NeuroImage 86 (2014) 503–513

Contents lists available at ScienceDirect

NeuroImage journal homepage: www.elsevier.com/locate/ynimg

Subthreshold muscle twitches dissociate oscillatory neural signatures of conflicts from errors Michael X Cohen a,b,⁎, Simon van Gaal a,c,d,e a

Department of Psychology, University of Amsterdam, The Netherlands Department of Physiology, University of Arizona, United States Inserm, Cognitive Neuroimaging Unit, Gif-sur-Yvette, France d Commissarìat à l'Energie Atomique, Neurospin Center, Gif-sur-Yvette, France e Radboud University Nijmegen, Donders Institute for Brain, Cognition and Behavior, The Netherlands b c

a r t i c l e

i n f o

Article history: Accepted 21 October 2013 Available online 1 November 2013 Keywords: Medial frontal cortex Theta Partial errors Oscillations Cognitive control Errors EEG Connectivity Time-frequency

a b s t r a c t We investigated the neural systems underlying conflict detection and error monitoring during rapid online error correction/monitoring mechanisms. We combined data from four separate cognitive tasks and 64 subjects in which EEG and EMG (muscle activity from the thumb used to respond) were recorded. In typical neuroscience experiments, behavioral responses are classified as “error” or “correct”; however, closer inspection of our data revealed that correct responses were often accompanied by “partial errors” — a muscle twitch of the incorrect hand (“mixed correct trials,” ~13% of the trials). We found that these muscle twitches dissociated conflicts from errors in time-frequency domain analyses of EEG data. In particular, both mixed-correct trials and full error trials were associated with enhanced theta-band power (4–9 Hz) compared to correct trials. However, full errors were additionally associated with power and frontal–parietal synchrony in the delta band. Single-trial robust multiple regression analyses revealed a significant modulation of theta power as a function of partial error correction time, thus linking trial-to-trial fluctuations in power to conflict. Furthermore, single-trial correlation analyses revealed a qualitative dissociation between conflict and error processing, such that mixed correct trials were associated with positive theta-RT correlations whereas full error trials were associated with negative delta-RT correlations. These findings shed new light on the local and global network mechanisms of conflict monitoring and error detection, and their relationship to online action adjustment. © 2013 Elsevier Inc. All rights reserved.

Introduction Several cognitive control processes, including response conflict monitoring and error processing, rely on brain structures within the medial prefrontal cortex (Nachev, 2006; Ridderinkhof et al., 2004b; van Veen and Carter, 2006). Response conflict arises when multiple response options are activated and only one must be selected, whereas error processing occurs when an incorrect response is made. Some have argued that conflict and errors are processed by the same neural system (van Veen and Carter, 2006; Yeung et al., 2004), on the basis of cognitive models and similar topographical distributions of EEG during conflict and error trials, and spatially overlapping patterns of activation in fMRI studies (Ridderinkhof et al., 2004a). Others have argued that errors and conflicts are processed by different neural systems (Falkenstein et al., 2000; Swick and Turken, 2002) and may recruit somewhat dissociable spatial regions within the medial frontal cortex (Mathalon et al., 2003; Nee et al., 2011; Ullsperger and von Cramon, 2001).

⁎ Corresponding author at: Department of Psychology, University of Amsterdam, The Netherlands. E-mail address: [email protected] (M.X. Cohen). 1053-8119/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.neuroimage.2013.10.033

Whether errors and conflict lead to the same neurocognitive process can be difficult to test empirically, because errors often occur when conflict is already present. Conflict, on the other hand, should be easier to isolate from errors. Typically, conflict effects are examined by comparing trials in which conflict is induced by the experiment with trials in which conflict is not induced by the experiment. This occurs, for example, in the Stroop task, when the word RED is printed in blue ink. A valid interpretation of condition differences relies on the assumption that subjects experienced response conflict in one condition and not in the other. Although there are clear behavioral condition differences that support this assumption–reaction times (RTs) are generally longer and error rates higher in conflict conditions–there is also thought to be conflict during conditions that supposedly contain no conflict (Coles et al., 2001), and there are fluctuations in cognitive control that affect how much conflict is experienced on each trial, depending on previous trial and other contextual events (Egner, 2007; Gratton et al., 1988). Thus, a more ideal way to test for conflict would be to measure it directly. One approach is to perform trial-to-trial brain-behavior analyses, wherein trial-varying brain activity is correlated with trial-varying RTs. However, RTs can vary across trials for a number of reasons unrelated to conflict, including general attention and other non-specific cognitive factors (Carp et al., 2011; Esterman et al., 2012; Weissman et al., 2006). Comparing

