The Mind-Wandering Brain: Analyses of behavioural indices of attention lapses

(1)

The Mind-Wandering Brain

Analyses of behavioural indices of attention lapses

Miriam C.L. Maan, Wouter Boekel, & Birte U. Forstmann University of Amsterdam

Department of Psychology October, 2014

(2)

Abstract

Mind wandering refers to a transient shift of attention away from the current task. Mind wandering can be studied on the basis of experience sampling, in which participants provide introspective reports on their degree of task-oriented attention, or by investigating markers in performance on a task. In the Sustained Attention to Response Task (SART), anticipations, omissions, commission errors and reaction time variability (RTcv) all mark an episode of mind wandering.

This study aims to further investigate these behavioural markers by studying how they relate to each other and aims to replicate the findings done in earlier studies (Cheyne et al., 2009). Next, this study will explore whether we can index the direction of attention with an algorithm that inserts probes at specific moments in the natural fluctuation of the RTcv time course. We were able to confirm the interdependence of the markers and could predict their occurrence based on RT. The algorithm could successfully be applied. This framework thereby lays the first steps in finding an objective correlate of mind wandering. We intend to use our results to lay the groundwork for a simultaneous EEG/fMRI/eye tracking experiment in which we aim to predict the occurrence of mind-wandering episodes with high accuracy.

(3)

Introduction

Attention has been one of the most intensely studied topics within psychology and cognitive neuroscience. However, task situations are not always attention demanding, and we often have the experience that our thoughts do not remain on a single topic for a long period of time (Smallwood & Baracacy, 2013, Smallwood & Davies, 2004 in Smallwood & Schooler, 2006; Christoff, Carriere, & Smilek, 2009). This tendency of the mind to drift away was first described in the late '90s by James (1980, in Christoff, 2012) and has been a domain of study in more recent research under the topic of spontaneous thought and mind wandering (Christoff et al., 2004, Smallwood & Schooler, 2006; Christoff, 2012). Despite its high occurrence (it is thought that mind wandering occurs about 30-50% of the time we spend awake (Bastian & Sackur, 2013; Killingsworth & Gilbert, 2010)), research has focussed on this topic only recently (Callard, Smallwood, Golcher, & Margulies, 2013). This was mainly due to theoretical controversy about the term used for mind wandering (Callard, Smallwood, Golcher, & Margulies, 2013). Task-unrelated thought, off-task thought, stimulus-independent thought, or internal thought were differentially used in the literature, and no coherency in the field existed (Randall et al. 2014). However, after the term mind wandering was used in some high impact papers, research on this topic merged and increased tremendously (Smallwood & Schooler, 2006).

Topics of study include research on the time course (Bastian & Sackur, 2013), content (Christoff et al. 2009) and control (Kane et al. 2007) over mind wandering, in relation to brain processes. Knowing which factors trigger mind wandering could help inform people who want to promote it, as can be the case in people who meditate, or who want to prevent themselves from mind wandering, as can be the case in people who are engaged in public safety, such as (truck) drivers or security officers. Even in less dangerous settings this knowledge might be important, such as for people working on the stock market or in the office. By understanding how a mind wandering episode is initiated, the consequences of an attention drift could be limited, as a lack of attention is related to an increased frequency of accidents and errors (Smallwood & Schooler, 2011). However, since mind wandering is also related to creative thinking, this information could be useful to promote episodes of mind wandering. Furthermore, knowing what happens in the brain when people make the transition from being focused on the task to being focussed on their own thoughts, or vice versa, could give us a better understanding of attention processes and consciousness in general, and would help us to better understand the diagnostics of for example depression, in which mind wandering seems a common symptom (Smallwood & Schooler, 2006).

(6)

However, despite its high occurrence and relevance to society, mind wandering has become a topic of study only recently (Smallwood & Schooler, 2006). One reason is that studying mind wandering is difficult. Not only does the identification of mind wandering depend on the individual's ability to monitor its attention, it also often occurs spontaneously and/or without awareness (Smallwood & Schooler, 2006). In early mind wandering studies, this was solved by using thought probes. These entailed asking for the content of thoughts at random times during the day (Smallwood et al., 2004). This way, results were not biased by the ability of people to monitor their attention. This experience sampling method is now a widely used method in mind wandering research, and many studies have shown the validity of this method (Christoff, Carriere, & Smilek, 2009; Smallwood et al., 2008; Kam & Handy, 2014).

Yet, not only the content, but also the duration and frequency of mind wandering are of interest. Experience sampling methods do not provide enough means to study these topics and therefore, researchers started using a task in which the duration of a mind wandering episode is related to frequency of performance errors and reaction times: the Sustained Attention to Response Task (SART; Robertson et al., 1997).

The SART is a go/no-go task, with an infrequent no-go-trial. In the SART, individuals have to respond with a button press to a series of digits (go-trials) and have to withhold this response on a target item (no-go-trial). The digit presentation is often short (200-400ms) and is followed by a blank screen in which a participant can still respond (inter-trial-interval; IRI). This IRI is often longer than the stimulus presentation (800ms-1000ms) and can have a fixed or a random interval (jitter).

The theory behind the SART is based on the idea that performance on the SART requires attentional resources. According to the executive function theory of mind wandering, cognitive resources are limited and, as both task related thoughts and mind wandering depend on the same resources, mind wandering interferes with task performance (Smallwood & Schooler, 2006). No-go errors (responding to a no go-trial, in other studies also known as commission errors (CE)) are seen as a failure of sustained attention as they represent a failure of response inhibition (Kam & Handy, 2014), (Cheyne, Carriere, & Smilek, 2009). Episodes in which participants make a lot of these errors, therefore, are known to reflect an episode of mind wandering.

Mind wandering also impairs working memory updating, further disrupting performance, since external information is not updated, leading to repetitive behaviour and misses of trials (omissions) (Kam & Handy, 2014). For that reason, series of short reaction

(7)

times (anticipations) on trials, or the miss of a trial, both reflect an episode of mind wandering.

SART performance shows several types of behaviour that either lead or follow a commission error. It is important to know about these specific behavioural markers since they provide an objective measure for the subjective experience of mind wandering. This is of essence in this domain of study because, even though people are able to use self-report measures to index their direction of attention, the subjective experience of mind wandering is variable, shows a lot of temporal variation and sometimes occurs without awareness (Koyama et al., 2003). Furthermore, by having an objective measure of mind wandering in the form of performance on a task, this type of measure could be used online to predict when an individual enters a state of mind wandering, before he or she even notices this shift him/herself. This would be a more powerful approach to predict occurrences of mind wandering, since it allows for possibilities of intervention, something that cannot be done using post-hoc self-report measures. The central purpose of the current study is to find such an objective quantification of mind wandering, based on behavior on the SART.

We will next review the theory about different states of (in) attention and address their relation to the behavioral markers. This theoretic framework will inform us on what type of behavior could possibly be tracked online and could represent an episode of mind wandering.

1.1.

