Predicting mind wandering based on behavioral and p
sycho-physiological
measures
January, 2016
Miriam C.L. Maan, University of Amsterdam, Amsterdam
Under supervision of: Alan Kingstone, University of British Colombia, Vancouver
Abstract
A significant body of literature supports the idea that the lack of attention classified as mind
wandering, is related to variations in skin conductance (SCL), performance (errors and
reaction time), blink rate, saccade amplitude, fixation duration and fixation frequency. Here,
we assessed how these variables could predict an episode of mind wandering in an experiment
using simultaneous data recording over the different channels (skin conductance, Eyelink,
performance). A logistic regression model is used to assess which variables could predict
mind wandering best. The data reveal that the odds of attending are predicted by variations in
skin conductance level and fixation duration, with increases in these variables relating to
enhanced attention, and task duration and the amount of commission errors (CE), which both
predict mind wandering. The decreasing effect of task duration and CE on attention could be
mediated by correct inhibitions, demonstrating that correct inhibitions have an alerting
function. These results corroborate to the hypotheses that mind wandering episodes can be
predicted based on behavioral and psycho-physiological measures, and that reward has a
mediating effect on attention.
2
Contents
1. Theoretical Background ... 3
2. Materials and Methods ... 8
2.1. Subjects ... 8
2.2. Stimuli and Procedure ... 8
2.2.1. Sustained Attention to Response Task ... 8
2.2.2. SART measures ... 9
2.2.3. Galvanic skin response recording and preprocessing ... 10
2.2.4. Ocular data recording and preprocessing ... 10
2.2.5. Statistical Analysis ... 10
3. Results ... 12
PART 1: Relation between SART performance, skin conductance, blink rate and attention ... 13
3.1. Attention-Level and SART performance ... 13
3.2. Attention-Level and Skin Conductance ... 16
3.3. Attention-Level and ocular changes ... 17
PART 2: Logistic Regression Model building ... 20
3.4. Conclusion: correlation analysis ... 22
3.5. Model entry of parameters ... 23
3.6. Model fit ... 23
3.7. Model validity ... 25
3.8. Logistic Regression - Model interpretation ... 25
PART 3: The possible mediating effect of CI, CE and skin conductance on attention ... 27
3.9. Model including CI... 27
3.10. Model Interpretation ... 28
3.11. Model including CE ... 28
3.12. Model including SCL ... 29
PART 4: Component Analysis ... 29
3.13. Combining Factors: exploratory factory analysis ... 29
4. Discussion ... 31
3
Predicting mind wandering based on behavioral and p
sycho-physiological
measures
1. Theoretical Background
In the real world, task situations are not always attention demanding, and we often experience
our thoughts drifting off while not remaining on a single topic for a long period of time. This
tendency of the mind to drift away is labelled as mind wandering and is defined as a state of
decoupled attention whereby thought is focused inward, and is not related to the current task
or activity (Smallwood & Schooler, 2006; Christoff et al., 2009; Douchet et al., 2012). The
study of mind wandering has recently gained much interest in the scientific community.
Methods of study have been validated as the phenomenon is relevant to a wide array of
disciplines, including neuroscience, philosophy and the clinical world
(Callard, Smallwood, Golcher, & Margulies, 2013). Mind wandering negatively impacts performance as individuals
cannot simultaneously focus on their internal train of thoughts and the task at hand (Bixler &
D’Mello, 2014), but is also related to creative thinking and problem solving (Smallwood &
Schooler, 2006). Understanding which factors predict a mind wandering episode, can be
useful for individuals who would like to promote it (meditation), or who want to prevent
themselves from a mind wandering episode, as could be the case for individuals engaged in
public safety or transport.
Previous research has mainly focussed on the study of the content (Christoff et al., 2009),
duration and frequency of mind wandering (Bastian & Sackur, 2013) and has led to insights
about the relation of mind wandering to several behavioral and p
sycho-physiological
variables, including performance errors and reaction time (Cheyne et al., 2009), oculometric
variations (Grandschamp et al., 2014) and variations in arousal level (Smallwood, 2004).
First, mind wandering is more likely to occur when the situation at hand is repetitive,
unvarying, and undemanding (Smallwood and Schooler, 2006). This principle is used in the
tasks aimed at measuring attentional disengagement, like the sustained attention to response
task (SART).
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
4
withhold this response on a target trial (no go trial). The task is repetitive and requires
sustained attention in order to perform well. This becomes harder towards the end of the task.
Task duration therefore, is an important predictor of mind wandering, and studies often find
that mind wandering reports increase with task duration (Smallwood, 2004).
Second, errors in performance and variations in reaction time (RT) and mean RT in the SART
relate to mind wandering episodes (e.g. Cheyne et al. 2009). The Executive Functioning
Theory of Mind Wandering states that cognitive resources are limited, and as both task related
thoughts and mind wandering use the same neural sources, mind wandering affects task
performance (Smallwood & Schooler, 2006). Failures of response inhibition on target trials
(commission errors; CE); repetitive behavior (anticipations; AN), speeding of response times,
reaction time variability (RTcv, cv = SD/mean) and the neglect of a trial (omission; OM) are
several types of behavior that occur because mind wandering impairs working memory
updating and executive control (Kam & Handy, 2014; Cheyne, Carriere, & Smilek, 209).
These behavioral indices have been linked to different states of attentional disengagement
(Cheyne, Carriere & Smilek, 2009). Brief episodes of mind wandering are related to reaction
time variability, while generic task inattention is linked to anticipations. Lastly, response
disengagement, where an individual is completely decoupled from conscious processing, is
related to omissions (Cheyne, Carriere & Smilek, 2009). The three states of attentional
disengagement are sequentially related; anticipations and omissions often occur in succession,
and behavior on the SART therefore, can give important insights on the depth, frequency and
duration of a mind wandering episode
Third, mind wandering has been related to changes in arousal level. Smallwood et al. (2004)
examined how differences in attention are related to task performance on the SART and
p
sycho-physiological
measures (heart rate and galvanic skin response). Increases in skin
conductance levels (SCL) are seen as a function of time on task for blocks in which
participants attend, but there is no difference in their SCL during mind wandering blocks
(Smallwood et al., 2004). Furthermore, researchers link activity in areas that are active during
mind wandering episodes to decreases in SCL (Nagai et al., 2004; Christoff et al.,
2012).During a biofeedback task, activity in the ventral medial prefrontal cortex (vmPFC) and
orbitofrontal cortex (OFC), extending to the right medial temporal lobe (TL) and right
temporal pole (TP), was associated with decreases in skin conductance levels (SCL) during
5
trials in which participants received relaxation feedback versus trials in which they received
an alerting feedback prompt.