504

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

brain activity-RT correlations between conditions helps minimize some of the general contributors to RT variance (Cohen and Cavanagh, 2011), but still, RT does not directly measure response conflict. Here we sought to measure conflict more directly, and dissociate it from full errors, by recording subthreshold muscle twitches (measured via electromyography; EMG) from the thumbs that subjects used to indicate responses. The idea is that if the subject twitches the muscle of the incorrect hand but then pressed the correct button, then both responses were partially activated, but only one was fully engaged (Coles et al., 1995). This arguably provides a more accurate measure of endogenous conflict compared to averaging all trials in which the experimenter hopes that the subject experienced conflict. Furthermore, the EMG data provide a single-trial estimate of the amount of conflict experienced, as measured either by the time lag between the onset of the muscle twitch of the incorrect response and the onset of the muscle used to press the correct button (hereafter: correction time), or by the strength of the EMG response. Subthreshold muscle twitches during correct trials are termed “partial errors” (Allain et al., 2009; Gratton et al., 1988). We here call the correct trials in which partial errors occur “mixed correct” (MC) trials, and contrast them with “pure correct” (PC) trials, in which only the thumb corresponding to the correct response was activated, and with “full error” (FE) trials, in which only the thumb corresponding to the incorrect response was activated. Without measuring EMG activity, subthreshold muscle twitches go undetected, and partial errors are classified as correct responses. Surprisingly few investigations have studied partial errors, although they are known to be slow (Coles et al., 1995; Szucs et al., 2009), elicit post-trial slowing (Allain et al., 2009), and elicit some electrophysiological activity associated with error processing (Burle et al., 2008; Carbonnell and Falkenstein, 2006; Endrass et al., 2008; Masaki et al., 2012). Partial errors have also been used to probe memory (Seymour and Schumacher, 2009). EEG studies of brain circuits that support conflict and error processing typically focus on the error-related-negativity (ERN) or the stimuluslocked N2 (Nieuwenhuis et al., 2003; Yeung et al., 2004) or on a conflict-modulation of the correct-trial-related ERN (called the CRN). Others have highlighted that activity in a broad theta-band range (~2–8 Hz), maximal over midfrontal scalp sites (typically maximal at electrode FCz), increases with conflict and errors, correlates with RT, and predicts post-error adjustments (Cavanagh et al., 2009; Cohen, 2011b; Cohen and Cavanagh, 2011; Hanslmayr et al., 2008; Nigbur et al., 2011; Trujillo and Allen, 2007; Yordanova et al., 2004). In fact, although the error-related activity is often labeled as theta-band, visual inspection of time-frequency plots often suggests that the error-related activity extends lower, into the delta band (Yordanova et al., 2004). This might simply reflect frequency smoothing resulting from timefrequency decomposition (Cohen, 2014), or it could reflect the errors actually being processed by neural networks that operate in the delta band (Yordanova et al., 2004). This is an important distinction because different but temporally overlapping cognitive processes might be dissociable in frequency bands (Cohen, 2011b). In other words, a frequency-band dissociation between conflicts and errors would provide evidence in favor of distinct neurocognitive processes for conflict processing versus error processing. A time-frequency approach to EEG data is particularly well suited for making this distinction, because different neurocognitive processes that occur in the same brain region could produce distinct patterns of temporal-frequency dynamics, while producing the same or similar BOLD response and ERP (Cohen, 2011a). Indeed, it appears that much of the time-frequency power related to conflict and error processing is non-phase-locked (Nigbur et al., 2011), and is only weakly correlated with ERP indices of conflicts and errors (Cavanagh et al., 2012; Cohen and Donner, in press; Trujillo and Allen, 2007). Thus, the purpose of this paper was to examine EEG oscillatory dynamics related to mixed correct trials (indexing conflict without errors) and full error trials (indexing a combination of errors and