Behavioural markers of mind wandering

As described above, performance on the SART provides several types of behaviour, seen in reaction time and frequency of errors that can be considered as a behavioural example of mind wandering (Smallwood et al., 2004). These behavioural indices as proposed by Cheyne et al. (2009) include variability in response times (RTcv, cv = SD/mean), anticipations, and omissions on go-trials. The behavioural markers are linked to different states of attentional disengagement.

1.1.1. Reaction time variability

Increased variability in RT is linked to the first state of task disengagement, occurent task inattention (Cheyne, Carriere, & Smilek, 2009). During State 1, the mind wanders only briefly, there is only a transient disengagement of attention away from the dynamic features of the task. Individuals can be aware of their attention drifting away and can experience a form of dual consciousness (Hester et al., 2005; Schooler, 2002, Schooler et al., 2005; Smallwood et al., 2007 in Cheyne, Carriere, & Smilek, 2009). State 1 can be disrupted by

(8)

errors and near misses, but action slips can still be detected before they become real errors. This results in either extremely fast or very slow responses, causing an increased variability in RT (Johnson et al., 2007 in Cheyne et al. 2009). Therefore, the coefficient of variability in RT is thought to be an index of State 1.

Increased RTcv has been found as an indicator of mind wandering in other studies (e.g. Bastian & Sackur, 2013). Bastian and Sackur (2013) presented a thorough analysis on RTcv and found that episodes of mind wandering were most likely to occur at peaks in RT variability.

1.1.2. Anticipations

The second task disengagement state, generic task inattention, is linked to anticipations. Anticipations are represented by very fast reaction times (<100ms) on go-trials. Within this short window of time, sensory processing of the stimulus did not occur, but rather reflects a speed-accuracy trade-off, since the probability of being accurate is high (around 90%) regardless of the speed of response. Anticipations are mainly encouraged to occur when the inter trial interval is fixed, due to the repetitive behaviour (Cheyne, Carriere, & Smilek, 2009).

State 2 is linked to automatic responding, as the individual remains aware of the general task environment, but has lost its sensitivity to moment-to-moment task variations (Cheyne, Carriere, & Smilek, 2009). People are often less aware of this type of disengagement. However they can become aware when a commission error is made; attention is then drawn inwards to process the error itself. This error-induced-task-relevant processing can further impair task performance; anticipations therefore are known to induce commission errors (Cheyne, Carriere, & Smilek, 2009).

1.1.3. Omissions

The third state, response disengagement, is linked to a complete decoupling from conscious processing, as attention is directed to inner thoughts (Cheyne, Carriere, & Smilek, 2009). During this state, the individual is unresponsive to almost all features of a task environment, except the most intrusive. The most common errors during this state are omissions. Omissions reflect a failure to respond to go stimuli: participants are unable to respond to the stimuli within the right temporal window (Cheyne, Carriere, & Smilek, 2009).

(9)

The three states of attentional disengagement are sequentially related. Individuals will move from one state to another state sequentially, and are more biased to enter one state if they are already in one of the states. Following this idea, Cheyne et al. (2009) showed that anticipations and omissions occur in succession, and that commission errors are followed by an increased probability of anticipations and omissions. Lastly, the researchers showed that RT speeding precede commission errors and RT slowing precede omissions (Manly et al., 1999 and Robertson et al., 1997 and Cheyne et al., 2009). In Experiment 1, we will try to replicate these findings.

Anticipations, commission errors, omissions and reaction time variability can serve as critical end and start points for mind wandering, especially since they can be studied on a stimulus-by-stimulus-basis (Koyama et al., 2003). However, although Cheyne et al. (2009) linked SART performance to general measures of mind wandering (MAAS-LO, ARCES), they did not assess mind wandering on a subjective basis during the SART. Therefore, it is unknown whether these errors also occur at the exact moment of a mind wandering episode. Online experience sampling methods could be used to gain more insight in the temporal variations of task engagement/disengagement and the relations to behaviour on the SART.

To assess the predictions made by Cheyne et al. (2009) we analysed a data set of the SART obtained in our lab which included thought probes at a pseudorandom interval, during which participants reported on the direction of their attention, using a 5-point scale (on-task to off-task). We show that all behavioural markers (anticipations, omissions, commission errors and RTcv) are indicative of an episode of mind wandering and therefore would be linked to the off-task probe category (cat 1).

In addition, the data consisted of blocks with and without jitter inserted between trials, which allowed us to test the hypothesis that monotonousness (i.e. no-jitter) of a task enhances the frequency of mind wandering.

Based on the findings from this experiment, and the results of Bastian and Sackur (2013) showing that continuous tracking of RTcv is an efficient way to detect mind wandering, we performed a second experiment in which we further investigated the link between mind wandering and RTcv, by developing an RTcv-tracking algorithm which inserts thought-probes in the SART online, at peaks of the RTcv time course. We show that this classification could be extended by also using decreased variability in RT to conceptualize an on task state in Experiment 3.

(10)

Finally, we will review some literature on the neural correlates of mind wandering and propose a simultaneous EEG/fMRI/eye tracking experiment in which we use the results from the three experiments regarding behavioral markers of mind-wandering, in conjunction with hypotheses derived from the literature with regard to the EEG, fMRI, and eye-tracking signal, to attempt to predict mind-wandering with high accuracy. This would be the first study which traces mind wandering in an online manner based on behavior of the participant.

2.

Experiment 1: Replication of Cheyne et al. (2009)

2.1.

Method

2.1.1. Participants

A total of 25 participants participated in the experiment in exchange for money or extra course credits (6 males; mean age = 23.14 years, S.D. = 4.28), 22 were right handed. Participants filled in an informed consent prior to the study and the study was approved by the ethics committee of the University of Amsterdam.

1.1.1. Stimuli and procedure

To assess the behavioural indices of mind wandering, participants performed the Sustained Attention to Response Task (SART; Robertson et al., 1997). Participants were instructed to respond to all the digits, except the digit 3. A response was valid if it was conducted before the next digit would appear on the screen. The SART was programmed in Presentation 16.5.

The SART procedure consisted of 2 sessions (jitter and no jitter) each containing 720 trials per session. The numbers 1-9 were presented 80 times each in the two sessions. During a trial, a fixation cross appeared on the screen for 250 ms, followed by a random number for 250 ms, followed by a black screen, with a fixed duration in the no jitter condition (900 ms) and a variable duration in the jitter condition (600 ms plus a random number drawn from (0,600). This resulted in average trial duration of 1400 ms.

Each session was preceded by a practice session of one minute which had jitters depending on the following session. The participant was given the option to practise again for the second session if he/she felt not ready.

Participants were instructed to report on the direction of their attention using a 5-point Likert scale, ranging from completely off-task to completely on-task.

(11)

Per session, thought probes were given at a random interval with an average of 36 trials (min 28, max 44) between probes, four of which were no go-trials. Participants had five seconds to select the appropriate answer on the probe.