Fourth, spontaneous blink rate could be seen as an important
psycho-physiological
indicator
of mind wandering. The frequency of blinks is increased in blocks where mind wandering is
reported, compared to blocks in which participants report to attend (Bristow, Frith, & Rees,
2005; Bristow, Haynes, Sylvester, Frith, & Rees, 2005; Ridder & Tomlinson, 1997;
Volkmann, 1986). This link between blink rate and mind wandering is further confirmed by
other studies: blink rate in the 5-s periods preceding a probe on which participants reported
mind wandering was increased, in contrast to probes in which participants answered they
were on-task (Smilek, Carriere & Cheyne, 2010). Grandchamp et al., (2014) also observed an
increase in the number of blinks per minute during episodes of mind wandering, when using
examinations of attention based on self report during a breath counting task.
Gaze stability is another
psycho-physiological
indicator of mind wandering. The amplitude of
eye saccades is increased during mind wandering episodes (Grandschamp, Braboszcz and
Delorme, 2014), and participants make more fixations (fixation count) during episodes of
probe caught mind wandering (Uzzaman and Joordens, 2011). Lastly, Reichle et al. (2010)
report longer fixation periods (fixation durations) during mindful, versus mindless reading,
indicating that gaze is less stable during mind wandering compared to attended episodes.
Before mentioned studies provide a list of possible variables that could be used to predict the
likelihood of attending/mind wandering during a sustained attention task. 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). Also, measures of blink rate, saccade amplitude, fixation count
and duration, or arousal levels can be used as indicators for mind wandering episodes. Until
now, these variables have mostly been studied in isolation, even though there are some studies
assessing multiple variables simultaneously (Smallwood, 2004; Grandchamp et al., 2014).
This study will use measures of performance, eye position and arousal to simultaneously
study the relation among these variables and mind wandering, in order to build a model that
predicts an episode of mind wandering. We hypothesise that all behavioral variables (CE, CI,
OM, AN, RTcv, mean RT), ocular variables (blinks, saccade amplitude, fixation duration,
fixation count) and skin conductance are predictive of a mind wandering episode. Task
duration is hypothesised to explain the most variance in attention, followed by performance
6
errors and variations in arousal level. This is hypothesized because links between behavior
and mind wandering have been established by a broad area of research, while other relations
(for example the relation between SCL and mind wandering or gaze stability and mind
wandering), is less explored.
Subsequently, understanding the causal relationship between mind wandering and
psycho-physiological variations could be used in the prevention or stimulation of mind wandering
episodes. For example, previous studies have found that commission errors could function as
an alerting signal, inducing reflective mind wandering, which ultimately prompts participants
to re-focus on the task (Cheyne, Carriere & Smilek, 2009). Commission errors are in fact
associated with post error slowing (PES), during which participants reflect on task behavior,
and are associated with greater SCLs (Zhang et al., 2012), providing further evidence for the
alertment function of errors.
Jonker et al, (2013) found that not only errors, but also correct inhibitions on target trials
could lead to these types of reflection processes. Correct inhibitions are followed by errors on
the task, indicating an episode of reflective mind wandering, after which a participant
re-engages to the task. Correct behavior has a rewarding function, and previous studies have
shown that reward can mediate attention (Peck et al., 2009), providing a possible explanation
for the positive effect of CI on task behavior.
Furthermore, previous studies have shown that by training individuals to internally increase
their arousal levels when cued, participants are able to make fewer errors on a sustained
attention task (O’Connell et al., 2008). The placebo condition was linked to a gradual
reduction of SCL with time and increased reaction time variability, and these effects were not
shown in the self alertment group.
These studies provide possible windows for intervention. Therefore, after finding evidence for
a causal relationship between performance, ocular variations, arousal and mind wandering, we
aim to assess how CE and CI could mediate the relation between mind wandering and ocular
variations in blink rate, saccade amplitude, fixation count, fixation duration and skin
conductance, and how SCL could mediate the relation between these ocular variations,
performance and mind wandering.
The layout of the paper will be as followed. First, we will replicate previous literature relating
performance on the SART, blink rate, gaze stability and skin conductance to mind wandering
7
episodes. We predicted that all behavioural markers (anticipations, omissions, commission
errors, mean RT and RTcv) are indicative of an episode of mind wandering, and that mind
wandering blocks are indicated by an increased blink rate, decreased gaze stability and a
refusal changes in skin conductance levels during off-task (mind wandering) blocks.
Subsequently, we will investigate how these variables can be used in a logistic regression
model predicting attention during a block as measured during the SART. As our
quantification of attention we will use the subjective classification of the level of attention of
a participant during a block, by predicting the answer on a probe question aimed to quantify
the level of attention either as on task or off task, based on performance, ocular variations and
arousal. For that reason, we will first assess which variables should be included in a model,
that both significantly predicts attention and improves model fit, using (partial) correlation
analysis.
Based on the findings of the second analysis, we aim to study how CE, CI and SCL mediate
the predictability of the variables in the logistic regression model, by including interaction
terms. Because of the rewarding function of CI, we predict that CI’s will mediate the relations
in our model the strongest. Lastly, we will conduct some exploratory analyses to study if we
can find components in our data on which several variables load, to study whether they
explain the same kind of variance, and could be grouped together. These analyses are
conducted as our simultaneous study allows for the examination of common components in
our data. We hypothesize that task duration and performance errors load high on one
component and skin conductance and CI and CE on another component, as they explain a
similar kind of variance in the data.
This study can contribute to knowledge about the co-variation of physo-psychological
measures and subjective reports and could increase construct validity of the concept of mind
wandering. Also, it could lead to a possible intervention mechanism if CE, CI or SCL can
mediate the effects of task duration, ocular changes or arousal and their predictability of mind
wandering. This is one of the first studies assessing multiple measures of mind wandering
simultaneously (Grandshamp et al., 2014).
8
2. Materials and Methods
2.1. Subjects
A total of 47 participants participated in the study. Due to problems with the data collection, two participants had to be excluded, such that the total amount of participants was 45 (17 male, Mage=21.2, SDage=2.56). All participants had normal or corrected to normal vision. Written informed consent was obtained prior to the study. The experimental protocol was approved by the ethical committee of the University of British Colombia.
2.2. Stimuli and Procedure
2.2.1. Sustained Attention to Response Task
In the present study, participants completed a sustained-attention to response task (SART) during which they intermittently provided assessment of task engagements. The SART was an adapted version from the one used by Robertson et al. (1997).
On every SART trial, a digit (1-9) was shown for 350ms, followed by a “-“ mask for 900ms (see figure 1). Each digit was presented equally often, such that 11% of the digits were target-trials (“3”) and 89% were non-target-trials. Each digit was presented in black against a grey background. The font style of the digits was comparable to a digital clock. This was done to minimize the change in luminance, to control for possible changes in the pupil size. The task was divided in 40 blocks with an average of 25 trials per block. Every block started with a 1500ms “-“ mask. The size of the digits varied randomly across blocks, with the fonts being sampled from a vector of three possible font sizes (10, 12.5 and 15). This was done to prevent participants from simply responding based on familiar features of a given stimulus. The display was viewed from approximately 50 cm. Participants were restricted from head movement by a chinrest used to collect the ocular data.
