for the three sessions: a) the first session (without thought probes, TPs), b) the second session (with TPs) and c) the third session (with TPs). Grey shading indicates the time window from 300 to 500 ms post-stimulus from which the mean amplitude was calculated for the statistical analyses. Coral shadings indicate the p-values obtained from paired samples t-tests comparing amplitudes between 0-back and 1-back tasks in a moving average window of 50 ms. The topographical maps (with the Pz electrode highlighted) in the middle show the scalp distributions of the ERPs for the 300e500 ms time window separately for each session and task condition. The line graphs on the right show the P3 amplitudes at the Pz electrode for the task conditions as a factor of elapsed time, i.e., number of trials from the last target trial. Note that the scales vary between plots and that the line graphs visualize the results but do not directly describe the slopes reported in the LMM analyses. Error bars show the standard error of the mean. cognition, b) Deliberate thought, c) Imagery and d) Emotion. Violin plots on the left show the distributions of component scores, including medians and ± 1.5 £ IQR for the 0-back (grey) and 1-back (black) condition across all participants. The line graphs in the middle show the component scores for the task conditions as a factor of elapsed time, i.e., number of trials from the last target trial. Note that the line graphs visualize the results but do not directly describe the slopes reported in the LMM analyses. Error bars show the standard error of the mean. **: p < .01. c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 trials, we also included electrode location at the parietal site (with levels: P3, Pz, and P4) as a fixed factor to investigate possible laterality effects for the P3 ERP response. To investigate the electrophysiological correlates of thought patterns reflected in the PCA components, we predicted the P3 amplitudes immediately preceding the thought samples with models that further included the component scores as fixed factors. To investigate behavioral correlates of thought patterns, we computed Spearman's rank correlations between individual average scores for task performance and thought probes. Further, we assessed the relationship between behavior and P3 amplitudes by adding P3 amplitude to the models explaining both accuracy and RTs. The analysis codes and data underlying statistics and figures can be found at https://github.com/jsimola/MW-EEG. Original behavioral and EEG data are stored at https://osf.io/879ex/.

3.1.
Task manipulation impacts behavioral accuracy, response times and task focus reports As a manipulation check for our task conditions, we first analyzed accuracy and response time (RT) data across all three sessions as well as participants' self-reports on task focus on the last two sessions where thought probes (TPs) were presented. Inter quartile ranges (IQRs) were calculated for the RT data and trials with the RT outside the range of Q1 e 1.5 Â IQR and Q3 þ 1.5 Â IQR were excluded (comprising 6.3% of the trials) from these analyses.
For the task focus reports collected from the sessions with TPs, random slopes for condition at the level of participants were removed from the model to obtain a non-singular fit. Adding session and elapsed time did not improve the model fit for the task focus reports. The main effect of task condition showed that participants reported to be more in on-task mode in the 1-back than in the 0-back task, b ¼ .20, 95%CI ¼ [.11, .29], SE ¼ .05, z ¼ 4.42 (Fig. 2c, Table 2). Together, these analyses showed that the 1-back task was linked to both lower accuracy and slower RTs than the 0-back task. Moreover, analysis of the task focus reports demonstrated that participants engaged in more on-task thoughts during the 1-back than the 0-back task.