To asses attention lapses in daily life, participants completed the Mindful Attention and Awareness Scale (MAAS). The MAAS consists of 15 items in which participants had to answer a collection of statements on a 6-point scale (1 = almost always, 2 = very frequently, 3 = somewhat frequently, 4 = somewhat infrequently, 5 = very infrequently, 6 = almost never}). The MAAS measures inattention (e.g. “I find myself preoccupied with the future or the past”) and the tendency for meta-awareness (e.g. ”I do jobs or tasks automatically, without being aware of what I am doing.”). The mean of these 15 items is calculated and reflects dispositional mindfulness and mind wandering (Brown, & Ryan, 2003).

1.1.1. Outlier rejection

In all the analyses described in Experiment 1, 2 and 3, we identified outliers as observations which differed more than 2.5 SD from its group mean.

1.1.2. Statistical inference

Bayesian statistics was used for statistical inference. Bayesian statistics are used to quantify evidence in favour of a hypothesis, by providing the probability for the data to have occurred under that hypothesis (Eguchi, 2008). For example, to investigate whether a correlation exists between two variables, two hypotheses are generated; the null hypotheses (r = 0) and the alternative hypotheses (r != 0). Bayes factors (BF01/BF10) represent the likelihood of the data

under these hypotheses. If the BF01 > 1 there is more evidence for the null hypotheses, and if

BF01 < 1 there is more evidence for the alternative hypotheses. This is because BF01 are

related; BF10=1/BF01 and BF01=1/BF10.

Interpretation of the strength of evidence against H0 can be scaled in multiple ways.

(12)

Table 1

Scaling method of Bayes factors as proposed by Jeffreys (1961) Bayes factor BF10 Interpretation

> 100 Extreme evidence for H1

30 100 Very strong evidence for H1

10 30 Strong evidence for H1

3 10 Moderate evidence for H1

1 3 Anecdotal evidence for H1

1 No evidence

1/3 1 Anecdotal evidence for H0

1/10 1/3 Moderate evidence for H0

1/30 1/10 Strong evidence for H0

1/100 1/30 Very strong evidence for H0

< 1/100 Extreme evidence for H0

An advantage of using Bayesian statistics over frequentist statistics is that Bayesian statistics provide a way of combining new evidence with prior beliefs, through the application of Bayes’ rule. This is in contrast with frequentists statistics, which do not take prior beliefs into account, but relies only on the evidence as a whole. Furthermore, Bayesian statistics do not overestimate the evidence against the null hypothesis, something that is more likely to occur when using frequentist statistics. (Edwards, Lindman, & Savage, 1963; Sellke, Bayarri, & Berger, 2001; Wetzels, Matzke, Lee, Rouder, Iverson et al., 2011).

1.1.

Results

1.1.1. Attention reports

On average across the two sessions, 7.9% of the occasions when a probe was presented, participants chose category 1, the extreme "off-task" probe answer, to classify their direction of attention (jitter = 7%, no jitter = 8.8%). For 16.8% of the probes, participants answered that they were "on-task" (jitter = 17.4%, no jitter = 16.2%). Category 2, 3 and 4 were chosen in 21.9% (jitter = 22.2%, no jitter = 21.6%); 29.3% (jitter = 30.0%, no jitter = 28.6%); and 24.1% (jitter = 23.4%, no jitter = 24.8%) of the time. Figure 1 shows a histogram of the probe answers.

(13)

Figure 1

Histogram of probe answers for the jitter and the no jitter condition combined (N=25)

1.1.2. The effect of jitter on mind wandering

We began our analyses by testing whether jitter had an effect on the rate of occurrence of the three behavioural markers (anticipations, omissions, commission errors) and the variability in RT (RTcv). We hypothesised that the frequency of mind wandering is increased in the no jitter condition, due to its monotonous nature, and that this condition therefore should contain more anticipations, omissions and commission errors. We further hypothesized that the variability in RT should be higher in the no-jitter condition, as literature shows mind wandering is related to an increased variability in response times.

The amount of anticipations, omissions, and commission errors was computed per participant over all the trials. RTcv was calculated over all the go-trials by dividing the SD by the mean RT of the go-trials. A Bayesian t-test was used for statistical inference.

Our analyses showed moderate evidence for the data to be likely under the null hypotheses, indicating that the amount of omissions was similar in the jitter and the no jitter condition (BF01= 7.14). Our analyses showed ambiguous Bayes factors with regard to the

hypothesis concerning the amount of anticipations and commission errors in the jitter and no jitter condition (BF01= 1.47, BF10= 1.58). The null hypotheses concerning the RTcv effect in

the jitter and no jitter condition was more strongly supported than the alternative hypotheses (BF01= 10.0), indicating that the RTcv was similar in both conditions.

Therefore, we conclude that it is possible to assume that jitter has no major effect on the frequency/duration of mind wandering as seen in the amount of omissions, anticipations

(14)

and commission errors made, and the RTcv effect. This allowed us to combine the two data sets in subsequent analyses.

1.1.3. Relations between performance errors and mind wandering

We continued our analyses by investigating the relations between performance errors (anticipations, omissions and commission errors) and mind wandering. As shown in the literature, off task probe reports should be preceded by an increased rate of anticipations, omissions and commission errors (Cheyne et al. 2009). We also were interested in the RTcv effect described by Bastian and Sackur (2013) and Cheyne et al. (2009) and wanted to check whether an increased RTcv was indicative of a mind wandering report.

For each participant, we dichotomized the answers on the thought probes. This decision was made based on visual inspection of the histograms of the probe answers. We decided to only use the extreme probe answers of the participants. Participants had to have at least two answers in a probe category before we included the category in our analyses. For example, when participant “X” has one answer on 1, eight answers on 2, six answers on 3, twelve answers on 4, and fifteen answers on 5, category 2 was used for the off-task condition and 5 for the on-task condition. While somewhat arbitrary, this approach allowed us to dichotomize the data while retaining individual variability.

In order to determine whether high RTcv predicts a mind wandering episode, we examined the RTcv in the eight trials preceding a probe. Similar to the data exclusion method of Cheyne et al. (2009) we first omitted answers on probes that were preceded by one or more no-go-trial in the eight trials preceding a probe. However, the small amount of data left implied we would not have enough power to detect an effect. Another method of exclusion, used by Bastian & Sackur (2013), discards a probe if the eight trails preceding a probe contain a no-go-trial or an omission. With this method of exclusion, however, the likelihood of finding an effect was again low, due to the small amount of data left after exclusion. Therefore we chose a more lenient method of exclusion and omitted probes which had more than one no-go trail in the eight trials preceding a probe. Bayesian t-tests were used for statistical inference.