Participants were instructed to respond as quickly and accurately as possible to the non target digits (go) by pressing the right shift-key, and to withhold a response on the target-trial (no go). They were instructed to place most emphasis on being accurate versus being quick (see Seli et al., 2012). After twenty practice trials, participants completed on average thousand experimental trials.
At the end of each block, an assessment of task engagement was taken. Upon presentation of each probe, the participants were asked to indicate with a key press whether they were (1) on-task, or (2) off-task. The same description of task engagement and disengagement was given as in Seli et al.
9
(2015), instructing participants that: “Being “on-task” is defined as thinking only about task-related things (e.g., one’s performance on the task, the relative time of the button press, etc.). “Off-task” was defined as thinking about internally focused, task- unrelated things (e.g., plans with friends, an upcoming exam, etc.). After each attention-probe, participants were asked to rate their confidence in the attention rating they just reported. They were presented with a 4-item confidence scale, indicating whether they were (1) not at all, (2) somewhat, (3) very and (4) extremely confident in their rating. This confidence scale was included as Seli et al. (2015) theorize that lower level of confidence would less accurately reflect mental experiences, and they found that for higher confidence ratings there is a stronger relation between mind wandering and task performance.
After 20 blocks, there was a short break, before participants could continue with the second part of the experiment. The total duration of the experiment was on average thirty minutes, including a practise session.
Figure 1: Experimental task
2.2.2. SART measures
For all the correct go-responses, the mean RT of every trial was calculated. If participants did not make a response on a go-trial (non-target), this was coded as an omission (OM). When participants responded within 100ms after the start of a non-target-trial, this was coded as an anticipation (AN). Responses made on target-trials, were coded as commission errors (CE). When participants correctly withhold their response on a target trial, this was coded as a correct inhibition (CI).
10
2.2.3. Galvanic skin response recording and preprocessing
Reusable skin conductance electrodes (8 mm, Bio-trace) were attached to the palm side of the index and middle fingers of the non-dominant hand. The signal from the electrodes was amplified with a DC-EXG- wideband amplifier and sampled on a rate of 256Hz. SCR data was filtered with a 1st order Butterwort filter. A bidirectional filter with low pass cut off frequency of 5/128 Hz was used and a bidirectional high pass filter with cut off frequency of 0.05/128 Hz. Data were meaned over stimulus-trials and fixation-stimulus-trials, such that every trial had two values, one mean for the fixation duration and one mean SCL value for the stimulus duration. The data was standardized using a z-transformation. The mean and standard deviation of all the SCL values are calculated for all the trials of a participant and a standardized SCL-value was obtained for every trial (z = x-mu/sd).
2.2.4. Ocular data recording and preprocessing
Eye movements, blink rate and pupil size were monitored with an Eyelink 1000 plus, eye tracker setup for the participants right eye. The eye tracker samples with a frequency of 1000Hz. Before each experiment, a 9-point calibration was performed. During the calibration, participants had to follow a white dot which moved randomly across each of the 9 points on the screen.
Message events were created to mark the start and end of a trial, and important events including display changes and button events. Data-viewer was used to obtain a trial report of the data on a trial-by-trial basis including duration, start time, average fixation duration, fixation count, pupil size (min, max, mean), saccade count, blink count, number of button presses and event messages. These files were used for analysis.
2.2.5. Statistical Analysis
Statistical analyses were performed using Rx64 3.2.0. The performance indicators (CE/CI/AN/OM) and mean reaction time were calculated for all blocks over the total duration of the experiment. Reaction time variability (RTcv) was measured over a window of five trials. Anticipations and reaction times on no-go trials were not included in our calculations for mean RT and RTcv.
General Linear Mixed Models using a hierarchical logistic regression model, were used to test for the predictability of attention answer, based on the behavioral variables, blink rate, saccade amplitude, fixation duration, fixation count and skin conductance levels. Using a hierarchical model allowed us to have a better estimation of the effect, as the error term is estimated at the subject level, improving generalizability of our model to other samples and the population. Independent variables were added
11
to our model based on the strength of the (partial) correlation between the variable and our dependent variable (attention-level) and improvement in model fit. Maximum likelihood was chosen as the method of estimation. Only main effects were included in the model. Next, mediation of CE, CI and SCL to the variables with attention will be assessed by including interaction terms in our final model. The three models will be assessed in terms of likelihood and model fit. Lastly, principal component analysis, including orthogonal rotations, will be conducted on all variables to assess possibilities for variable reduction in our dataset.
12
3. Results
In the next sections, we will first examine the relation between task performance and mind wandering, followed by an examination of the relation between skin conductance and mind wandering, and lastly the assessment of the relation between the eye variables (blink rate, saccade length, fixation duration and fixation frequency) and mind wandering. This first part is used to examine differences in on-task and off-task blocks through univariate analysis such as t-tests and analysis of variance in repeated measures. Another important aspect is to identify outliers (described in more detail later) that might become apparent when the data values are separated into groups. These analyses will be conducted to replicate previous literature and will be informative about the reliability of the relation of mind wandering and these behavioral and psycho-physiological.
In the second part of the analysis, we will build a hierarchical logistic regression model to predict the direction of attention in a block. We choose to use the probe answer as our main dependent variable, as, even though that measure is subjective, it is the best indicator of an individuals’ attention, as we have today.
We hypothesize that the likelihood that a participants is mind wandering/attending to a block is related to the behavior on the task (CE/ AN/ OM/ CI/ RTcv/ mean-RT), their skin conductance and ocular changes (blink rate, FD, FC and SAC) and can be predicted from these variables.
These two parts will inform us on 1) how mind wandering can be detected in error rate and reaction time, ocular changes and skin conductance levels; 2) how rumination processes can be predicted based on the same variables.
Third, we were interested in how CE, CI and skin conductance mediate the predictability of the other variables to study whether the harmful effects of for example task duration or increased fixation duration could be mediated by behavior on the target (CE/CI) or differences in skin conductance. These analyses are conducted as previous research has found that CE and CI have an alerting function, as they relocate attention back to the task (Jonker et al., 2014;
Zhang et al., 2012
) and increased skin conductance levels have been found to be related to increased sustained attention and decreased error rate (O’Connell, 2009).Lastly, we conducted some exploratory analysis and used principal component analysis to study if our behavioral measures, skin conductance and our ocular variations would explain similar variance in our data.
13
In these sections, we will use the term on-task to refer to blocks in which participants answered they were attending, and the term off-task to refer to blocks in which participants answered they were mind wandering.