Identification of thought patterns
Next, we examined the MDES data to understand the patterns of thinking across both tasks. Following prior studies (Konishi et al., 2017;Sormaz et al., 2018;Turnbull et al., 2019a, b) we decomposed the trial level experience sampling reports (730 reports in total) using PCA and found four dimensions of thought (Fig. 3). See Supplementary material S1 (Fig S1, Tables S1-S3) for rationale behind the application choices concerning the PCA. Together these components explained 64.4% of variance in the MDES data (Table S2). Based on varimax rotated component loadings (Fig. 3, Table  S3), the first component described thoughts that were not focused on the task but were high on content reflecting oneself, other people, and future, and described as intrusive. We refer to this component as off-task episodic social content. The second component reflected thoughts that were detailed, in the form of words, and directed towards the task. We refer to this component as deliberate and detailed task focus. The third component reflected thoughts that were in the form of images and concerned past and other people. The fourth component reflected the emotional quality of ongoing experience. We labelled the components (in order of decreasing explained variance, Table S2) as "off-task episodic social cognition" (Fig. 3a), "deliberate thought" (Fig. 3b), "imagery" (Fig. 3c) and "emotion" (Fig. 3d). These components are similar to those seen in parallel studies conducted in York (e.g., Sormaz   c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 across experiments and settings. Notably, component 1 had the lowest loading on the on-task score (Table S3), as well as high loadings on personally relevant content (self, future, person) and so corresponds to the state of off-task selfgenerated mental content that are often considered important features of the mind-wandering state (Smallwood & Schooler, 2015). We then tested the task condition effects on component scores with models that included task condition as a fixed factor and random intercepts for participants and random slopes for condition at the level of participants as the random factors. These analyses showed that all component scores except the scores for emotion differed between task conditions: off-task episodic social cognition, b ¼ À.34, 95% CI ¼ [À.55, À.12], SE ¼ .11, t ¼ À3.09; deliberate thought, b ¼ .51, 95%CI ¼ [.24, .79], SE ¼ .14, t ¼ 3.65; and imagery, b ¼ À.25, 95% CI ¼ [À.40, À.10], SE ¼ .08, t ¼ À3.34 (Fig. 3).
For deliberate thought, adding task condition (X 2 (1) ¼ 11.22, p ¼ 8.1e-4), but not session and elapsed time, improved model fit compared to the random slope model. The main effect of task condition, b ¼ .51, 95%CI ¼ [.24, .79], SE ¼ .14, t ¼ 3.65, indicated more deliberate on-task thought in the 1-back condition (Fig. 3b). This pattern is seen in this paradigm in both prior behavioral (Turnbull et al., 2019a) and in brain imaging studies (Turnbull et al., 2019b). As the deliberate on-task thought component loads high on thinking in the form of words, this result is in line with a pattern shown earlier , whereby thinking more in words during the 1-back task was associated with more efficient task performance, suggesting a beneficial strategy in the more demanding 1-back task.
For the imagery component, adding condition (X 2 (1) ¼ 9.66, p ¼ .0019) and elapsed time (X 2 (1) ¼ 4.53, p ¼ .033) significantly improved model fit compared to the random slope model, while adding session or the interaction terms did not improve the model fit. The analysis showed an effect of task condition, b ¼ À.25, 95%CI ¼ [À.40, À.11], SE ¼ .07, t ¼ À3.39, with reduced imagery type of thought in the 1-back condition (Fig. 3c). The reduction in imagery in the 1-back task is also consistent with earlier findings using this paradigm . Further, an effect of elapsed time, b ¼ .05, 95%CI ¼ [.003, .09], SE ¼ .02, t ¼ 2.08, showed increased imagery as a function of time irrespective of the task condition (Fig. 3c).
Task condition did not affect the emotion component (Fig. 3d). Adding session and elapsed time did not improve model fit. To obtain a non-singular model, random slopes for condition at the level of participants were removed from the model.
Next, we investigated the behavioral correlates of these thought patterns by computing correlations between individual average scores for thought patterns and accuracy separately for the two task conditions. These analyses indicated a correlation between accuracy and deliberate thought in the 1-back task (Spearman's r(26) ¼ .332, p ¼ .012). The result indicated improved accuracy in the cognitively demanding task when participants reported more on-task deliberate thoughts. This finding is also in line with Turnbull et al. (2020) who reported that thinking in words was associated with more efficient task performance in the 1-back but not in the 0back condition. No significant correlations were found between RTs and any of the thought patterns.

3.3.