Bayesian t-test statistics showed moderate support for the alternative hypotheses, indicating that omissions occur more often in blocks followed by an off-task probe answer versus blocks followed by an on-task probe answer (BF10= 3.29). We found anecdotal

(15)

answers, compared to on-task probe answers (BF10= 2.59). Furthermore, Bayes factors

showed extreme evidence in favour of the alternative hypotheses, supporting the idea that the amount of commission errors is increased in off-task blocks, compared to on-task blocks (BF10= 1.45e+7). Lastly, the data extremely supports the hypotheses that RTcv is higher

preceding an off-task-, compared to an on-task answer (BF10= 1124).

To conclude, these analyses show that performance errors are indeed linked to mind wandering, as was found in the study of Cheyne et al. (2009). The amount of commission errors and an increased RT variability are the most indicative of a mind wandering episode.

1.1.4. Relations between RT and performance errors

Next, we aimed to assess some other predictions made by Cheyne et al., (2009) about the relationships among the task errors and the effect of reaction time on commission errors, anticipations and omissions. They will be described in the next section.

1.1.4.1.

Commission errors

First, Cheyne et al. (2009) have shown that RTs preceding commission errors are all below baseline RT (RT speeding), compared to RTs before correct inhibitions. We tested whether the mean RT of the trial preceding a commission error was decreased compared to the mean RT of a trial preceding a correct inhibition. To do so, a Bayesian t-test was used for statistical inference.

We found extreme evidence in favour of our alternative hypotheses (BF10=

9.497e+91), supporting the hypotheses that RT speeding on t-1 is indicative of a commission error. This is in accordance with Cheyne et al. (2009).

The mean RT preceding and following commission errors and correct inhibitions is shown in Fig. 2, the mean RT on t-1 for a commission error is faster, than the mean RT for a correct trial.

In Fig. 2, another result of Cheyne et al. (2009), namely one-trial speeding of RT after a correct inhibition, can be observed. Although never tested, this is probably a result of the experimental paradigm. No-go trials could not be directly followed by another no-go trial, and therefore participants could already execute a response on the t+1 trial, before directly monitoring the stimulus.

(16)

Figure 2

Mean RT of the four trials before and after a target trial showing RT speeding before a commission error

1.1.4.2.

Anticipations

Next, we analysed the mean RT before anticipation with the aim to replicate the findings done by Cheyne et al. (2009). In their study, they found a near-mirror image around an anticipation, showing linear decline in RTs around an anticipation. We hypothesized that the same effect was also present in our data-set.

Fig. 3 shows the mean RT of the four trials immediately preceding and following an anticipation on t0. The image shows an near-mirror image, as predicted by Cheyne et al. (2009).

Figure 3

(17)

Although we never tested the hypotheses that the RTs preceding and following an anticipation are below the mean RT, they seem to be rather fast. We hypothesized that this could probably be due to an increased probability of anticipations occurring in the surrounding trials. This idea was never tested by Cheyne et al. (2009), even though they tested for the probability of omissions around an anticipation.

To assess the conditional probabilities of anticipations in the four trails preceding and following anticipations, a Bayesian t-test was used for statistical inference on all the trial intervals. We hypothesised the probability of an anticipation to increase towards t0. According to this hypothesis, the probability of an anticipation is higher for 1 compared to t-2 or t-3.

Fig. 5 shows the mean probabilities of anticipations and the Bayes factors for the four trials preceding and four trials following an anticipation on t0..

Figure 5

Probability of an anticipation in the four trials before and after an anticipation. Bayesfactors included.

Bayesian inference indicated extreme evidence in favour of the alternative hypotheses stating that the probability of an anticipation is higher in the t-1 compared to the t-2 trial (BF10=

9.75e+4), t-1 compared to t-3 trial (BF10= 3.48e+11), and t-1 compared to t-4 trial (BF10=

2.10e+15).

We found anecdotal evidence for our alternative hypotheses, stating that there is a difference in the amount of anticipations between t-2 and t-3 (BF10= 1.14), and found

(18)

between t-2 and t-4 (BF10= 24.54). We found moderate evidence in favour of the null

hypotheses, supporting the idea that the amount of anticipations in t-3 and t-4 is similar (BF01= 9.09).

From these analysis, we conclude that the closer to t0, the evidence for a difference in the amount of anticipations is stronger.

Next, we compared the probability of anticipations in the trials directly following an anticipation with the trials next to it. We found extreme evidence in favour of the hypotheses stating that the probability of anticipations was increased in the trial directly following an anticipation, compared to the t+2, t+3 and t+4 trial (BF10= 8.45e+3, BF10=6.56e+11, BF10=

2.03e+13). We found moderate evidence for the hypotheses stating that there was a difference in the probability of an anticipation between t+2 and t+3 (BF10=5.05), and found evidence in

favour of the null hypotheses stating that the probability of an anticipation was similar for t+3 and t+4 (BF01=18.18). For the hypotheses concerning the difference in anticipations between

t+2 and t+4 was found strong evidence; they differ in the probability of anticipations (BF10=

19.47).

Overall it seems that the trials immediately preceding and following anticipations, all show an increased probability of an anticipation, the closer they come to t0. Therefore we conclude that anticipations are clustered. This finding was never reported by Cheyne et al. (2009), and adds to the current knowledge on the topic of interest. A series of anticipations could indicate a fundamental form of action slip, in which there is no moment-to-moment-task relevant stimulus monitoring during a short period of time. This shows a qualitative different mode of processing, as predicted by the three state model of inattention (Cheyne et al., 2009).

1.1.5. Omissions

In addition, we were interested in the mean RT preceding and following an omission on t0. Cheyne et al. (2009) found that there was slowing of reaction times on the one trial preceding an omission. To assess whether slowing of RTs occurred on the t-1 trial, a Bayesian t-test was used to test for a difference in mean RT for all the trial lags. It was hypothesized that the mean RT of t-1 should be higher compared to the other trial lags.

(19)

Figure 6

Mean RT of the four trials preceding and following an omission

We found extreme evidence in favour of the alternative hypotheses stating that the mean RT in t-1 is higher than the mean RT in t+1 (BF10= 2.03e+4). This also confirms

another prediction made by Cheyne et al. (2009), namely that there is a decrease in RT in the trial following t0. We found anecdotal evidence for our hypothesis stating that the mean RT of t-1 is higher compared to t-2 and t-3 (BF10 = 2.53, BF10 = 1.706) and moderate evidence for

the hypothesis stating that the mean RT on t-1 differed from the t-4 trial (BF10=3.96). We

found moderate evidence in favour of the null hypotheis stating there was no difference in RT between t-2 and t-3 (BF01 = 13.16) and t-2 and t-4 (BF01 = 8.62) and t-3 and t-4 (BF01 = 6.54).

Therefore, we conclude that our data is in accordance with Cheyne et al (2009), the response on t-1 is generally slower, compared to t-2,t-3, and t-4, which fall in the same range of response times. There is a clear transition from RTs in the normal range to a very long RT followed by an omission. This makes increased RT’s are indicators of omissions.

