UNIVERSITY OF AMSTERDAM
RESEARCH PROJECT I
Different Decision Modes: Are they Based on
Distinct Cognitive and Neural Mechanisms?
Author Joaquina Couto Supervisor dhr. dr. M. (Martijn) Wokke CoAssessor and UvA representative dhr. prof. dr. K.R. (Richard) Ridderinkhof Date: October 31, 2014 UvA student number: 10551395 Email: joaquinacouto@gmail.com Master’s programme: Brain and Cognitive SciencesAbstract
Different cognitive states may be adopted when making decisions. Many daily decisions emerge out of fast intuitive judgements in which little or no explicit knowledge seem to be present. In contrast, there are also decisions which require explicit deliberation, exploration of alternative actions and their consequences. In these cases, people make use of metacognitive knowledge and memory processes to regulate their own decisions. Despite the general consensus on the existence of different decision modes and metacognitive judgments, there is still a huge amount of disagreement when it comes to the cognitive and neural mechanisms underlying them. In addition, it is still not clear how these processes are related to each other. The present study aims to investigate how cognitive decision styles and confidence levels vary in terms of associated neural oscillations. We focus on the effects that feedback has depending on decisions mode (intuitive vs. rational), and on low or high levels of confidence. Although our findings are suggesting that different decision modes and confidence levels are based on similar cognitive and neural mechanisms, future studies are necessary to fully ascertain this claim. Keywords: Decisionmaking, Metacognition, Negative feedback, Neural oscillations
Introduction Many daily decisions emerge out of fast intuitive judgements in which little or no explicit knowledge seem to be present. Decisions involving some kind of expertise seem particularly to occur under these circumstances. Fire officers, for example, are able to recognize a situation as of a kind faced previously and rapidly retrieve a schema that provides a solution. There are, however, also some decisions which require explicit deliberation, exploration of alternative actions and their consequences. Decisions involving specialized knowledge in which the decision maker is still unfamiliar with may be a prime example of this type of decision. Taking the situation of fire officers as example, novices do not own sufficient knowledge to act in a rapid and effective way as experts do, so they usually depend on explicit analytic reasoning to accomplish their duties (Evans, 2008). Importantly, neither of these decision modes are perfect and they may result in errors. While intuitive decisions may fail when novel problems are presented, deliberative decisions may fail due to a wrong hypothesis testing (Evans, 2008). In order guide current and future behavior, people make use of metacognitive judgments and memory processes to accurately evaluate their own decisions. These judgments help people to avoid making the same errors twice and to evaluate whether they have enough information on which to base a reliable decision (Yeung, & Summerfield, 2014). Despite the general consensus on the existence of different decision modes and metacognitive judgments (e.g. low or high levels of confidence), there is still a huge amount of disagreement when it comes to the mechanisms underlying them. The most typical conceptual framework regarding decision modes with varying amounts of metacognition comes from the generic dualsystem theory, which dichotomizes two decision systems. System 1 (S1) is usually interpreted as being intuitive, unconscious, rapid, and automatic, whereas system 2 (S2) is regarded as being more deliberative, conscious, slow, and controlled (Evans, 2008, Price, & Norman, 2008, Glöckner, & Witteman, 2009). Based on this conception, most of the neuroscientific studies have focused on analysing reportable perceptual consciousness in an allornone manner (Gaillard, Dehaene, Adam, Clémenceau, Hasboun, Baulac, Cohen, & Naccache, 2009, Dehaene, & Changeux, 2011). In these studies, consciousness is assumed to reflect a discrete transition from a
state of no awareness to a state of full awareness (Miller, & Schwarz, 2014). Such findings have shed some light on both cognitive and neural mechanisms concerning consciousness access in perceptual decision making. However, they are still based on a dualprocessing approach.
Further approaches have advanced with different perspectives and have tried to soften the dichotomy between the two systems. Interventionist models, for example, have suggested that intuitive and deliberate decisions rely on the same basic automatic S1 processes, and that they may become complemented by additional deliberate operations from S2 if circumstances require a deliberate decision mode (Evans, 2008, Horstmann, Ahlgrimm, & Glöckner, 2009). In this respect, Glöckner and Witteman (2009) identified different types of processes underlying intuition which may or may not involve consciousness from S2. According to the authors, some intuitive decisions are based on automatic informationsampling processes (accumulative intuition), others however, involve a transformation of the accumulated information into mental representations, which although unconscious, its result enters awareness (constructive intuition). In the same line of thought, Price and Norman (2008) suggested intuitive processes (from S1) as informative conscious feelings (from S2) without conscious access to its antecedents. To support their idea, the authors focused in cases in which decisions are based on metacognitive judgments, such as the Feeling of Knowing (Koriat, 2000), or the Feeling of Rightness
(Thompson, Turner, & Pennycook, 2011), where we are capable of recognizing a correct response though we cannot currently recall it in an explicit manner, and suggested them to function as an intermediate point between conscious and unconscious processes. In addition to the disagreement on whether the mechanisms underlying S1 and S2 are based on two completely distinct processes or on two ends of a continuum, it is still not clear how decision modes and metacognitive judgments are related to each other. The present study aims to investigate how different decision modes and confidence levels vary in terms of cognitive processes and neural oscillations. We focus on the effects of feedback depending on decisions made intuitively or rationally, and whether participants experience low or high levels of certainty. Our main questions are a) how are neural dynamics affected by negative feedback in the different decision modes and degrees of certainty, b) how do learning by negative feedback evolve over time and vary depending on decisions made intuitively/rationally, or on low/high levels of confidence, and c) how are decision modes and certainty related to one another.