Parietal P3 was larger for the 1-back task and increased over time To confirm that our task elicited the P3 component central to our investigation, we plotted the scalp topographies for the ERP amplitudes averaged over participants and all NT trials at the time window of 300e500 ms separately for the sessions and the task conditions (Fig. 4, middle panel). The scalp distributions indicated that the response amplitudes between 300 and 500 ms post-stimulus were strongest over parietal electrodes, particularly at the midline parietal scalp site (Pz), where the P3 response is typically maximal (Linden, 2005;Polich, 2012) and prior fMRI studies have highlighted a relationship to off-task thought (Turnbull et al., 2019b).
We next explored how evoked brain activity in this region varied across the task contexts. P3 amplitudes at parietal electrodes for NT stimuli were averaged over task conditions for each session and participant (Table 3). Adding task condition (X 2 (1) ¼ 19.32, p ¼ 1.10e-5) and subsequently session improved model fit (X 2 (2) ¼ 45.74, p ¼ 1.17e-10) compared to the random slope model but adding the interaction terms did not influence model fit. P3 amplitudes were larger for the 1back relative to the 0-back task, b ¼ 1.31, 95%CI ¼ [.81, 1.80], SE ¼ .25, t ¼ 5.18 (Fig. 4, Table 3). Further, the P3 amplitudes were larger in the 1st session compared to both 2nd and 3rd The P3 responses were also larger in the 3rd session compared to the 2nd session, b ¼ À.58, 95%CI ¼ [À1.01, À.15,], SE ¼ .22, t ¼ À2.62 (Fig. 4, Table 3). c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 Adding electrode to the model did not improve model fit, indicating no significant effect of laterality for the P3 responses. Therefore, we focused all following ERP analyses on the P3 response detected at the midline Pz electrode.
Analysis indicated that evoked responses were largest in the first session where there were no TPs (Fig. 4, Table 3). To understand whether this change reflected a change in the landscape of brain activity induced by experience sampling, we tested the effect of task for each session separately. These analyses showed that P3 amplitudes at parietal electrodes were larger for the 1-back relative to the 0-back task in all ]. This analysis is important because experience sampling could potentially disrupt the natural pattern of brain activity that would otherwise emerge during task processing. Our results, however, indicate that although presenting TPs resulted in decreased P3 amplitudes, the effect between the tasks remained implying that TPs did not disrupt the broader landscape of brain activity during the task performance.
Next, we investigated how the amplitudes of parietal P3 responses varied over time. P3 amplitudes to NTs in single trials were labeled by the number of NT-trials since the last target trial. Outlier amplitudes were removed by the IQRmethod (comprising 1.7% of the trials). To obtain a nonsingular model, random slopes for condition at the level of participants were removed from the model. Adding condition (X 2 (1) ¼ 38.75, p ¼ 4.82e-10), session (X 2 (2) ¼ 43.49, p ¼ 3.60e-10) and elapsed time (X 2 (1) ¼ 68.45, p < 2.20e-16) improved the model fit while adding the interaction terms had no effect on the model fit. The results indicated main effects of task condition, b ¼ 1.55, 95%CI ¼ [1.07, 2.03], SE ¼ .25, t ¼ 6.34, with enhanced P3 amplitudes in the 1-back task (Fig. 4, left panel), and with elapsed number of trials, b ¼ .72, 95%CI ¼ [.55, .89], SE ¼ .09, t ¼ 8.30 (Fig. 4, right panel). Moreover, session influenced P3 amplitudes, whereby the amplitudes were stronger in the session without TPs compared to both sessions with TPs .50], while the P3 amplitudes did not differ between the sessions including TPs. Thus, in addition to the effects of task condition and session, this analysis indicated an effect of elapsed number of trials (Fig. 4, right panel), with gradually increasing P3 amplitudes over elapsed time.