1.1.6. Omissions, anticipations and commission errors

Another set of predictions made by Cheyne et al. (2009) that we tested, concerned the occurrence of the behavioural indices after a commission error. Specifically, we predicted that the amount of anticipations and omissions in the trial directly following a commission error is increased, compared to the amount of anticipations and omissions after a correct inhibition. The data were analysed with a Bayesian t-test.

(20)

We found extreme evidence supporting the alternative hypotheses, indicating that the amount of anticipations is increased after a commission error (BF10= 108). Furthermore, we

found moderate support for the alternative hypotheses stating that the amount of omissions is increased after a commission error, compared to a correct inhibition (BF10= 7.74). These

results confirm the results of Cheyne et al. (2009).

A possible explanation for the first effect could be that during a series of anticipations, no sensory processing of the stimulus takes place, therefore, the motor response is already executed, before the participant processes that there is a no-go-trial to which they should not respond. A commission error therefore could correspond to an anticipation on a no-go-trial. This idea would be supported by our earlier analyses in which we found that anticipations often occur in succession.

On the other hand, both effects could also be explained in terms of reactive mind wandering; when an error (CE) is made, participants are preoccupied with their thoughts to process the error, which could induce further errors, leading to anticipations and omissions in the following trial.

1.1.7. Anticipations and Omissions

Lastly, we wanted to assess the sequential associations of anticipations and omissions. We had two reasons to be interested in this effect. First, to confirm the study of Cheyne et al. (2009). They showed that the probability of an anticipation/omission was increased in the preceding and following trials, if t0 was either an anticipation or an omission. Second, we were interested in whether the probability of an anticipation was specifically increased in the t+1 trial, if t0 was an omission. If this effect was indeed present, we are intended to believe that anticipations on t+1, could also reflect the missed response (omission) on t0.

Anticipations surrounding an omission: We used a Bayesian t-test to assess whether the chance of having an anticipation after an omission is increased versus an anticipation occurring before an omission, and compared the conditional probabilities of the t+1 and t-1 trial.. We found extreme support for our hypothesis (BF10= 20259.76). Fig. 7 shows the mean

probabilities of an anticipation occurring in the trials around an omission on t0

Earlier analyses in this study showed that anticipations seem to cluster. If the anticipation in t+1 is indeed an anticipation, and not a response on the t0 trial, than it should cluster with the t+2,3,4 trials. The probability of an anticipation should therefore also be increased in these trials. We used a Bayesian t-test for statistical inference.

(21)

Data showed moderate to strong support for the null hypotheses indicating that the probabilities for anticipations occurring in the t+2, t+3 and t+4 were similar (BF01 (t+2 vs.

t+3) = 10.75, BF01 (t+2 vs. t+4) = 8.06, BF01 (t+3 vs. t+4) = 8.06). However, data supported

the alternative hypotheses stating that the probability of an anticipation was increased in the t+1 trial compared to the t+2, t+3 and t+4 trial after an omission (BF10= 5.25, BF10 = 5.25,

BF10= 12.64). This indicates that the probability of an anticipation is specifically increased

for the t+1 trial, indicating that it does not reflect an anticipation, but rather the response on the t0 trial.

We could not confirm the finding of Cheyne et al (2009) stating that the probability of an anticipation is also increased in the preceding trials when an omission occurs on t0, representing a state transition. The differences in the probabilities of having an anticipation in the t-1, t-2, t-3 and t-4 trials did not differ (BF01 (t-1 vs. t-2) = 21.32, BF01 (t-1 vs. t-3) =

14.16, BF01 (t-1 vs. t-4), = 15.95; BF01 (t-2 vs. t-3) = 5.56; BF01 (t-2 vs. t-4) = 6.76; BF01 (t-3

vs. t-4) = 12.30). Given that the change of having an anticipation following an omission is specifically increased in the t+1 trial, we conclude that this response is probably not an anticipation, but rather reflects a delayed response on the t0.

Figure 7

(22)

Omissions surrounding an anticipation: Another prediction we wanted to assess, concerned the conditional probabilities of omissions around an anticipation on t0. Cheyne et al. (2009) found the conditional probability of omissions to be increased in the surrounding trials, showing a near mirror-image pattern around an anticipation on t0.

Figure 8

Mean probabilities of omissions in the trials preceding and following an anticipation

Fig. 8 shows the mean conditional probabilities of the omissions for four trials preceding and four trials following an anticipation on t0.

Based on Cheyne et al. (2009) we hypothesized that in the trials immediately preceding (t-1) and following (t+(t-1) anticipations, the probability of an omission should be increased. We tested this hypothesis by comparing the difference in the amount of omissions over the surrounding trials and used Bayesian t-tests to test the strength of the effect.

We found extreme evidence in favour of our null hypotheses stating that there is no difference in the amount of omissions between all the trials preceding an anticipation (BF01

(t+1 vs. t+2) = 8.81; BF01 (t+1 vs. t+3) = 22.23; BF01 (t+1 vs. t+4) = 29.32; BF01 (t+2 vs. t+3)

= 46.51; BF01 (t+2 vs. t+4) = 55.56; BF01 (t+3 vs. t+4) = 33.90); and following an anticipation

(BF01 (t-1 vs. t-2) = 64.94; BF01 (t-1 vs. t-3) = 95.23; BF01 (t-1 vs. t-4) = 97.09; BF01 (2 vs.

t-3) = 50.76; BF01 (t-2 vs. t-4) = 52.35; BF01 (t-3 vs. t-4) = 27.74).

Our data do not fully support the results of Cheyne et al. (2009), who predicted that anticipations and omissions occur in succession. We could only find evidence for an increased

(23)

amount of anticipations on t+1 after an omission, and an increased amount of omissions on t+1 after an anticipation. If states indeed occur in succession, then the chance of having an anticipation/omission, should also be higher in the t-1 trial.

There could be two possible reasons why we could not replicate Cheyne et al. (2009) in their findings; (a) omissions and anticipations occur in succession as predicted by Cheyne et al. (2009), but we did not find it in our dataset; (b) our data is correct in the sense that the amount of anticipations/omissions is only increased in t+1 trial, and an alternative explanation should be found instead of the qualitative state transition model.

We have reasons to believe explanation (b) is correct. If anticipations and omissions occur in succession, then the amount of anticipations should also be increased in the t+2,3,4- trials, as anticipations seem to cluster. This was however not the case in our dataset, and therefore we hypothesize that the t+1 anticipation actually reflects a response on t0. However, Cheyne et al. (2009) never reported clustering of anticipations and therefore we cannot fully exclude the first explanation (a). Why the chance of having an omission after an anticipation is higher in the t+1, could be due to the experimental paradigm. In the paradigm we used, participants could press a button within a certain window, but only the first button press is registered as a response. It could be that participants had an anticipation in the t0, and then pressed again in the same t0 window, resulting in no response in the t+1 window. Our data did not allow us to test this hypotheses, so this result remains open for discussion.

1.2.