Although our study is exploratory in nature, some hypothesis are made in accordance with studies investigating neural mechanisms in response to typical negative feedback processing. We expect that intuitive and rational decisions will involve similar neural oscillations. In particular, we predict a frontocentral
feedbackrelated negativity (FRN) (Luu, Tucker, Derryberry, Reed, & Poulsen, 2003), an increase in thetaband oscillations in midfrontal areas (Cohen, Elger, & Ranganath, 2007, Cavanagh,
Zambrano‐Vazquez, & Allen, 2012), and an increase in longdistance synchrony within theta range
between midfrontal brain areas and sensorimotor areas (van de Vijver, Ridderinkhof, & Cohen, 2011) for both decision modes following a negative feedback. Although we predict rational and intuitive decisions to share similar topographic distributions, we predict rational decisions to elicit higher amounts of neural activity, namely greater FRN, stronger theta power, and stronger synchronization, in comparison to intuitive decisions. This last prediction is made in the light of interventionist models, as decision modes are hypothesized to be different in quantitative terms, but not in qualitative terms. The same predictions are made for low and high levels of confidence, although no strong a priori hypotheses are made regarding whether high levels of confidence will elicit higher amounts of neural activity or not. As for learning, some hypothesis are also made in accordance with studies investigating neural mechanisms in response to typical negative feedback. We expect the last blocks to contain greater FRN
(Cohen, & Ranganath, 2007, Philiastides, Biele, Vavatzanidis, Kazzer, & Heekeren, 2010), stronger
theta power (Cavanagh, Frank, Klein, & Allen, 2010, Cohen, Wilmes, & van de Vijver, 2011), and strong synchronization (Luft, Nolte, & Bhattacharya, 2013) in comparison to the first blocks. In
behavioral terms, less suboptimal decisions and more optimal decisions are expected in the last blocks as a result of learning. In this case, we also expect more rational decisions to be made in the last blocks. No clear predictions are made for confidence levels in this respect. Materials and Methods Participants In total, thirtyone subjects participated in the experiment. However, while one subject was immediately excluded due to misinterpretation of task instructions, fourteen/seventeen participants were excluded due to insufficient number of trials per condition (<20 intuitive/rational decisions or low/high confidence levels, respectively, either before or after artifact rejection). We ended up with a sample of thirty participants (8 male, 22 female, Mage = 22.67 years, age range: 1930 years) for the behavioral analysis. As for the EEG data, only fourteen participants (3 male, 11 female, Mage = 23.29 years, age range: 2030 years) were used to analyse decision modes, and sixteen (4 male, 12 female, Mage =
23.44 years, age range: 2027 years) to confidence levels. The subjects were recruited from the website https://www.lab.uva.nl/spt/ and had the opportunity to choose either a monetary reward or an amount of credits (ECTS) in order to fulfill course requirements. Informed consents were signed prior to the start of the experiment. The experiment and procedures were approved by the ethics review board of the Department of Psychology, University of Amsterdam. Behavioral task In order to test our hypotheses, a probabilistic implicit learning task was used (see Fig. 1). This task consists of six blocks, each of them containing 50 trials. In each trial, participants had to track a submarine as long as possible using the computer mouse (see Fig. 1A). The movements of the submarine could be horizontal, vertical or diagonal, each of these associated with different probabilities of danger. That is, while some moves were safe (vertical and horizontal moves), others were unsafe (diagonals), and the probability of the submarine being attacked (see Fig. 1B) was higher when dangerous moves were made (in a cumulative fashion). Participants were not told about these rules beforehand, such that they had to learn which moves were dangerous and which ones were not by themselves. At some point, participants would become skilled in avoiding danger, and would be able to make optimal decisions. That is, track the submarine long enough to collect a high amount of points, but stop tracking the submarine as soon as the trajectory became too dangerous (too many dangerous movements have been made). To stop the trial, participants just had to press the “emergency” space bar. At the end of each trial, participants were asked to indicate whether their decision was based on an intuition or rational thoughts. Next, on a scale from 1 to 5, participants were asked to mention how certain they were about their decision. In case of being attacked (suboptimal decisions), the procedure was the same. First, participants had to indicate whether their decision not to stop the trial was based on intuitive or rational thoughts, and second, how sure they were about their decision not to stop the trial. As the attacks occurred before participants’ reports, they were asked to try to be as reliable as possible in their responses. That is, they were instructed to mention whether they were very confident about their decision, even if that decision have resulted in an attack. By recording EEG signals while participants were performing the task, we were able to investigate to what extent intuitive and rational decisions, and low (13) and high (45) levels of confidence vary in terms of neural processing, especially, when