We further investigated the relationship between P3 amplitudes and task performance. We added the single-trial P3 amplitude to the models explaining both accuracy and RTs. First, we tested P3 effects on behavior with models including only P3 amplitude as a fixed factor. These analyses showed that larger P3 amplitudes were associated with improved accuracy, b ¼ .03, 95%CI ¼ [.002 .05], SE ¼ .01, t ¼ 2.14. However, adding P3 amplitude to the model including task condition and session as fixed factors, i.e., to the model that best predicted accuracy did not improve the model fit, indicating no association between accuracy and P3 amplitudes. In addition, RTs were affected by the P3 amplitudes, b ¼ À.85, 95% CI ¼ [À1.45 -.25], SE ¼ .31, t ¼ À2.77, whereby RTs were negatively associated with the P3 amplitudes (Fig. 5a). When the P3 amplitude was added to the model that best predicted RTs, i.e., to the model that included the interaction between task condition and elapsed time as a fixed factor, the effect of P3 amplitude on RTs remained, b ¼ À.64, 95%CI ¼ [À1.24 -.03], SE ¼ .31, t ¼ À2.07, along with the interaction between task condition and elapsed time, b ¼ À9.48, 95%CI ¼ [À17.53 -1.43], SE ¼ 4.11, t ¼ À2.31.

3.4.
Relation between parietal P3 and ongoing thought patterns Finally, we pooled data from the sessions including TPs and investigated whether the PCA components reflecting different patterns of ongoing cognition (Fig. 3) were predictive of the P3 amplitudes that immediately preceded the thought probes. Outlier amplitudes were removed by the IQR-method (comprising 1.7% of the trials). Adding off-task episodic social cognition improved the model fit compared to the model with task condition as a fixed effect (X 2 (1) ¼ 5.70, p ¼ .017). For the model including off-task episodic social cognition and task condition as fixed effects, adding elapsed time improved model fit (X 2 (1) ¼ 12.85, p ¼ 3.37e-4), while adding session, the 2-way interaction between component score and task condition, or the 3-way interaction between component score, task condition and elapsed time did not improve the model fit. This model showed that both the off-task episodic social cognition component score, b ¼ À.77, 95%CI ¼ [À1.52, À.02], SE ¼ .38, t ¼ À2.02, and elapsed time, b ¼ .87, 95%CI ¼ [.40, 1.35], SE ¼ .24, t ¼ 3.61, but not condition, influenced the P3 amplitudes for the NT trials preceding thought probes. Similar to analyses without thought probes, the P3 amplitudes increased with elapsed time. Critically, the model indicated a negative association between P3 amplitudes and component scores for the off-task episodic social cognition that did not vary with task context (Fig 5b and c). The component scores for other patterns of thought had no effect on the P3 amplitudes.
For consistency with prior studies, we also analyzed whether the P3 amplitudes differed between on-and off-task thinking. To do so we added participants' response to the first MDES item "My thoughts were focused on the task I was performing" (Table 1) as a fixed factor to the model including task condition as a fixed effect. This significantly improved the model fit (X 2 (1) ¼ 5.45, p ¼ .020). Adding elapsed time (X 2 (1) ¼ 13.20, p ¼ 2.80e-4) further improved the model fit while neither session, the 2-way interaction between task focus and condition, nor the 3-way interaction between task focus, condition and elapsed time affected the model fit. The results showed that task focus, b ¼ .87, 95%CI ¼ [.05, 1.68], SE ¼ .42, t ¼ 2.08, and elapsed time, b ¼ .88, 95%CI ¼ [.41, 1.35], SE ¼ .24, t ¼ 3.65, but not condition affected the P3 amplitudes immediately preceding the thought probes. The model indicated a positive association between task focus and P3 amplitudes, whereby increased on-task focus was associated with larger P3 amplitudes, in line with earlier reports (Barron et al., 2011;Kam et al., 2021;Smallwood, Beach, et al., 2008). Importantly, this result supports our analysis using PCA that indicates that off-task processing impacts on P3 amplitude in a manner that is similar across contexts. In addition, similar to prior analyses, the P3 amplitudes increased over time.
To rule out the possibility that the association between episodic social cognition and P3 amplitudes were dependent c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 on the choice of number of components, we report results for a PCA with three components in Supplementary material S2. These data supported similar interpretations for the components than our original analysis. Critically, we also replicated the association between P3 amplitude and episodic social cognition.