PART 1: Relation between SART performance, skin conductance, blink rate and attention
3.1. Attention-Level and SART performance
The purpose of this analysis is to demonstrate the relation between attention rating (on vs. off) and the two outcome measures of the SART; RT and error rate, to investigate whether we could confirm previous literature (e.g. Cheyne et al., 2009). A paired design was used to analyse the conditions. We predict that over all individuals, the mean difference between on- and off-task blocks for the performance errors is not zero. We compared the mean amount of commission errors (CE), correct inhibitions (CI), anticipations (AN), and omissions (OM) across the on-task and off-task blocks for all participants except the first five, as we could not analyse their RT trial data due to a coding error in the Matlab-script.
All the paired differences have a normal distribution, except for the amount of omissions. Paired differences are calculated by subtracting the mean of the variable (e.g. CE/CI/AN/OM/mean-RT/RTcv) in the on-task condition from the mean amount of this variable of the off-task condition for every participant.
A paired two sampled t-test yielded a significant difference between these two conditions for commission errors, t(39) = 5.3259, p-value <0.01, 95% confidence interval for the paired difference: 0.2742999- 0.6102287 d’ = 1.705365), demonstrating that the amount of CE was higher in off-task blocks (M= 1.409, sd=0.7266), compared to the on-task blocks (M=0.9668, sd=0.4767). Furthermore, there was a significant difference in the amount of anticipations for on (M=0.2260, sd=0.416) and off-task blocks (M= 1.327, sd=0.6493; t(39) = - 7.5733, p-value <0.01, 95% confidence interval of the difference: 0.8065853 - 1.3944361, d’=-2.4253), with the amount of anticipations being higher in off-task blocks. The amount of correct inhibitions was higher in the on-task condition (M=1.666, sd=0.5077) compared to the off-task condition (M= 1.3265, sd=0.6493; t (39) = -5.53366, p-value <0.01, 95% confidence interval: -0.4637287 - - 0.2155635; d’=1.77219). These findings confirm previous literature that relates these behavioral makers mind wandering episodes and attention (Cheyne et al., 2009).
14
However, omissions were not specifically related to off-task blocks. As the mean amount of omissions was not normally distributed, and the skewness was too high to use the Wilcoxon Signed rank test (skew =2.2057, Z=4.5722, p<0.05), we chose to remove the outliers, which fell outside the interquartile range and removed them based on a practical and substantive standpoint: the rate of omissions was unexpectedly high for these participants, this might have been because of a recording error, it seems like they pressed the wrong button during one or more blocks, resulting in a high amount of omissions. After removing the outliers, the data was normally distributed, so we could conduct a paired-sample t-test. This test indicated that there was no difference between the on-task (M=2.047507, sd=1.363873) and off-task condition (M=2.02805, sd=1.425191) in the amount of omissions (t(39) = 1.2404, p=0.2222, 95%-confidence interval of the difference =-0.0965838, 0.40286060; d’= 0.4024389).
After outlier removing (for which we followed the same procedure as for the omission data), the RTcv difference histogram was normally distributed. There was a difference between the on-task (M=0.3916, sd=0.1196) and off-task condition (M=0.4296, sd=0.1736), the RTcv was higher for the off-task condition (t(38) = -2.6283, p<0.05, 9% = -0.0673508/-0.008742433, d’=-0.8527331). This was in accordance with previous literature (Bastian & Sackur, 2013; Cheyne et al. (2009)).
There was no difference in the mean RT between off-task (M=0.4110725, sd=0.07911) and on-task blocks (M=0.41264, sd=0.0759) (t(39) = 0.3113, p=0.7574, 95%CI: -0.00867059/0.01182295, d’ = 0.09970).
Furthermore, there was no significant difference in the amount of target trials for off-task (M=2.739, sd=0.296), compared to on-task blocks (M=2.6329, sd=0.296) t(39) = 1.3795, p=0.1756, -0.04784462 0.25308109; d’=0.441633).
Table 2:Mean, standard deviation, t-value and effect size for the difference in SART behavior for the
on-task and off-task blocks
SART behavior Mon SDon Moff SDoff t d’
Commission Errors 0.9668 0.4767 1.409 0.7266 5.325 1.705365 Anticipations 0.226 0.4159 1.3265 0.6493 - 7.573 -2.4253 Omissions 2.048 1.3639 2.028 1.4251 1.2404 0.40244 Correct Inhibitions 1.666 0.5077 1.3265 0.6494 -5.533 1.77219 RTcv 0.3916 0.1196 0.4296 0.1736 -2.6283 -0.8527331 Mean RT 0.4126 0.0760 0.4111 0.0791 0.3113 0.09970 Target Trials 2.6329 0.2412 2.7355 0.2962 1.379 0.441633
15
Figure 2: Summary of SART performance errors
Mean, standard deviations of commission errors, anticipations, correct inhibitions, and RTcv for on and off-task blocks.
-0.5 0 0.5 1 1.5 2 2.5 Commission Errors Anticipations Correct Inhibitions RTcv on-task off-task
16
3.2. Attention-Level and Skin Conductance
Next, we were interested in how skin conduction varied as a function of Attention-Level. Previous studies found that skin conductance increased with task duration in blocks were participants reported they were on-task, but there was no difference in the SCL in blocks were mind wandering was reported (Smallwood et al., 2004).
We chose to not use a paired design as skin conductance levels drift with task duration, and instead used a repeated measures ANOVA to analyse the data. We hypothesised that the SCL increases with task duration for on-task blocks, but not for off-task blocks. We subtracted the mean over the last five trials in a block from the baseline measure taken directly before the start of a block, and we expected this difference to be positive for on-task blocks.
We first made a plot of the data to visualize the effect of probe answer on the difference vector (diff = last 5 – BL).
Figure 3: In the graph, a blue line
represents the linear approximation of the difference of BL and last 5 trials for on-task (lower) and off-task blocks (upper) for every participant during the total duration of the experiment. The grey area represents the standard error of this estimate. Negative values indicate a decrease in SCL during a block, positive values indicate an increase in SCL during a block (as diff= last – BL), the y=0 line is added to the plot to improve visual understanding of the data.
As can be seen in the plot (figure 3), off-task blocks are in approximately the first fifteen blocks represented by positive difference values; indicating the SC increases in off-task blocks. This could possibly be explained by the co-occurrence of errors and mind wandering, errors induce reflective mind wandering and are associated with post error slowing and increases in SCL
(Zhang et al.,
17
2012),
. This pattern is reversed in the later blocks; the difference in those blocks is most often negative: indicating a decrease in SCL during these blocks. This could be because participants engage in deeper depths of inattention, and errors go unnoticed. However, not all participants seem to follow this pattern; some participants have negative difference points in the first half and positive difference points in the second half of the experiment.During on-task blocks, the opposite seems to happen; during the first half of the experiment, the difference between baseline and last five trials is often negative: indicating a decrease in SCL during those blocks. In later blocks, however, the difference is mostly positive, indicating an increase in SCL during those blocks. Overall, these patterns seem to be in coherence with our hypotheses that the SCL increases for on-task blocks with task duration, but that this effect is not present in the off-task blocks.