implicit learning fails and an attack occur. A. B. Fig. 1. The Yellow Submarine Task. A, The submarine moves in accordance with different movements (horizontal, vertical, and diagonal), each one associated with different probabilities of danger. B, If too many dangerous movements are made, the probability of predators attack the submarine is higher. EEG data recordings EEG data were recorded and sampled at 1048 Hz using a BIOSEMI Active Two electrode system. Sixtyfour scalp electrodes were measured, and four ocular electrodes (two HEOG and two VEOG) were used to detect horizontal and vertical eye movements. After acquisition, sampling rate was resampled to 512 Hz and the data was filtered using a lowpass filter of 50 Hz and a highpass filter of 0.5 HZ. EEG segments contaminated with eye blinks were detected and removed across recordings on the basis of the Independent Component Analysis (ICA). Other artifacts (e.g. noisy data) were further detected and removed by visual inspection, though artifact rejection based on ICA was also done when necessary in this phase. As we were interested in negative feedback, data epochs of 3.5 seconds (1.52 s) were selected relative to the moment when participants were attacked. This epoch window ensured that edge effects would not contaminate our timewindow of interest. Epoched data was further divided according to intuitive or rational decisions (decision modes), and low or high levels of confidence (confidence levels). All the preprocessing steps were done using Brain Vision Analyser and the EEGLAB toolbox in Matlab (Delorme, & Makeig, 2004). Eventrelated potentials (ERPs), timefrequency representations (TFR) and phase synchrony were also computed in Matlab using FieldTrip toolbox (Oostenveld, Fries, Maris, & Schoffelen, 2010).
EEG statistical analysis Statistical analyses of ERPs were performed by comparing the mean of ERP values from postfeedback timewindows (100500 ms). A paired ttest was used in order to analyse intuitive vs rational and low vs high conditions. To analyse decision modes (rationalintuitive) vs confidence levels (highlow), an independent ttest was used instead. As for the timefrequency representation (TFR) and phase synchronization, power values were firstly normalized with respect to a prestimulus baseline (−5000 ms). After that, statistical analyses were conducted. In case of TFR, average power values from postfeedback timewindows (100500 ms) were entered into a 2 (decision mode: intuitive or rational) × 4 (channel poolings: [FCz, FC1, FC2, Fz, Cz], C3, C4, Oz) and a 2 (confidence levels: low or high) × 4 (channel poolings: [FCz, FC1, FC2, Fz,Cz], C3, C4, Oz) repeated measures ANOVAs, fortheta (47 Hz) and deltaband (24Hz) separately. In order to analyse decision modes (rationalintuitive) vs confidence levels (highlow), a mixed ANOVA was applied instead for both frequency bands: 2 (conditions: rationalintuitive or highlow) × 4 (channel poolings: [FCz, FC1, FC2, Fz, Cz], C3, C4, Oz). In case of phase synchronization, average values of coherence from postfeedback timewindows (100500 ms) were entered into a 2 (decision mode: intuitive or rational) × 4 (channel pairs: [FCzC3], [FCzC4], [FCzOz]) and a 2 (confidence levels: low or high) × 4 (channel pairs: [FCzC3], [FCzC4], [FCzOz]) repeated measures ANOVAs, fortheta (47 Hz) and deltaband (24Hz) separately. Again, to analyse decision modes (rationalintuitive) vs confidence levels (highlow), a mixed ANOVA was applied for both frequency bands: 2 (conditions: rationalintuitive or highlow) × 4 (channel pairs: [FCzC3], [FCzC4], [FCzOz]). We included deltaband in our analysis because, despite theta being the most reported frequency band among research investigating feedback processing (Luft, 2014), deltaband has also been associated with prediction error (Bernat, Nelson, Holroyd, Gehring, & Patrick, 2008). The Greenhouse–Geisser correction was used to adjust the degrees of freedom when necessary. It is important to mention that the statistical tests described above were not the ones initially in mind. We intended to use a within subjects design to look at differences between decision modes and confidence levels and how those interact to each other. This analysis would be performed by only applying repeated measures ANOVAs for ERPs, TFR and phase synchronization. However, due to an insufficient number of trials for several participants, the use of such statistical tests was not possible. Under the circumstances, we decided to use, on the one hand, a within subject design to look at