For consistency with the analyses for P3 amplitudes, we investigated the relationship between task performance and the P3 peak latencies. We found no relationship between accuracy and P3 latencies. However, the RTs were affected by the P3 latencies, b ¼ À.07, 95%CI ¼ [À.14, À4.64], SE ¼ .03, t ¼ À1.97, whereby longer P3 latencies were associated with shorter RTs. Contrary to the P3 amplitude results, this effect was no longer significant when added to the model that included the interaction between task condition and elapsed time as fixed factor, i.e., the model that best predicted RTs.
Finally, we investigated the relationship between thought probes and P3 peak latencies. The component scores for other patterns of thought except for the emotion component had no effect on P3 latencies. We found that adding elapsed time (X 2 (1) ¼ 4.48, p ¼ .034), but not the task condition, session nor the interaction between component score and elapsed time, improved the model fit compared to the model including emotion component score as a fixed effect. This model showed that stronger emotion component scores were associated with longer P3 latencies, b ¼ 7. 67,95%CI ¼ [.19,15.16], SE ¼ 3.82, t ¼ 2.01, and that P3 latencies increased with elapsed time, b ¼ 5. 05,95%CI ¼ [.37,9.73], SE ¼ 2.39, t ¼ 2.11.

Discussion
Individuals prioritize self-generated personally relevant thoughts when external task demands are low and attend externally when external tasks require continual monitoring Smallwood & Schooler, 2015). Attending to topics that are unrelated to the task in hand are associated with reductions in evoked responses to perceptual signals and task-relevant information (Barron et al., 2011;Bozhilova et al., 2021;Kam et al., 2021Kam et al., , 2022Smallwood, Beach, et al., 2008;Villena-Gonz alez et al., 2016). However, it remains unclear whether this pattern is comparable across tasks, a question that is particularly important when considering prior studies using fMRI in this paradigm Turnbull et al., 2019a, b), which identified that regions of dorsal parietal cortex exhibit similar reductions in off-task thought across both conditions, while regions of dorsolateral prefrontal cortex instead show negative correlations with off-task thought in the 1-back condition and a positive correlation in the 0-back condition. The primary aim of our current study, therefore, was to understand whether the observed pattern of reduced P3 during off-task thought is consistent across both tasks or instead varies between tasks. In our study, the non-target events are irrelevant in the 0back condition, allowing periods when attention can be devoted to off-task mental content with no consequence for task performance. In contrast, in the 1-back condition, target responses are based on the previous non-targets indicating that continuous monitoring of the stimulus stream is important for task accuracy. Corroborating earlier reports (Konishi et al., 2017;Sormaz et al., 2018;Turnbull et al., 2019b), we found that when task demands were low, participants reported more thought patterns related to off-task episodic social cognition and mental imagery. By contrast, when task demands were high, participants described more on-task deliberate thoughts.
Importantly, we found that stimulus-evoked parietal P3 responses were attenuated for low relative to high external task demands. Because the low demand task was associated with task-unrelated thoughts, this finding is consistent with prior reports of reduced parietal P3 amplitudes during periods of off-task thought (Barron et al., 2011;Kam et al., 2021;Smallwood, Beach, et al., 2008). When P3 responses were analyzed with respect to trials immediately preceding participants' self-reported thought patterns, off-task episodic social cognition was associated with diminished P3 amplitudes independent of the external task demands. This result suggests that the reduction of the P3 was broadly consistent across both conditions. Prior fMRI studies help understand the likely neural basis of this effect. Studies demonstrate that neural activity in dorsal parietal cortex is associated with ontask thought regardless of the external task demands Turnbull et al., 2019a, b). Since the P3 sources are localized to parietal cortex (Linden, 2005;Polich, 2012), our study builds on these findings by establishing that neural activity in these regions is likely to vary in its correlation with external input in manner that is dependent upon whether attention is internally or externally focused (see also Kam et al., 2021). Our study, therefore, provides direct support for a growing literature implicating parietal cortex in externally focused states.