Summary of Experiment 1

To summarize, the present results seem in general to be consistent with the hypotheses made based on the study of Cheyne et al. (2009). The four behavioural indices, anticipations, omissions, commission errors and RTcv are all associated with reports of mind wandering. Furthermore we could indicate that reaction times slow down before an omission, making a long RT a specific indicator of an omission, and therefore mind wandering. We also found that RTs are faster before a commission error compared to a correctly inhibited response. This finding is consistent with prior findings and interpretations of the attention states named by Cheyne et al. (2009). The result that commission errors are likely to be followed by an omission or an anticipation is striking, since they represent opposite extremes of the RT. However, this indicates that there is bi-directionality between mind-wandering and attention related errors. It might be possible that reactive-mind wandering reflects attempts to reengage with the task, but that this has a brief adverse effect on performance in the case of

(24)

anticipations. The collapse represents a cessation of task analysis and, in the case of omissions, a failure of response preparation.

We furthermore showed that anticipations are often clustered, a finding that has not been found in other studies before. Lastly, anticipations occur more often in the t+1 trial if t0 is an omission, and we found similar results for the probability of an omission on t+1, if t0 was an anticipation. We could, however, not explain this effect.

Besides confirming the study of Cheyne et al. (2009) and exploring the relations of the behavioural markers, we wanted to investigate whether we could find a technique that could be a robust way of inferring the direction of attention of a participant, online. Therefore, in this study we will next propose a technique to assess the subjective experience of mind wandering exactly at the moment where performance errors are made. Before, these two types of behaviour (performance errors and thought reports) were seen as two separate indexes of mind wandering. We hypothesise RTcv to be an ideal marker for the classification of mind wandering.

First, a peak in RTcv is linked to a type of attentional disengagement, occurent task inattention, a state of inattention linked to brief and unstable lapses of attention, influenced by errors and near misses. Second, increased variability in RT has been found to be an indicator of mind wandering in earlier analyses done by Cheyne et al. (2009, 2012) and Bastian & Sackur (2013). This relation between high RTcv and mind wandering is confirmed in the present study. Third, anticipations and omissions occur in succession. Since omissions are preceded by an increased RT, this series of events can lead to increased variability in RT, further linking RTcv to other types of attentional disengagement and mind wandering. Therefore, online tracking of RTcv would be a useful paradigm for tracing drifts in the level of attention. Other advantages of using RTcv is that it can be tracked consistently on a trial by trial basis, something that is harder for the other behavioural markers, increasing the power of this type of measurement.

We will show that by tracking the reaction time variability during the SART, increased variability of RT can be linked to an episode of mind wandering as reported by the participant. This study thereby follows the trend set by Bastian and Sackur (2013), of having a continuous tracking of RTcv as a way to detect mind wandering. Having an online marker of mind wandering based on behaviour, instead of introspection, is the first step in making the study of mind wandering more objective and less dependent on the individual's ability to monitor its attention and awareness of their drift.

(25)

We performed a second experiment in which we further investigated the link between mind wandering and RTcv, by developing an RTcv-tracking algorithm which inserts thought-probes in the SART online, at peaks of the RTcv time course.

1.

Experiment 2: RTcv-tracking algorithm

1.1.

Method

A total of 10 participants (1 male; mean age 24,1 year, S.D. = 3,42) participated in the experiment in exchange for money (10 euro). They were all right handed, and had normal to corrected-to-normal vision. Participants filled in an informed consent prior to the experimental procedure. The study was approved by the ethics committee of the psychology department of the University of Amsterdam.

All participants completed the Sustained Attention to Response Task (SART). The paradigm was similar as the one described in Experiment 1, except that the stimulus was now visible for 400 ms (vs. 250 ms) and that participants had 6 seconds (vs. 5 seconds) to respond to the probe question. Also, we used a two-point Likert scale for the probe question. This was done because dichotomization of Likert data resulted in loss of data in the earlier analyzed experiments.

Furthermore, before the start of a session, participants were given the definition of the two attention states, based on literature (Smallwood & Schooler, 2011; Cheyne et al. 2009; Bastian & Sackur, 2013). On-task was described as: "Being on-task means that your attention is focussed on the task. Your brain is processing task related information. You are not distracted by environmental or internal stimuli.".

Off-task was described as; "Being off-task means that your attention is not primarily focussed on the task, but rather is focussed on internal information, such as memories or personal relevant information. When you are in an off-task state, you don't process sensory information deeply, you might be aware that there are stimuli appearing on the screen, but don't process their higher meaning (I should not respond to a 3)."

We developed an algorithm which aimed to present a probe during the SART based on natural fluctuations of the RTcv time course. The probe either had to appear when a local

(26)

maximum in RTcv was reached (peak probe) or when the interval between two probes became too long (max probe).

More specifically, the task of the algorithm was to analyze two variables online: (a) the value of the RT, and (b) the amount of trials passed since a probe was presented (Ntrials).

Based on the RT data, the algorithm had to calculate the RTcv over a window of eight trials (i.e, at every new trial the RTcv was computed by taking the standard deviat ion of the RT of the previous 8 trials, and dividing it by their mean RT), and had to put this RTcv-value in a vector. The RTcv-vector had to be arranged in size, and had to be updated and re-arranged every new trial. The aim of these tasks is to extract the RTcv-value that is equivalent to the 15th percentile of the current RTcv-vector,which we will refer to as the threshold value (RTcvthreshold). This threshold value was determined after the first 20 trials, and was updated

every trial.

Another variable that had to be analyzed online was (c) whether the threshold was crossed (C0/1/2). This parameter could have three values (0, 1, 2) , since the parameter was

reset after the threshold was crossed a second time,.

Based on the current value of the RTcv, the RTcvthreshold, Ntrials and C0/1/2, the

algorithm determined whether or not to present a probe.

A probe was presented if two requirements are met. if the minimum amount of trials passed since the last probe is 15 (Ntrials = 15), the algorithm presents a probe if the threshold is

crossed the second time (C = 2) . We will refer to this type of probe as a peak probe. Second, the algorithm presents a probe if the time in between probes has reached a maximum of 45 trials. We will refer to this type of probe as a max probe.

Fig. 9 shows an example of the performance data (RT) and the value of the RTcv (red) and RTcv threshold (green) based on this data. If the threshold is crossed a second time, a peak probe (orange) is presented. If the amount of trials is too long (45) a max probe is presented (blue).

(27)

Figure 9

Example of RTs on trials, RTcv, and a threshold determined by the algorithm. The orange arrow represents a peak probe, the blue one a max probe

After the two sessions of the SART, the participants filled in a small questionnaire in order to assess their awareness of the algorithm.

1.2.

Results

1.2.1. Attention reports

On average, participants (N = 10) had to answer 33 probe questions (range 26-36), 18,6 of these probes were triggered by the algorithm (peak probes, range 13-31). Participants answered that they were off-task on 37,3% of the probes. We excluded one participant from analyses, because the percentage of off-task answers was too low (3%), resulting in a total of nine participants. Participants were unaware of the algorithm and assumed that the probes appeared at a pseudo-random interval.