differences between intuitive and rational decisions, as well as between low and high levels of confidence, while on the other hand, a between subjects design to look at the interactions between decision modes and confidence levels. This gravely limited our possibilities on finding out how confidence and decision modes are related to each other concerning the third main question of our study. Also, due to the insufficient number of trials, we were not able to investigate the second main question of our study in neural terms how do learning by negative feedback evolve over time and vary depending on decisions made intuitively/rationally, or on low/high levels of confidence. For those reasons, we just reported the results regarding our first main question, how are neural dynamics affected by negative feedback in the different decision modes and degrees of certainty (which corresponds to the within subject design) and the behavioral results regarding our second main question. Results Behavioral data On average, participants were attacked on 23% of the trials. Fig. 2 displays the percentages of the attacks in which participants decided intuitively or rationally (see Fig. 2A), and when they reported low or high levels of confidence (see Fig. 2B). It is clear that when participants were attacked decisions were mostly made intuitively (M = 67.77, SE = .20) than rationally (M = 31.99, SE = .20), t(29) = 4.913, p < .001. We can also see that attacks occurred most of the times when participants reported low levels of confidence (M = 61.97, SE = .23) than high levels (M = 37.73, SE = .23), t(29) = 2.880, p = .007.
A. B. Fig. 2. Behavioral results to negative feedback. A, Percentage of different decision modes that resulted in attacks. B, Percentage of different confidence levels that resulted in attacks. In Fig. 3, we can see the percentages of attacks that occurred on the three first and last blocks for different decision modes (see Fig. 3A) and confidence levels (see Fig. 3B). A twoway ANOVA revealed a significant main effect of decision modes, F(1,29) = 24.13, p < .001, r = .45, but no significant main effect of blocks F(1,29) = 1.94, p = .174. Interestingly, a marginal significant interaction effect between decision modes and blocks was observed F(1,29) = 2.81, p = .105. Further ttests revealed that the percentage of attacks based on intuitive decisions was higher in the first blocks (M = .399, SE = .040) than in the last blocks (M = .278, SE = .038). This difference was towards a statistically significant difference, t(28) = 1.989, p = .057. In contrast, there is no significant difference on the percentage of attacks based on rational decisions between the first blocks and last blocks, t(28) = .173, p = .864. A twoway ANOVA also revealed a significant main effect of confidence levels, F(1,29) = 8.30, p = .007, r = .22, and no significant main effect of blocks, F(1,29) = 1.87, p = .183. In this case, no significant interaction effect between confidence levels and blocks was observed, F(1,29) = 1.88, p = .181.
A. B. Fig. 3. Behavioral results for learning by negative feedback. A, Percentage of different decision modes that resulted in attacks in the beginning and in the end of the task. B, Percentage of different confidence levels that resulted in attacks in the beginning and in the end of the task. EEG data ERP responses to negative feedback Fig. 4A shows feedbacklocked ERPs over the midfrontal electrodes (FCz, FC1, FC2, Fz, Cz) for trials containing intuitive and rational decisions, as well as the difference wave (rational decisions intuitive decisions). As we can see from the figure, rational decisions (red lines) elicited a larger FRN component (M = 35.29, SE = 7.17) than intuitive decisions (blue lines) (M = 31.31, SE = 5.75). A dependent ttest revealed that this difference was trending towards a statistically significant difference, t(13) = 2.096, p = .056. The topography of the activation averaged over the identified FRN timewindow (transparent grey box) is presented in Fig. 4B. We can see a difference in heads topographic maps, namely a distribution towards midfrontal electrodes.
A. B. Fig. 4. ERP responses to negative feedback. A, FRN over the midfrontal electrodes (FCz, FC1, FC2, Fz, Cz) for intuitive and rational decisions, as well as the difference wave (rational decisions intuitive decisions). B, Respective topographic distributions from 300400 ms after negative feedback. As for the trials in which participants reported low and high levels of confidence (see Fig. 5A), we observed a larger midfrontal FRN component when participants reported high levels of confidence on their decisions (red lines) (M = 25.79, SE = 4.92) than low levels (blue lines) (M = 22.82, SE = 4.65). This difference was statistically significant, t(15) = 2.638, p = .019. We can see that the distribution of FRN also occurred differently in topographic maps, again, towards midfrontal electrodes (see Fig. 5B).