Collectively, these analyses provide important support for the notion that attentional decoupling may be a necessary feature for the integrity of an internal train of thought (Smallwood, 2013). According to this perspective, it is argued that when we focus attention away from events in the here and now and instead on self-generated material, task evoked responses are decreased because they reflect a source of information that is incompatible with the current train of thought and thus would derail this type of experience. This view is also consistent with a growing body of evidence from fMRI. For example, this paradigm has established that when participants retrieve information during the target trials in the working memory task, and we must attend internally, activity in visual cortex is reduced relative to the 0-back task, when an internal focus in not necessary (Murphy et al., , 2019. More generally, studies have shown that retention, but not encoding during working memory is enhanced by decoupling between regions of transmodal cortex (i.e., the default mode and frontoparietal cortex (Zhang, McNab, Smallwood, & Jefferies, 2022)). Finally, in the context of reading, the experimental induction of a pattern of autobiographical memory retrieval leads to a disruption of the pattern of coupling along the ventral stream that typifies successful reading (Zhang et al., 2022a). Together, with the current evidence, these studies suggest that decoupling from external input may occur when we focus internally, and importantly this effect can at times be beneficial to task performance (i.e., Murphy et al., 2018;Murphy et al., 2019;Zhang et al., 2022a).
Although our study helps disentangle the theoretical interpretation of the reduced evoked neural responses during off-task thought, it leaves open several important questions. First, it remains unclear what consequence the act of introspection, that is necessary for gaining insight into cognition via experience sampling, has on neural processing . Compared to data undisturbed by thought probes, we observed reductions in accuracy and in the P3 amplitudes in the second session of the experiment where thought probes occurred for the first time, but less marked changes in the third session. However, we found that differences in evoked responses between tasks were preserved across sessions with and without thought probes. Although these data do not rule out the possibility that introspection elicits changes in neural activity, they suggest that these do not disrupt the broader neural landscape. Moreover, it is possible that at least part of the effects of experience sampling arise because participants take time to adjust to the probes, because we found more evidence of disruption in the second than the third session of our task. As understanding internal states is increasingly becoming a focus of cognitive neuroscience (Gonzalez-Castillo, Kam, Hoy, & Bandettini, 2021) it is important that paradigms are developed to assess the consequences of experience sampling on brain activity, as well as through the development of indirect markers for these states that can access them without the need for introspection.
Second, our study adds to an emerging body of evidence (Smallwood, Obonsawin, & Reid, 2002;Turnbull et al., 2019a) that indicates that variations in the passage of time alter the balance of internally and externally derived thoughts. In an analysis of how our trial structure was linked to the association between ongoing thought and task relevant indices of brain function, we found that elapsed time after the last target trial, where a manual response was required and attention had to be focused on the task, affected RTs and thought patterns for imagery and off-task episodic social cognition. Both RTs and thoughts related to episodic social cognition decreased with time particularly when external task demands were high, whereas imagery increased with time irrespective of the task demands. Moreover, the amplitudes and peak latencies of the parietal P3 response showed a gradual increase over time. These results indicate a complex pattern of modulation of behavior, brain activity, performance and thought patterns with the passage of time, an interesting question for future work to explore. In terms of ongoing experience, elapsed time impacted on both off-task episodic thought and imagery, neither of which were task relevant components. c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 One possibility, therefore, is that variations in attention with the passage of time is related to the time it takes for the brain to shift attention away from events in the environment to entertain experiences unrelated to the immediate environment. This could potentially relate to the presence of slow fluctuation in activity over time within the mammalian nervous system (e.g., Leopold, Murayama, & Logothetis, 2003) and/or the strategic deployment of attention after task processing because of the variable foreperiod