1.2.2. Online RTcv as a measure for mind wandering

In the next section, I will first describe some analyses we ran that checked whether the algorithm could fulfil its use properly.

First, we wanted to investigate whether the algorithm could present the probes on the right moments of the RTcv fluctuation. To present a probe, the algorithm had to keep track of the RT, calculate the RTcv and RTcv-threshold value, track the amount of trials since the last probe was presented and track whether a probe was presented recently or not.

Based on visual analyses of the performance data, all tracking variables (Ntrials , C),

(28)

compared this to the RTcv values generated by the algorithm. We found the RTcv values to be similar. To check whether the threshold sampling method worked, we calculated the RTcv over the eight trials before a probe and compared these values for the two probe types (max vs. peak). We hypothesized the RTcv value to be higher for the peak probes. A Bayesian t-test was used for statistical inference.

We found extreme evidence in favour of the alternative hypotheses (BF10= 7169510)

indicating that the RTcv was higher before a peak probe compared to a max probe. This analyses confirmed that our algorithm could online track RTcv and trigger probes after peaks in RTcv.

In Fig. 10, an example is given of the RTs, RTcv, and RTcvthreshold values. As shown,

the green dots represent probes that are presented because the algorithm reached a local minima in RTcv, the blue dots represent probes that are presented because the time between two probes was too long.

Figure 10

Example of the RT, RTcv, threshold and locations of probes of a participant

1.2.3. RTcv and probe answers

Next, we were interested in whether we could find the expected RTcv effect online. We hypothesized that the RTcv was increased for off-task probe answers, compared to on-task probe answers. We again calculated the RTcv before an off-on-task and on-on-task probe answer and compared the two conditions. A Bayesian t-test was used for statistical inference.

(29)

We found moderate evidence supporting the null hypotheses (BF01= 3.13), indicating

that there was no difference in the RTcv between off-task and on-task probe answers across all participants.

This result came somewhat unexpected, since it is not in accordance with the existent literature (e.g. Cheyne et al. 2009; Bastian & Sackur, 2013) and the analyses we did on off-line data. It might be that the result is correct, but we first explored some alternative explanations for this finding. We conducted some follow up analyses to investigate what factor could have possibly caused this effect.

1.2.4. Factors influencing probe answers

Based on our earlier analyses on the data obtained in our lab, we hypothesized that probe answers could be influenced by two other factors, namely the amount of commission errors made in a block, or by the length of a block.

We hypothesized that the amount of commission errors was increased before an off-task probe answer. Therefore, we calculated the number of commission errors made in the ten trials before a probe and compared this amount per participant for off-task and on-task probe answers. A Bayesian t-test was used for statistical inference.

We found strong evidence in favour of our alternative hypotheses (BF10= 72.3)

indicating that the amount of commission errors is increased before an off-task probe answer. Second, we hypothesised that the blocks followed by a max probe were longer than the blocks followed by a peak probe. Since mind wandering reports seem to increase as a function of task duration (Smallwood & Schooler, 2006), the max probe, which we implemented to detect the online attention states of the participant, therefore could now catch the off-line behaviour, instead of online behaviour.

We found moderate evidence for the hypotheses supporting the idea that the block length was increased before a max probe compared to a peak probe (BF10= 4.26). This might

be a possible explanation for why participants filled in they were ‘off-task’ instead of ‘on- task’ on the max probes.

2.2.5 Conclusion

Concluding from above, even though the basic tasks the algorithm had to fulfil were met (calculate RTcv, track trials, track probes, insert probes after a peak in RTcv and after 45

(30)

trials), the algorithm did not catch the online behaviour with the max probe types. Also, the probe answers were influenced by the block length and the amount of commission errors made in the trials before. Therefore, we decided to adjust the algorithm.

We hypothesised that as high variability in reaction times is linked to mind wandering, then low variability in RT should be related to task appropriate behaviour, since the participants responds consistently within the right temporal window on the digits on the screen, as is expected during the SART.

We decided to replace the max probes for trough probes, probes given after a trough in RTcv, to catch the ‘online’ behaviour of the participants. Another pilot study was conducted to investigate whether the algorithm could successfully be applied.

2.

Experiment 3: the final algorithm

2.1.

Method

A total of six participants (four males; mean age 25.8 years, S.D. = 2.83) participated in the experiment in exchange for money (10 euro). Participants filled in an informed consent prior to the experimental procedure. The study was approved by the ethics committee of the psychology department of the University of Amsterdam.

All participants completed the SART. The paradigm was similar as the one used before, except that this algorithm presented the probes both after a peak in RTcv and after a trough in RTcv.

Presentation of a trough peak was based on the same parameter settings as for a peak probe (current RTcv, Ntrials, C, RTcvthreshold), except that this time for the threshold the 80th

percentile was used; a probe was presented if the current RTcv was similar or lower than the 20% lowest RTcv-values. The peak threshold was put to the 20th percentile. We changed this setting in order to allow a sufficient number of probes to be presented per participant.

The last eight RTcv values of session 1 were used as input for the threshold in session two, such that the old threshold was the start threshold for session 2.

(31)

The probe question asked participants to rate their attentional state based on a 5-point Likert Scale, instead of a 2-point scale. This change was made because it allowed us to performcorrelation analyses. Furthermore, Likert scales allow people to have a certainty rating included in their attention reports, allowing for individual differences in confidence about their attention rating.

Participants filled in a questionnaire which checked whether participants were aware of the algorithm.

2.2.

Results

2.2.1. Attention Reports

Across the 2 sessions, participants chose the "off-task" probe answer (cat 1, cat 2) on 10.1% and 19.1% of the probe questions, and an "on-task" probe answer (cat 4, cat 5), 31,1% and 12.4% of the occasions. The middle category was chosen 27.3% of the time. Participants were unaware of the algorithm. Fig. 11 shows a histogram of probe answers for all the participants.

Figure 11

Histogram of probe answers

2.2.2. Algorithm

The first analyses we conducted, checked whether the factors we found in earlier analyses to influence the probe answers, did not influence the current data.

(32)

Earlier analyses in this study showed that there was an influence of block length on the answers on the probes, as longer blocks were more often answered with an off-task probe answer.

Since we adjusted the algorithm in such a way to overcome this problem, we hypothesised that the block length of the peak and through probes would be similar this time. We calculated the amount of trials per block and used a Bayesian t-test for statistical inference.

We found anecdotal evidence in favour of our null hypotheses indicating that there was clear proof for a difference in block length between peak and trough probes (BF01= 1.23).

Therefore, we concluded that our algorithm worked and that probe answers could no longer be biased by block length.

Earlier analyses furthermore showed, that blocks with more commission errors were more likely to be answered with an off-task probe answer. We wanted to check whether this effect was also present in the current data. However, the amount of commission errors, is biased by the amount of no-go-trials in a block. Therefore, we did a partial correlation in which we corrected the amount of commission errors for the amount of no-go-trials in a block and correlated the residuals with the probe answers.