A. B. Fig. 5. ERP responses to negative feedback. A, FRN over the midfrontal electrodes (FCz, FC1, FC2, Fz, Cz) for low and high levels of confidence, as well as the difference wave (high confidence levels low confidence levels). B, Respective topographic distributions from 300400 ms after negative feedback. Oscillatory responses to negative feedback In Fig. 6, TF activity for intuitive (see Fig. 6A) and rational decisions (see Fig. 6B) is displayed, as well as their respective topographical distributions. We can see that there is an increase in theta and delta power following negative feedback in the time period marked by the black frame for both intuitive and rational decisions. Repeated measures ANOVAs revealed no significant main effect of decision modes in theta, F(1,13) = 1.92, p = .183, or delta power, F(1,13) = .74, p = .406. There was, however, a significant main effect of channels location at both theta F(3,39) = 12.78, p < .001, r = 50, and deltaband, F(1.68,19.18) = 29.20, p < .001, r = .69. This result suggests that different poolings of channels had different effects on these two frequency bands regardless of what decision mode was adopted. To break down this interaction, posthoc tests were performed comparing channels C3, C4, Oz and a cluster of electrodes (FCz, FC1, FC2, Fz, Cz), our area of interest. Posthoc comparisons using Bonferroni test revealed that theta and delta power were significantly higher in the midfrontal electrodes (FCz, FC1, FC2, Fz, Cz) than in the channels C3 (p =.004 for theta, p =.002 for delta), C4 (p
<.001 for theta, p <.001 for delta ), and Oz (p =.056 for theta, p <.001 for delta). No significant difference was observed between C3 and C4, C3 and Oz, and C4 and Oz at both frequency bands. This increase in theta and delta power over the midfrontal area for both decision modes can also be evidenced in Fig 6A and Fig 6B, in the topographic plots. No significant interaction effect between decision modes and channels location was observed at theta, F(3,39) = 1.84, p = .155, or delta, F(3, 39) = 1.31, p = .285. A. B. Fig. 6. Oscillatory responses to negative feedback. A, Timefrequency for intuitive decisions and respective topographic distributions. B, Timefrequency for rational decisions and respective topographic distributions. Timefrequency is computed for a pooling of midfrontal channels (FCz, FC1, FC2, Fz, Cz). The TFR for trials in which participants reported low and high levels of confidence are presented in Fig. 7A and Fig. 7B, respectively. As in the previous analysis, we observed an increase in theta and delta power from 200 to 500 ms following an attack, either for participants who reported low or high levels of confidence. Again, repeated measures ANOVAs revealed no significant main effect of confidence levels in theta, F(1,15) = .20, p = .660 or delta power, F(1,15) = .77, p = .395. Nevertheless, a significant main effect of channels location was observed at theta F(3,45) = 8.17, p < .001, r = .34, and deltaband, F(1.52,22.84) = 14.35, p < .001, r = .49. Posthoc comparisons revealed that theta and delta
power were significantly stronger in the in the pooling of channels [FCz, FC1, FC2, Fz, Cz] than in the channels C3 (p =.005 for theta, p =.003 for delta), and C4 (p <.001 for theta, p =.002 for delta). In contrast to theta, delta power was also significantly stronger in the midfrontal area compared to Oz (p =.018). No significant difference was observed between C3 and C4, C3 and Oz, and C4 and Oz. Topographic plots show, indeed, that this increase in theta and delta power occurred over the midfrontal area for both low (see Fig. 7A) and high levels of confidence (see Fig. 7B). Again, there was no significant interaction effect between confidence levels and channels location at theta, F(3,45) = .75, p = .530, or delta, F(3,45) = .13, p = .943. A. B. Fig. 7. Oscillatory responses to negative feedback. A, Timefrequency for low confidence levels and respective topographic distributions. B, Timefrequency for high confidence levels and respective topographic distributions. Timefrequency is computed for a pooling of midfrontal channels (FCz, FC1, FC2, Fz, Cz). Intersite phase synchronization responses to negative feedback Fig. 8 shows the effects of decision modes and intersite synchronization at theta (see Fig. 8A) and deltaband (see Fig. 8B) from repeated measures ANOVAs. As it is shown in the figure, there was no significant main effect of decision modes at theta, F(1,13) = 1.33, p = .270, or at deltaband, F(1,13) = .01, p = .919. Although there was no significant main effect of channel pairs at theta, F(1.16,15.04) =
1.20, p = .299, there was a significant main effect of channel pairs at deltaband, F(2,26) = 6.14, p = .007, r = 326. This result indicates that different channel pairs had different effects on delta synchronization regardless of what decision mode was applied. Posthoc comparisons using Bonferroni test revealed that delta synchronization was significantly higher between FCz (midfrontal area) and C3 (contralateral motor area) than between FCz and C4 (ipsilateral motor area) (p =.041). No significant difference was observed in synchronization between other channel pairs, namely FCz with C3 compared to FCz with Oz, and FCz with C4 compared to FCz with Oz. No significant interaction effect between decision modes and channel pairs was observed at both frequency bands, F(1.45,18.80) = .89, p = .397 for theta, and F(2,26) = 1.02, p = .376 for delta. A. B. Fig. 8. Intersite phase synchronization responses to negative feedback. A, Main effects of decision modes and intersite synchronization at thetaband. B, Main effects of decision modes and intersite synchronization at deltaband. The main effects of confidence levels and intersite synchronization at theta (see Fig 9A) and deltaband (see Fig. 9B) are displayed in Fig. 9. There was no significant main effect of confidence levels, F(1,15) = .82, p = .380, no significant main effect of channel pairs, F(2,30) = 1.66, p = .208, and no significant interaction effect between confidence levels and channel pairs at thetaband, F(1.14,17.15) = .36, p = .585. As for delta, although there was no significant main effect of confidence levels, F(1,15) = .01, p = .913, and no significant main effect of channel pairs, F(2,30) = 2.59, p = .091, the interaction effect between confidence levels and channel pairs was trending towards a statistically significant difference, F(2,30) = 3.17, p = .057.