To analyse the data, we used repeated measures ANOVA. Block counter was entered as a covariate variable in the analysis as the mean value of the skin conductance in a block depended on duration of the experiment, and we were not interested in this effect. We excluded the first block and the first block after the break for all participants, as these blocks difference scores were for many participants greater than two SD’s of the mean, and analysed a total of 38 blocks for every participant.
A repeated measures ANOVA indicated a interaction effect of block counter and attention-level (F(1,37) = 2.141, p=9.01e-05), indicating that the difference score was mediated by attention-level. The mean increase during a block for off-task blocks (M=0.0082706) was lower, compared to on-task blocks (M= 0.2617475). We also found a main effect of block (F(37)=3.233 p=3.58e-10). , indicating a bigger difference in skin conductance levels for later blocks.
3.3. Attention-Level and ocular changes
Subsequently, we were interested in the relation between our ocular measures and Attention-Level, to assess the predictions made in previous literature about the relation between these variables and mind wandering. We used a paired design to compare the blink rate, fixation duration, fixation count and saccade length for the on and off-task blocks for all the individuals.
A histogram indicated there was an outlier in the blink condition, we removed this participant from the analysis, after outlier removal, the data became normally distributed so we could conduct a paired t-test on the data.
A paired t-test indicated that the number of blinks was higher for off-task blocks (M=13.41245, sd=8.1422) compared to on-task blocks (M=12.31082, sd=7.635047) (t(43) = 3.7402, p<0.05, 95% =
18
0.5076337/1.d’= -1.68775). People intend to blink more in blocks were they report being off-task. This finding is in accordance with previous literature (e.g Grandshamp et al., 2014; Smilek et al., 2010) and our hypotheses.
The fixation count data was not normally distributed. There was no skew present in the data so we could conduct a Wilcoxon Signed Rank test (skew=0.53941, z=1.59250, p=0.1113). There was a significant difference between the number of fixations made for on-task (Mdn = 53.55, IQR= 13.67639 and off-task blocks Mdn=54.5, IQR = 17.1)(V=753, Mdn=1.888, p<0.05, 95%-confidence interval: 0.5296366- 3.2291667), the number of fixations was higher in the off-task condition, compared to the on-task blocks.
The fixation duration (ms) was higher in the on-task (M=15979.264, sd=5059.152) compared to the off-task blocks (M=1498.45, sd=5134.5) (t(44)=-5.1511, p<0.05). This is in accordance with our hypotheses and previous literature: Reichle et al. (2010) predict fixation durations to be longer for mindful versus mindless reading. The saccades-data was skewed to the right (skew=1.9083, z=4.3449, p<0.05), therefore we chose to transform the data using a log transformation. After transformation, the data became normally distributed, so we could conduct a paired t-test. This t-test indicated a difference between the log on-task sac vector and log off-task sac vector (t(44) = 7.3339, p<0.05, 95% = 0.3347701/0.5884805, d’=-1.668461, M=0.461623, sd=0.422224). The log saccade length was higher in the off-task condition, compared to the on-task condition.
To conclude, measurements of fixation duration, fixation count and saccade amplitude together indicate that the eye gaze of participants is less stable in the off-task condition, as people make more fixations, have shorter fixation durations and make longer saccades.
Table 3: Mean, standard deviation, t-value and effect size for the ocular variables for the on-task and
off-task blocks
SART behavior Mon SDon Moff SDoff t/V d’
Blink Rate 12.311 7.635 13.412 8.142 3.740 -1.6877
Fixation count 53.55 17.1 54.5 3.673 753 -
Fixation Duration (ms) 15979.26 5059.152 1498.45 5134.5 -5.1511 -1.5531
19 -5000 0 5000 10000 15000 20000 25000 Fixation Duration on-task off-task 0 20 40 60 80 100 Fixation count on-task off-task 0 10 20 30 40 50 60 Saccade length on-task off-task
Figure 4: Bar graphs indicating blink rate, fixation count, fixation duration and sac per participant.
Blue bars represent the on-task blocks, red bars represent the off-task blocks.
3.4. 0 5 10 15 20 25 Blink Rate on-task off-task
20
PART 2: Logistic Regression Model building
Besides confirming previous literature, the aim of this paper is to build a logical regression model to investigate whether the behavioral variables, skin conductance levels and ocular changes can predict attention in a block.
Our hypotheses are that the likelihood that a participant is on-task or mind wandering, is related to the behavior on the task (errors/AN/OM/CI/ mean-RT/ RTcv), skin conductance and ocular changes (blink rate, FD, FC and SAC). Before building a regression model, the dependency between the multiple predictor variables is examined using a correlation matrix.
We chose to center the block counter variable at block #5, to make the interpretation of our logistic regression model more meaningful, as we did not have a block #0, and we expected that at this block, all participants would be familiar enough with the task to give a reliable answer at the attention question, and we would expect their skin conductance levels to be task specific (and not be related to pre-task behavior). Lastly, we chose to use the untransformed version of the saccade amplitude variable, as a logistic regression model does not require multivariate normality. We used the data in which outliers were removed as predictor estimates in logistic modelling are sensitive to outliers. In table X, the results of the correlation tests are presented. Each cell on the right of the diagonal contains the correlation coefficient between each variable and the others. Pearson correlation was used, and to correct for multiple comparisons, pair wise comparisons were made on p-corrected values using Holms. The distribution of each variable is shown on the diagonal. On the bottom of the diagonal are the bivariate scatter plots with a fitted polynomial line. In the cells on top of the diagonal are the correlation values, including the p-values of the t-tests conducted on the correlation coefficients.
21
Figure 4. Bivariate profiling of relationships between variables: presents the scatter plots matrix of
selected variables (block counter, block errors, number of CE, number of CI, number of targets, number of anticipations, number of omissions, RTcv, SCL, blink rate, fixation duration, saccade amplitude and fixation count
22
Block counter positively correlated with block errors (AN: (r=0.17, p<0.05) OM: (r=0.071, p<0.05; CE: r=0.035 p=ns), and negatively with the amount of CI in a block (r=-0.25, p<0.05), This indicates that participants make more errors towards the end of the task. Block counter correlated negatively with SCL (r=-0.17, p<0.05), indicating that skin conductance decreases with task duration. Block counter correlated positively with the amount of blinks (r=0.16, p<0.05), saccade length (r=0.15, p<0.05) and fixation count (r=0.099, p<0.05), but negatively with fixation duration (r=-0.27, p<0.05) indicating gaze becomes less stable towards the end of the task.