We found a positive correlation (r = 0.21, BF10 = 42.65) between the amount of

no-go-trials and probe answer, indicating that if there are less no-go-no-go-trials in a block, participants are more likely to be off-task then on-task. Apparently, no-go-trials keep participants focussed on the task. This might be because when a no-go-trial appears on the screen, participants become aware of the task conditions again and actively try to prevent their minds from wandering.

2.2.3. RTcv effect

The main effect we were interested in was whether off-task probe answers were related to mind wandering. To assess whether the algorithm could successfully relate RTcv to mind wandering, we examined the relation between RTcv and probe answers. We expected a negative correlation between RTcv and probe answers, lower probe answers (indicating being off-task) should be related to a higher RTcv.

We calculated the correlations between RTcv and probe answer per participant and used a Bayesian correlation- t test to assess whether the direction of the correlation was different from zero.

(33)

An overall negative correlation (mean r = - 0.15, BF10 = 5.19) existed between RTcv

and probe answer, indicating that the RTcv was increased before an off-task probe answer. This confirmed our initial hypothesis and validated the approach of using experience sampling online.

However, since we trace the RTcv online, the algorithm could influence the behaviour of the participants. It is unclear whether participants are aware of their variability in their response times, and thus whether there is a causal role in their introspection. Perhaps peaks in RTcv have a subjective counterpart, and thus relate in a way to mind wandering.

To test whether participants could be aware of their variability in reaction times, we invited the six participants back to the lab, and used the data of the algorithm as input for the new experiment such that the probes appeared at the same time interval as in Experiment 3, but this time were independent of the algorithm (random).

Three participants could be tested again. We calculated the correlations between RTcv and probe answer for the random condition and the algorithm condition and checked whether the effect occurred in the same direction (negative).

We found the effect to be similar for both conditions, indicating that the algorithm worked similar to a random probe insertion. However, the power of this result is low, due to the low amount of participants we were able to test, but indicates that we can possibly use the algorithm in our follow up study.

2.3.

Conclusion

Concluding from above, the algorithm could successfully be applied. Increased RTcv is related to more off-task probe answers of the participants. However, one factor for which we cannot control with the use of the algorithm, is the amount of no-go-trials in a block. When this amount is increased, participants are more intended to answer they are on-task. However, since we did not find substantial evidence against the use of the algorithm, we will not discard its use, and will continue to use it in a follow up experiment.

Now we found a way for on-line detection of mind wandering, the next step is to study the role of certain brain areas in mind wandering. Knowledge of which brain areas are active during mind wandering can give more insight in the role of executive functions, attention and/or consciousness/awareness in mind wandering. Furthermore, understanding the neurobiological markers of mind wandering is a necessary step in understanding human thought (Christoff, 2012) .

(34)

The next section will function as a review on the literature and will summarize, although not directly, the hypotheses we have concerning the neural correlates of mind wandering. I will review this literature because in the study that will follow up the behavioural analyses done in this study, we will try to use simultaneous EEG-fMRI and eye tracking, to allow the full characterization of the evolution of the neural processes as they occur during mind wandering.

3.

Behavioural markers and the brain

Researchers interested in studying the neural correlates of mind wandering, often use a similar paradigm as described above (SART), and apply this paradigm in the scanner (fMRI). Both performance errors and attention reports can be linked to neural activations.

4.1 functional MRI

Christoff et al. (2009) used this paradigm in combination wih fMRI to investigate which networks are recruited during both performance errors and before reports of mind wandering. They were mainly interested in the role of executive functions and the default mode network, an area related to states of low cognitive demand.

In their study, they showed that before reports of mind wandering, participants made more SART errors compared to periods of being on-task. This is in accordance with Cheyne et al. (2009) and our study. Furthermore, Christoff (2012) demonstrated that three brain systems are related to mind wandering; the default mode network (DMN), the executive functioning system (EF) and the sensory cortices. I will describe these systems in the next sections.

The default mode network: The default mode network includes the medial PFC, posterior cingulate/precuneus region, and the temporoparietal junction (Christoff, 2012). Both reports of mind wandering and performance errors are preceded by activations in the default mode network. Furthermore, activations in the DMN have been related to self reports of mind wandering, behavioural errors and attention lapses (Mason et al., 2007b; McKiernan et al., 2006 in Christoff, 2012). Therefore, there is strong evidence that activations in the DMN are related to mind wandering.

(35)

The executive network: In addition to the DMN, activations were observed in the executive network (EF), consisting of the dorsolateral prefrontal cortex (dlPFC) and the anterior cingulate cortex (ACC), during episodes of mind wandering, compared to episodes of being on-task (Christoff, Carriere, & Smilek, 2009). This finding seems surprising because it was first thought that the DMN and the EF work in opposition, since EF activation is seen in tasks with high cognitive demands (Christoff, 2012). However, mind wandering uses cognitive resources, like working memory and executive control, and therefore reduces performance on the primary task.

Kam et al. (2014) showed that mind wandering mainly impairs two executive functions, working memory updating and response inhibition, influencing performance on the SART. The parallel recruitment of the DMN and the EF is also observed during naturalistic film viewing (Golland et al., 2007 in Christoff, 2012) and creative thinking (Kounious et al., 2006,2008, Subramaniam et al., 2009 in Christoff, 2012). Functional connection analyses showed that the co-recruitment of the DMN and the EF occurred in the presence of a significant positive functional connectivity between the networks (Christoff, 2012) .

The sensory cortices: A third system related to mind wandering are the sensory cortices. Christoff, 2012 found a negative correlation between the default network and some sensory regions, including the visual and somatosensory cortices. This is in accordance with the perceptual decoupling hypotheses of mind wandering, stating that during mind wandering, external information is not processed deeply and therefore task performance diminishes (Smallwood et al, 2011).

4.2. EEG studies

Analyses of the psycho-physiological precursors of lapses of sustained attention have further been done with EEG. O’Connell et al. (2009) used the continuous temporal expectancy task (CTET) in which participants had to monitor the duration of a frame and had to identify a target frame which has a longer duration compared to the non target frames. Event related activity was compared over three time intervals; immediate target related processing, short term epochs preceding targets (4 s), and long-term epochs preceding targets (30 s) and compared for correctly detected targets (hits) and undetected targets (misses).

O’Connell et al. (2009) found that 20 seconds before the occurrence of a miss, a slow drift in the alpha-band amplitude (8-14 Hz) occurred at posterior electrodes. Furthermore,

The Mind-Wandering Brain: Analyses of behavioural indices of attention lapses

Table of contents

Introduction

Behavioural markers of mind wandering

Experiment 1: Replication of Cheyne et al. (2009)

Method

Results

Commission errors

Anticipations

Summary of Experiment 1

Experiment 2: RTcv-tracking algorithm

Method

Results

Experiment 3: the final algorithm

Method

Results

Conclusion

Behavioural markers and the brain

4.1 functional MRI

4.2. EEG studies