A. B. Fig. 9. Intersite phase synchronization responses to negative feedback. A, Main effects of confidence levels and intersite synchronization at thetaband. B, Main effects of confidence levels and intersite synchronization at deltaband. Discussion In the present study, we investigated how different decision modes and confidence levels come about in time during an implicit learning task and how these vary in terms of cognitive and neural mechanisms. We focused on the effects that feedback has depending on decisions made intuitively or rationally, and on low or high levels of certainty. Overall, the findings we got suggest that different decision modes and confidence levels are based on similar cognitive and neural basic mechanisms. Nevertheless, future studies are necessary to fully ascertain this claim. Behavioral effects of negative feedback We found that attacks occurred mostly when participants were deciding intuitively or with low levels of confidence. Although these results do not seem to support the interventionist models, as optimal decisions were found to rely more often on analytic reasoning, than on intuitive automatic processes (Evans, 2008), we cannot dismiss the fact that these findings may be based on biased reports. The task we used was developed in a way that the attacks occurred before participants’ reports on whether their decisions were based on intuitive/rational thoughts and low/high levels of confidence. Participants were instructed to be as reliable as possible in their responses. Even so, at the moment of their reports, they already knew their decisions not to stop the trial had not been successful, and this
information could have influenced their judgments somehow. As for the behavioral results regarding learning, we observed relatively the same amount of attacks across the blocks, regardless of being based on intuitive/rational decisions, and low/high levels of confidence. These results may be interpreted in light of recent findings within implicit learning paradigms. According to them, participants who have implicitly learned a solution to a problem, may still perform the task at chance (Norman, Price, Duff, & Mentzoni, 2007, Price, & Norman, 2008). Thus, even if participants have implicitly learned (regardless of decision modes and confidence levels) that diagonal movements were more dangerous than vertical/horizontal ones, this learning could have not been reflected in the percentage of attacks. Perhaps, analysing the average of points in the beginning (bloks 1/2/3) and in the end of the task (blocks 4/5/6) would allows us to make more accurate inferences about learning by negative feedback. Electrophysiological effects of negative feedback We observed a feedbackrelated negativity (FRN) component, as well as an increase in thetaband oscillations in midfrontal areas for both decision modes and confidence levels following a negative feedback. We also observed an increase in thetaband synchronization between midfrontal electrodes and other channels, namely in the motor areas (C3/C4) and occipital areas (Oz). Interestingly, although we did not observed a significant difference in theta synchronization between different pairs of channels, synchronization between midfrontal areas (FCz) and motor related electrodes ipsilateral to the hand tracking (C4) appeared to be somehow suppressed in comparison to synchronization between the other pairs of channels, between FCz and C3 (contralateral motor area), and between FCz and Oz especially when analysing decision modes. Altogether, these results are in line with studies investigating neural mechanisms in response to feedback. Previous work in this field has shown a frontocentral voltage deflection in the ERP following the presentation of negatively valenced feedback around 250300 ms (Luu et al., 2003, Nieuwenhuis, Holroyd, Mol, & Coles, 2004). Growing literature has also identified frontal midline theta oscillations (48 Hz) in the generation of those ERPs during conflict and error response (Cavanagh et al., 2012, Cohen, et al., 2007, Cohen, Elger, & Fell, 2009). And, although scarce, most of studies addressing intersite synchronization have also suggested midfrontal brain areas to communicate with sensorimotor areas in theta frequency range in response to negative feedback
(Cavanagh, Cohen, & Allen, 2009, van de Vijver et al., 2011). A large body of research has particularly proposed midfrontal areas to communicate with contralateral sensorimotor areas (C3/C5) not with ipsilateral sensorimotor areas (Luft, 2014), and this may perhaps explain why we observed a kind of suppression in synchronization between FCz and C4 (ipsilateral motor area). Our results are also in line with the interventionist models as different decision modes and confidence levels shared similar topographic distributions, namely the midfrontal topographic distribution of FRN and thetaband oscillations/synchronization, as the literature of typical feedback also suggests. In addition, intuitive/rational decisions, and low/high levels of confidence shared relatively the same amounts of TFR. This result refutes our hypothesis