Skin conductance negatively correlated with the number of anticipations (r=-0.056, p<0.05), this is probably because, when anticipations occur in succession, they represent a state of offline awareness and decreased alertness/arousal. SCL was positively related to the number of targets (r=0.10, p<0.05) and the number of CI (r=0.095, p<0.05). This could represent the alerting function of targets, which could help to refocus and direct attention back to the task (Jonker et al., 2012). SCL was not related to CE (r=0.15, p=0.554). This is probably because targets go unnoticed, explaining the error (CE) on this trial, and the lack of an increase in skin conductance. SCL negatively related to saccade length (r=-0.054, p<0.05) and fixation count (r=-0.073, p<0.05), but was not related to the other eye variables; blink rate (r=-0.028, p=0.256) and fixation duration (r=0.038, p=0.13).
All eye variables (blink rate, fixation count and saccade amplitude), except fixation duration (r=-0.27, p<0.05), were positively related to block counter (respectively: r=0.16, p<0.05; r=0.099, p<0.05; r=0.15, p<0.05), indicating that eye gaze became less stable with task duration. Blink rate correlated with the number of omissions (r=0.084, p<0.05), indicating that closure of the eye and the miss of a trial go together. Fixation duration with the number of CI (r=-0.074, p<0.05), number of targets (r=0.065, p<0.05), number of OM (r=-0.057, p<0.05), RTcv (r=-0.091, p<0.05). Saccade amplitude correlated with all behavioral measures; number of CE (r=0.068, p<0.05), number of CI (r=-0.055, p<0.05), number of anticipations (r=0.084, p<0.05), number of omissions (r=0.098, p<0.05), and RTcv (r=0.13, p<0.05), while fixation count only correlated with the number of CI (r=-0.053, p<0.05) and the number of CE (r=0.06, p<0.05). SCL was negatively related with saccade amplitude (r=-0.054, p<0.05), and fixation count (r=-0.073, p<0.05). These analyses indicate that increased fixation count, saccade amplitude and blink rate relate more to mind wandering and increased fixation duration more to attended episodes.
3.4. Conclusion: correlation analysis
From our correlation analysis, we conclude that some but not all variables are related. We have low multi-collinearity among our predictor variables; non of the predictor variables are highly correlated to each other(r>0.30). Therefore, specifying the correct logistic regression model can be done based on
23
the before mentioned variables, without affecting the reliability and stability of the regression coefficients, which could happen if the variables highly correlate.
3.5. Model entry of parameters
Next, the order in which the predictors would enter the model should be determined, and we assessed partial correlations between the variables and attention level. A partial correlation is the correlation that is unexplained when the effects of other variables are taken into account. Our previous analysis indicated block counter had the highest correlation with attention, and therefore we calculated the correlation of other variables with attention, when taking block counter into account. Amount of CE in a block had the highest partial correlation with attention (r=-0.176, t(1556) = -7.0733, p<0.01). We continued with this procedure, but now take into account both the effect of block counter and CE Now, mean RT had the highest partial correlation with attention (r=-0.107, t(1556)=-7.07, p<0.01). We continued this procedure and found significant partial correlations for fixation duration and attention (r=0.09, t(1556) = 3.575, p<0.01) and SCL and attention (r=0.05, t(1556) = 2.13, p<0.01). After that, none of the variables had a significant partial correlation with the dependent variable.
Table 4: variables, t-statistic partial correlation coefficients and significance values on df = 1556. The
partial correlations are calculated for the variables with the variables in the above columns partialled out.
Variable Estimate Statistic Significance Block counter -0.188 -7.0844 2.11e-12
Nr CE -0.17666 -7.0733 1.51e-12 mean-RT -0.10787 -4.274577 1.91e-05 FD 0.09497 3.7571 0.0001718 SCL 0.0540 2.131611 0.03308 AN -0.04008 -1.5790 0.11433 3.6. Model fit
Our next step involved creating a logistic regression model to predict attention based on the independent variables. The data is a series of repeated measures in which we try to model the variable Attentionlevel, which is the binary probe answer given by the participants, based on the factors block counter, commission errors, mean RT, fixation duration and SCL. Our first step involved the selection of predictors, based on (partial) correlations with the dependent variable and the factors (see previous paragraph). Next, logistic regression models were built, in which we sequentially enter variables based
24
on their incremental predictive power over variables already in the model. A chisquare-test compares the reduction in the residual sum of squares, to investigate whether the new model significantly improves deviance reduction. When a predictor could not improve model fit, it was left out of the model and a new variable was entered and assessed for model improvement. The results can be found in table 5. We included all factors as fixed and the intercept as random (1|subject). We used the lme4-package in R to conduct the analysis.
Table 5: Possible logistic regression models, AIC, BIC, loglikelihood and Chisquare tests. The
chisquare test represents the test of the N’th model with the N-1 Model. After Model 5, non of the predictors could significantly improve model fit. Models 6 to 11 are compared for model fit to model 5. Cells in grey represent the minimum values of AIC, BIC and logLik.
Logistic Regression Model AIC BIC logLik Chisq.
1) AT = B1 * block_ctr + (1|subject) 1996.9 2013.0 -995.48
2) AT = B1 * block_ctr + B2 * CE + (1|subject) 1927.8 1949.2 -959.89 71.18 3) AT = B1 * block_ctr + B2 * CE + B3*mean-RT + (1|subject) 1927.3 1954.0 -958.63 2.52.ns 4) AT = B1 * block_ctr + B2 * CE + B3* FD + (1|subject) 1920.6 1947.4 -955.31 9.1697
5) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + (1|subject) 1914.1 1946.2 -951.07 8.47
6) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*AN+ (1|subject) 1915.1 1952.2 -950.54 1.062 7) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*CI+ (1|subject) 1914.8 1952.3 -950.42 1.306 8) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*OM+(1|subject) 1915.9 1953.4 -950.96 0.2136 9) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*FC+(1|subject) 1916.0 1953.4 -950.99 0.1584 10) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*SAC+ (1|subject) 1915.3 1952.7 -950.63 0.8744 11) AT = B1 * block_ctr + B2 * CE + B3* FD + B4*SCL + B5*blinks+(1|subject) 1915.5 1952.9 -950.74 0.6562
We first build a model based on block counter as a predictor, as block counter had the highest correlation with attention (Model 1). In our second model, we included the amount of commission errors as a predictor in our model. Analysis of deviance indicated that the second component could improve model fit (chi(1)=71.18, p<0.01). A third model was build adding mean RT to the predictors. Entry of mean RT however, did not improve model fit (chi(1) = 2.52, p=ns). Mean RT was removed from the model and fixation duration entered the Model. This improved model fit, compared to Model 2 (chi(1) = 71.18, p<0.01). SCL could also significantly improve model fit (chi(1) = 8.47, p<0.01). All other variables were assessed (AN, CI, OM, FC, SAC, blinks), but they could not improve model fit. Therefore, we concluded the model including block counter, CE, fixation duration, and SCL provided the best fit on our data (AIC=1953.8 BIC=1991.2, logLik =-969.85). This model had both the lowest AIC and BIC score, compared to the other models, indicating that they agreed on the preferred model.