sampling regime we used in our study (e.g., Polzella, Ramsey, & Bower, 1989). It would be fruitful for future studies to develop approaches that can explore the underlying mechanisms behind these changes, for example through variations in task relevant motivation (Seli, Schacter, Risko, & Smilek, 2019), or through the application of attention enhancing drugs, such as modafinil (Ikeda et al., 2017). It is worth noting that while some studies have used short windows (e.g., around 6 s Konu et al., 2020;Sormaz et al., 2018;Turnbull et al., 2019b) others have used windows around 15 or 20 s preceding the probes (Andrillon, Burns, Mackay, Windt, & Tsuchiya, 2021;Baird, Smallwood, Lutz, & Schooler, 2014;Kam et al., 2021;Tusche, Smallwood, Bernhardt, & Singer, 2014). Based on our study, we argue that it is unlikely to be appropriate to assume that the attentional state will remain stationary over an analysis window of this length. Recent work has begun to examine mind wandering episodes using machine learning techniques (Groot et al., 2021;Mittner et al., 2014;Wang et al., 2020;Wang, Poerio, et al., 2018). It may be possible in the future to leverage machine learning approaches to fine tune the appropriate window size for the study of the behavioral and neural correlates of ongoing thought patterns. In lieu of the results of these analyses, our study suggests that future work should utilize a shorter rather than longer window, if possible, when exploring links between brain activity and ongoing thought.
Third, we found that all other components except the component reflecting the emotional quality of the ongoing experience were modulated by external task demands. Along with prior studies (e.g., Turnbull et al., 2019b) we showed that externally demanding tasks reduce the self-generation of personally relevant information and increase a detailed focus on the task, whereas thought patterns reflecting emotional content seem less sensitive to the task context (see Konu et al., 2020). Participants' inability to regulate emotional thought patterns is partly consistent with reports demonstrating that patterns of unpleasant or intrusive thoughts are associated with higher levels of depression (Konu et al., 2021;Ottaviani & Couyoumdjian, 2013) and negative moods (Killingsworth & Gilbert, 2010). Together, these results suggest that while selfgenerated thought may be adaptive and allow the individual to escape the here and now (Mooneyham & Schooler, 2013), it may also have maladaptive consequences in terms of emotional thoughts that are hard to suppress, even when the external task would require so. Broadly consistent with this view are the results of the exploratory analysis examining links with P3 latency. This analysis shows that longer P3 latencies are linked to faster responding, and that shorter latencies occur when participants reported unpleasant intrusive patterns of thought. It is possible, therefore, that highly negative patterns of ongoing thought, compromise the detailed analysis of external information in a way that is different from off-task episodic thought. However, our analysis of P3 latency was exploratory and so we recommend that this pattern should be replicated in a future study before firm conclusions are drawn on the impact of unpleasant patterns of thought on the processing of external input.
In conclusion, our study provides novel evidence confirming the hypotheses that attentional decoupling is a necessary feature of self-generated mental content (Smallwood, 2013). When the external task did not require continuous monitoring, participants engaged more in off-task episodic social cognition and mental imagery, while on-task deliberate thoughts were more common when the external task demanded continuous updating of the incoming information. At the same time, parietal P3 responses decreased for low external task demands suggesting reduced perceptual processing during off-task thoughts. Importantly, however, we found that off-task episodic social cognition was associated with diminished P3 responses independent of the task context. Together, these findings establish P3 as a general feature of an external task related focus and suggest that attentional decoupling from this information is necessary for certain types of self-generated attentional states.

Declarations of competing interest
None.

Data statement
Analysis code and data underlying figures, statistics and main conclusions will be deposited at a repository and made publicly available as of the date of publication. Any additional information required to reanalyse the data reported in this paper is available from the lead contact upon request.

Open practices
The study in this article earned Open Data and Open Materials badges for transparent practices. The analysis codes and data c o r t e x 1 5 9 ( 2 0 2 3 ) 3 9 e5 3 underlying statistics and figures can be found at https:// github.com/jsimola/MW-EEG. Original behavioral and EEG data are stored at https://osf.io/879ex/