that rational decisions would elicit stronger theta power than intuitive decisions although different decision modes were expected to rely on the same basic mechanisms (in qualitative terms), rational decisions were still expected to require additional amounts of brain activity (in quantitative terms). Interestingly, intuitive/rational decisions, and low/high levels of confidence shared different amounts of ERPs, with intuitive and high levels of confidence eliciting higher FRN. This difference was more significant when analysing confidence levels, and appeared to be particularly prominent around the timewindow 250350 ms following an attack. Although we were not initially interested in analysing deltaband, we decided to do so, because deltaband has also been associated with prediction error (Bernat, Nelson, Holroyd, Gehring, & Patrick, 2008), and because visual inspection of TFR plots showed predominant power in this range in the midfrontal area. Similarly to the results obtained at thetaband oscillations, we observed an increase in deltaband oscillations/synchronization towards midfrontal areas for both decision modes and confidence levels following a negative feedback. In case of deltaband, we actually found a significant difference in synchronization between different pairs of channels (see above), with higher synchronization occurring between midfrontal and contralateral motor areas (C3) compared to ipsilateral motor areas (C4). Altogether, these findings support the literature which has suggested a role of deltaband in feedback processing. While theta has only been associated with negative outcome valence, delta has been associated with a variety of stimulus attributes, including outcome valence (mostly positive valence), outcome magnitude (large or small), and outcome probability (unlikely or likely) (Bernat et al., 2008). Although we only have analysed negative feedback which is supposed to
elicit thetaband, delta activity could be well explained by the secondary feedback processes, such as outcome magnitude and probability, that are part of our implicit learning task. The results regarding deltaband are also in line with the interventionist models. In this frequency range, different decision modes and confidence levels not only shared similar topographic distributions, but also relatively the same amounts TFR and synchronization. Limitations The fact that we did not observe significant differences in amounts of TFR and synchronization for intuitive/rational decisions and low/high levels of confidence at both theta and delta bands should be interpreted with caution. A small sample size and reduced number of minimum trials stipulated per condition (20 trials) could have well justified the statistical insufficient power that we obtained (for more details, see Material and Methods). The reduced minimum number of trials was also quite problematic for a suitable choice of statistical tests (for details, see Material and Methods). The fact that we had to use, on the one hand, a within subject design to look at differences between intuitive and rational decisions, as well as between low and high levels of confidence, and on the other hand, a between subjects design to look at the interactions between decision modes and confidence levels, gravely limited our possibilities on finding out how confidence and decision mode are related to each other concerning the third main question of our study. The reduced number of trials also limited our possibilities on finding out how learning by negative feedback evolve over time and vary neurally depending on decisions made intuitively/rationally, or on low/high levels of confidence concerning the second main question of our study. We planned to analyse this by splitting the six blocks into two time bins like we did for the behavioral analysis (for more details, see Results). This would allow us to see if the FRN, theta power and theta synchronization would be stronger in the last blocks compared to the first blocks. Unfortunately, we were not able to further test those hypotheses. Despite all the methodological limitations, we consider this study to convey important evidence in favor of interventionist models. This evidence can serve as an incentive for future researchers to end up the disagreement on which cognitive and neural mechanisms underlie different decision modes and metacognitive judgments. Future studies using more robust methodologies are still needed, however. To successfully overcome the methodological limitations faced in our study, we suggest future studies to
recruit more participants. Changing the task in a way that attacks would be more frequent is also a possibility. This could be done, for example, by changing the probabilities which are associated with danger. Another suggestion is repeating the task with one or two extra sessions. Altogether, these additional precautions would ensure a higher number of total trials, and consequently, an increase of the threshold would be possible. Thereafter, a within subjects analysis could be carried out, and more accurate interpretations would be possible.