25
3.7. Model validity
For logistic regression, there is no concept that corresponds to the predictive efficiency or can be tested in an inferential framework (like Rsquared; Peng, Lee, Ingersoll, 2002). Therefore we will study the deviance reduction. The model has performed well, the deviance was reduced by 117.842 points, on 5 degrees of freedom, for a p-value of p<0.01, indicating that our model is an improvement over the null model. The table in figure 6 shows, the reduction in deviance for each term, added sequentially first to last, showing significant reduction in deviance for all factors included in the model.
Table 6: Deviance Reduction
DF Deviance Resid Df resid Dev pr(>chi)
Null 1555 2139.3
Block counter 1 49.145 1554 2090.2 2.377e-12
CE 1 48.993 1553 2041.2 2.59e-12
Fixation
Duration 1 12.894 1552 2028.3 0.0003297
SCL 1 6.810 1551 2021.5 0.0090642
3.8. Logistic Regression - Model interpretation
Logistic regression predicts the logit of an event outcome from a set of predictors. Because the logit is the natural log of the odds (or probability/[1–probability]), it can be transformed back to the probability scale. The resultant predicted probabilities can then be revalidated with the actual outcome to determine if high probabilities are indeed associated with events and low probabilities with non-events (no attention/Mind wandaring) (Peng, Lee, Ingersoll, 2002).
In logistic regression, we directly estimate the probability of an event occurring (because there are only two possible outcomes for the DV), so we can write down the estimate the probability of a subject attending as:
Prob(Attention==1) =
=
Where z is the linear combination of: Z = β 0 + β 1x1 + β 2x2 + β3x3 + β 4x4
26
Such that: Z =0.7269 – 0.02753 * Block counter + 0. 4835*CE -0. 00004466*Fixation Duration + 0. 462 *SCL
Table 7: Fixed effects, including predictor estimate, standard errors, z-values and pvalues for the
predictors in Model 5
PE SE z-value p
Intercept 0.7269 0.318 2.283 0.02246
Block counter -0.02753 0.0053530 -5.142 2.72e-7
CE -0.4835 0.005969 -8.099 5.55e-16
FD 4.466e-5 1.455e-5 3.070 0.00214
SCL 0.462 0.1697 2.736 0.00623
As our model is centered at block =5, the odds of attending in block 5 were 0.6741, so 67%.
This indicates that when block counter increases one unit (with 0 = block number 5), and all other factors remain constant, then Z =0.69937; and the probability of attentional involvement is 1/(1+e^-0.75173) = 0.66804. So the likelihood of attending in block 6 was 66.89%. If we moved ten blocks then the likelihood would only be 61.1%, and towards the end of the experiment (block 35) to 47.5%, indicating that half of the participants was attending to the task at the end of the experiment.
Similarly, when commission errors increases one unit, and all other factors remain constant, the probability of attending is, 0.3583, so 35.83%, indicating that if an CE occurred during block 5, a participant was more likely to mind wander. If participants had not inhibited their response on a target trial twice during a block, the odds of attending were only 0.2561, so 25.6%, and decreased to 17.51% for 3 commission errors and to 11.57% for four commission errors indicating that CE decrease the likelihood of attending on a block severely.
If fixation duration increases one second, and all other predictors remain constant, the chance of attending in block 5 becomes 0.6838, so 68.38% .When fixation duration increased from 0 to 10 seconds, the odds of attending became 76.37%.
Lastly, when SCL increases one standard deviation, and all other predictors remain constant, the odd of attending become 0.7665, so 76.6%, and when SCL increases two standard deviations, the odds become as high as 83.9%.
We conclude that this model could provide some interesting insights in the relationship between attention, task duration, erros, fixation duration and skin conductance levels. Next, as both block
27
counter and CE are negatively related to attention and FD and SCL positively, we will use this model to study how CE, CI and SCL interact with these factors in their predictability of attention in a block.
PART 3: The possible mediating effect of CI, CE and skin conductance on attention
Previous research has indicated that by internally increasing arousal levels, performance errors decreased and attention was sustained. Similarly, CI and CE are known to have an alerting function and they could possibly re-locate attention back to the task
(Zhang et al., 2012; Cheyne et al.,
2009; Jonker et al., 2012)
. We aim to assess these hypotheses in our data, and used our model as a reference to study the effect of CI, CE and SCL on the predictability of our independent variables on three new models.3.9. Model including CI
First, a model that included interactions for CI and CE, CI and FD and CI and SCL and all main effects was assessed (AIC =1904.9, 1963.7, loglik=-941.4). This model performed better then the model without interactions (chisq(5) = 19.266, p=0.00171. The Deviance Reduction compared to a null model was 140.4894, this is significant on 9 degrees of freedom (p<0.01), indicating that we are better in predicting attention using an interaction term for CI
Table 8: Deviance Reduction D F Deviance Resid Df resid Dev pr(>chi) Null 1555 2139.3
Block counter 1 49.149 1554 2090.2 2.377e-12 CI 1 7.280 1553 2082.9 0.006971 CE 1 41.721 1552 2041.2 1.053e-10 Fixation duration 1 12.886 1551 2028.3 0.000331 SCL 1 6.821 1550 2021.5 0.009008 CI*block_ctr 1 5.117 1549 2016.4 0.02368 CI*CE 1 15.798 1548 2000.6 7.04e-5 CI*FD 1 0.006 1547 2000.5 0.938277 CI*SCL 1 1.714 1546 1998.8 0.19052
28
3.10. Model Interpretation
There was an interaction between block counter and CI (PE=-1.089e-02, SE= 4.078e-03, z=-2.671,p= 0.007552). When block counter increased with 10 units, without a CI, the odds of attending would be 57.22%, however, when a CI occurred, this would be 59.1%. There also was an interaction between CE and CI (PE = 1.449e-01, SE=4.799e-02, z= 3.019,p=0.002537). When a CE occurred, but no CI, the odds of attending were 44.18%, but when also a CI occurred, the odds were 52.4%. Lastly, the interaction between CI and SCL (PE=-1.828e-01, SE=1.095e-01, z=-1.669,p= 0.095168) was almost significant, indicating that when SCL changed one SD, the odds of attending were 76.62%, but when one CI occurred, the odds of attending became 76.68%. The main effect of block counter reached no significance (PE=-0.011, SE=0.0089, z=-1.359, p=0.174), neither did the main effect of CI (PE=0.1857, SE=0.1443, z=1.624, p=0.1042), indicating that the effect of a CI is different over blocks, as shown by the significant interaction term.
Table 9: Predictor Estimates, Standard Errors, z-values and p-values of the predictors in the
interaction model. Estimate SE Zvalue p Intercept 0.401 0.3814 1.051 0.2931 Block counter -0.01100 0.00809 -1.359 0.17425 CI 0.1857 0.1443 1.624 0.1042 CE -0.6348 0.0855 -7.421 1.16-13*** FD 5.134e-5 1.54e-5 3.334 0.000856*** SCL 0.7863 0.2935 2.679 0.007385** CI*Blockcounter -0.01089 0.00407 -2.671 0.007552** CI*CE 0.1449 0.0479 3.019 0.002537**
CI*FD -3.97e-6 4.5e-6 -0.879 0.37951
CI*SCL -0.1828 0.1095 -1.669 0.09516.
3.11. Model including CE
We assessed models including interaction terms with the same variables but now for CE. However, non of the interactions was significant, and model fit was not improved compared to Model 5. For a model assessing the mediating effect of CE, non of the interactions was significant at the 0.05 level (AIC=1917.9, BIC=1966.0, loklik=-949.9). CE does not mediate the effects of block counter, FD and SCL.
29
3.12. Model including SCL
A model including interaction effects for SCL also could not improve model fit (AIC = 1917.0, BIC=1965.1, loklik=-949.5). When we studied the significance of predictors, non of the interaction terms reached significance, indicating that increases in arousal do not mediate the effect of errors, block duration and fixation duration as assessed in our experiment.
PART 4: Component Analysis
3.13. Combining Factors: exploratory factory analysis
As we collected data from multiple psycho-physiological measures, we were interested in whether the variables could be combined and studied whether we could find components in the data that could describe the data using fewer variables. Increasing the variable to subject ration (N/-ratio) could improve cross validity for our model to other data (Stevens, 2009) and PCA was conducted using an orthogonal (varimax) rotation.
We choose to only include components according to Kaisers criterion ( Eigenvalues >1 (Stevens, 2009). Loadings of the variables onto the components were assessed on theoretical and practical significance, based on the technique proposed by Stevens (2009; chapter 11). For the fourthy-five subjects in this experiment, loadings of >2(0.361) = .722 are indicated as significant. Table X only shows loadings of variables on our components. Each PC corresponds to the collective behavior of the original variables.
As can be seen in table 9, every component, except the first, is loaded by one variable. This indicates that the variables each are unique and cannot be described by their combined behavior. No sets of variables could be indentified that are highly interrelated.
Table 9: loadings on the PCA components for all variables
PC2 PC3 PC7 PC11 PC5 PC8 PC10 PC6 PC4 PC9 PC1 Nr CE 0.02522 -0.1765 0.0791 -0.0038 0.0281 0.00167 0.9720 0.0316 0.0123 -0.1196 0.0076 Nr CI 0.0357 0.9567 -0.1057 0.0009 -0.0136 -0.1220 -0.1897 -0.0216 0.0454 0.0786 0.0307 Nr AN 0.0998 -0.1020 0.9743 -0.0119 -0.0005 0.0758 0.0771 0.0345 -0.0264 -0.0904 -0.0055 Nr OM 0.9648 -0.1503 -0.0771 0.0579 0.0353 0.0316 -0.0109 0.0279 0.0000 -0.0997 0.0309 RTcv 0.8333 0.3410 0.3543 -0.0293 -0.0627 -0.0066 0.0611 0.0704 0.0214 -0.0334 -0.1059 Mean RT -0.1072 0.0720 -0.090 -0.0204 -0.0206 -0.0289 -0.1175 0.0200 -0.0201 0.9790 0.0146 SCL 0.0120 0.04119 -0.0243 -0.00938 -0.03467 -0.0784 0.0112 -0.0219 0.9945 -0.0191 0.0069
30 blink 0.0340 0.00060 -0.0144 0.960588 -0.0248 0.0615 -0.0037 0.1136 -0.0102 -0.0212 -0.2389 Block ctr 0.0236 -0.1149 0.0746 0.061736 0.0384 0.9756 0.0022 0.0677 -0.0831 -0.0291 -0.1045 FD -0.0397 0.03156 -0.0125 -0.30584 -0.2773 -0.1326 0.0099 -0.0791 0.0085 0.0179 0.8945 SAC 0.0643 -0.0173 0.0367 0.108064 0.0931 0.0667 0.0301 0.9815 -0.0221 0.0199 -0.0637 FC -0.0068 -0.0151 -0.0039 -0.02254 0.9697 0.0370 0.0287 0.0952 -0.0373 -0.0210 -0.2117 PC2 PC3 PC7 PC11 PC5 PC8 PC10 PC6 PC4 PC9 PC1 Nr CE 0.9719 Nr CI 0.9566 Nr AN 0.9743 Nr OM 0.9648 - RTcv 0.8333 Mean RT 0.9790 SCL 0.9945 blink 0.9605 blockctr 0.9757 FD 0.8945 SAC 0.9815 FC 0.9696
Table 10: SS loadings, proportion of explained and cumulative variance of principal components one to eleven.
PC2 PC3 PC7 PC11 PC5 PC8 PC10 PC4 PC9 PC1 SS loadings 1.66 1.12 1.11 1.04 1.01 1.01 1.01 1.00 1.00 0.93
Prop var 0.14 0.09 0.09 0.09 0.09 0.08 0.08 0.08 0.08 0.08
31
4. Discussion
The present study investigated how skin conductance, blink rate, fixation duration, fixation
count, saccade amplitude, performance and reaction time on the SART relate to mind
wandering. We have described a method of analysis which allowed us to estimate the
likelihood of attending based on the predictors block counter (task duration), CE, fixation
duration and skin conductance.
Our data revealed that the odds of attending are decreased when task duration and amount of
CE increased, and the odds of attending were increased with increases in SCL and fixation
duration. Our model significantly reduced deviance compared to a null model. A model
including interaction terms for CE and CI, FD and CI and SCL and CI, improved model fit
and showed that the effects of task duration, CE and SCL (on a 0.10 level) were mediated by
correct inhibitions, indicating that CI’s have an alerting function. This mediating effect was
not found for CE or SCL, and we could not confirm previous research relating CE and
increases in SCL to improvements in sustained attention. Lastly, we established that all of
these variables were unique, a component analysis showed that all eleven variables loaded on
one component, indicating that their combined analysis is redundant.
This study has some practical and theoretical implications. On the theoretical side, we
contribute to the idea that CI-trials increase the likelihood of attending as they mediate effects
of task duration and commission errors. Furthermore, our data add to the body of evidence
relating performance errors to episodes of mind wandering, and confirm relations found
between ocular changes and mind wandering, and skin conductance and mind wandering. On
the practical side, we showed that it is possible to predict a mind wandering episode based on
task duration, CE, fixation duration and skin conductance. This information could be used in
situations in which mind wandering is maladaptive, for example during car driving. We
suggest that a system could be build that tracks attention based on skin conductance and
fixation duration, that could alert people when their attention drifts of. This should be done
using stimuli on which the participant can behave well, as alerting stimuli on which the
participant makes an error (like CE), do not mediate the effect of attention and time on task.
32