References Bernat, E. M., Nelson, L. D., Holroyd, C. B., Gehring, W. J., & Patrick, C. J. (2008, August). Separating cognitive processes with principal components analysis of EEG timefrequency distributions. In Optical Engineering+ Applications (pp. 70740S70740S). International Society for Optics and Photonics. Cavanagh, J. F., Cohen, M. X., & Allen, J. J. (2009). Prelude to and resolution of an error: EEG phase synchrony reveals cognitive control dynamics during action monitoring. The Journal of Neuroscience, 29(1), 98105. Cavanagh, J. F., Frank, M. J., Klein, T. J., & Allen, J. J. (2010). Frontal theta links prediction errors to behavioral adaptation in reinforcement learning.Neuroimage, 49(4), 31983209. Cavanagh, J. F., Zambrano‐Vazquez, L., & Allen, J. J. (2012). Theta lingua franca: A common mid‐frontal substrate for action monitoring processes.Psychophysiology, 49(2), 220238. Cohen, M. X., Elger, C. E., & Fell, J. (2009). Oscillatory activity and phase–amplitude coupling in the human medial frontal cortex during decision making.Journal of Cognitive Neuroscience, 21(2), 390402. Cohen, M. X., Elger, C. E., & Ranganath, C. (2007). Reward expectation modulates feedbackrelated negativity and EEG spectra. Neuroimage, 35(2), 968978. Cohen, M. X., & Ranganath, C. (2007). Reinforcement learning signals predict future decisions. The Journal of neuroscience, 27(2), 371378. Cohen, M. X., Wilmes, K. A., & van de Vijver, I. (2011). Cortical electrophysiological network dynamics of feedback learning. Trends in cognitive sciences, 15(12), 558566. Dehaene, S., & Changeux, J. P. (2011). Experimental and theoretical approaches to conscious processing. Neuron, 70(2), 200227. Delorme, A., & Makeig, S. (2004). EEGLAB: an open source toolbox for analysis of singletrial EEG dynamics including independent component analysis. Journal of neuroscience methods, 134(1), 921. Evans, J. S. B. (2008). Dualprocessing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol., 59, 255278. Gaillard, R., Dehaene, S., Adam, C., Clémenceau, S., Hasboun, D., Baulac, M., ... & Naccache, L. (2009). Converging intracranial markers of conscious access.PLoS biology, 7(3), e1000061. Glöckner, A., & Witteman, C. (2010). Beyond dualprocess models: A categorisation of processes underlying intuitive judgement and decision making.Thinking & Reasoning, 16(1), 125. Horstmann, N., Ahlgrimm, A., & Glöckner, A. (2009). How distinct are intuition and deliberation? An eyetracking analysis of instructioninduced decision modes (No. 2009, 10). Preprints of the Max
Planck Institute for Research on Collective Goods. Koriat, A. (2000). The feeling of knowing: Some metatheoretical implications for consciousness and control. Consciousness and cognition, 9(2), 149171. Luft, C. D. B. (2014). Learning from feedback: The neural mechanisms of feedback processing facilitating better performance. Behavioural brain research,261, 356368. Luft, C. D. B., Nolte, G., & Bhattacharya, J. (2013). Highlearners present larger midfrontal theta power and connectivity in response to incorrect performance feedback. The Journal of neuroscience, 33(5), 20292038. Luu, P., Tucker, D. M., Derryberry, D., Reed, M., & Poulsen, C. (2003). Electrophysiological responses to errors and feedback in the process of action regulation. Psychological Science, 14(1), 4753. Miller, J., & Schwarz, W. (2014). Brain signals do not demonstrate unconscious decision making: an interpretation based on graded conscious awareness.Consciousness and cognition, 24, 1221. Nieuwenhuis, S., Holroyd, C. B., Mol, N., & Coles, M. G. (2004). Reinforcementrelated brain potentials from medial frontal cortex: origins and functional significance. Neuroscience & Biobehavioral Reviews, 28(4), 441448. Norman, E., Price, M. C., Duff, S. C., & Mentzoni, R. A. (2007). Gradations of awareness in a modified sequence learning task. Consciousness and Cognition,16(4), 809837. Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J. M. (2010). FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational intelligence and neuroscience, 2011. Philiastides, M. G., Biele, G., Vavatzanidis, N., Kazzer, P., & Heekeren, H. R. (2010). Temporal dynamics of prediction error processing during rewardbased decision making. Neuroimage, 53(1), 221232. Price, M. C., & Norman, E. (2008). Intuitive decisions on the fringes of consciousness: Are they conscious and does it matter. Judgment and Decision Making, 3(1), 2841. Thompson, V. A., Prowse Turner, J. A., & Pennycook, G. (2011). Intuition, reason, and metacognition. Cognitive psychology, 63(3), 107140. van de Vijver, I., Ridderinkhof, K. R., & Cohen, M. X. (2011). Frontal oscillatory dynamics predict feedback learning and action adjustment. Journal of cognitive neuroscience, 23(12), 41064121. Yeung, N., & Summerfield, C. (2014). Shared Mechanisms for Confidence Judgements and Error Detection in Human Decision Making. In The Cognitive Neuroscience of Metacognition (pp. 147167). Springer Berlin Heidelberg.