conflict). We pooled data from four different experiments, including 64 subjects. The same analyses were applied to all datasets in order to highlight commonalities in conflict and error processing across a range of cognitive control and perception tasks (Riesel et al., 2013). Our analyses focus on features of time-frequency dynamics that have been observed in previous conflict and error studies, including frequency band-specific power, phase synchronization (a measure of frequency-resolved functional connectivity), and trial-to-trial correlations between EEG dynamics and behavior dynamics. Methods Subjects Seventy-eight subjects from the University of Amsterdam community participated in these studies in exchange for course credit or 14 Euros. Each study was approved by the local ethics committee and subjects signed an informed consent document. Subjects had normal or corrected-to-normal vision and were self-reported free of neurological disorders and history of physical head trauma. EEG data collection EEG/EMG acquisition and analysis procedures was the same across all four studies. EEG data were acquired using at least 512 Hz from 64 channels placed according to the international 10–20 system, and from both earlobes. Electromyographic (EMG) recordings were taken from the flexor pollicis brevis muscle of each thumb using a pair of surface electrodes, placed on a subject-by-subject basis approximately 5 mm apart on the thenar eminence. Offline, EEG data were high-pass filtered at 0.5 Hz and epoched from −1.5 to +2 s surrounding stimulus onset of each trial. All trials were visually inspected and those containing facial EMG or other artifacts not related to blinks were manually removed. Independent components analysis was computed using eeglab software (Delorme and Makeig, 2004), and components containing blink/oculomotor artifacts or other artifacts that could be clearly distinguished from brain-driven EEG signals were subtracted from the data. All data were scalp Laplacian transformed prior to analyses (Kayser and Tenke, 2006). Scalp Laplacian is a band-pass spatial filter (effectively a highpass spatial filter for 64 electrodes) that minimizes volume conduction by removing spatially broad and therefore likely volume conducted activities. This approach has been validated for investigating inter-electrode synchronization (Srinivasan et al., 2007; Winter et al., 2007), and is an appropriate method for examining synchronization dynamics of large-scale cortical networks during error/conflict monitoring (Cohen, 2011b; Nigbur et al., 2011; van de Vijver et al., 2011). The units of the data after this transform are μV/cm2, although time-frequency power data here were converted to decibel (see below). Tasks Data from four tasks were pooled together. In Task 1 (Cohen and van Gaal, 2013), subjects performed a visual discrimination task in which they had to report whether a briefly presented target stimulus was a square or a diamond by pressing a left or right response button (counterbalanced across subjects). The target (17 ms) was followed by a metacontrast mask (200 ms, 67 ms after target offset). Auditory performance feedback was given at button press. Tones were presented on both correct and error trials, and the mapping of tone pitch to correct/error was counter-balanced across subjects. The intertrial-interval was fixed at 1017 ms. In Task 2 (Vissers et al., 2013), subjects performed a color-motion variant of a Simon task, in which a dot kinetogram with low-coherence moving blue or green dots was presented. Subjects responded to motion direction or dot color in blocks of trials, signaled by instructions. Conflict arises when, for example, blue dots that require a right-hand response are moving leftward.

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

Stimuli were onscreen until button press, and a 1000 ms inter-trial interval separated trials. In Task 3 (Cohen and Ridderinkhof, 2013), subjects performed a color-location Simon task, in which they responded according to the color of a circle (e.g., red circles require a left-hand response) while ignoring the location (left or right of fixation). Four colors were used to counter-balance conflict and stimulus repetition effects. Circles were onscreen for 34 ms, and a variable inter-trial-interval between 700 and 1200 ms separated trials. In Task 4 (Cohen, 2011a, 2011b), subjects performed an auditory/visual Simon task, in which a tone was presented to the left or the right ear, and a square was simultaneously presented on the left or right side of fixation. In different blocks, indicated by instruction cues, subjects responded to the tone or visual stimulus location while ignoring the other stimulus modality. Stimuli were onscreen/in ear for 50 ms, and a 1000–1500 ms intertrial-interval separated trials. Non-overlapping results from Task 1 (Cohen and van Gaal, 2013), 2, 3 (Cohen and Ridderinkhof, 2013), and 4 (Cohen, 2011b) have been published elsewhere; additional methodological details can be found in these papers. No results presented here have been or will be presented in other publications.

changes in EMG variance. A partial error was marked when the Z-normalized derivative of the EMG on the hand not used to make the response exceeded two standard deviations in the time between stimulus onset and the actual button press. The peak of the partial error must also have been more than two times the largest peak from −300 ms to stimulus onset; this eliminated trials in which noisy EMG produced apparent partial errors. EMG onset was then defined as the first time point in which the Z-normalized EMG derivative exceeded a Z-value of 2. Visual inspection of random subsets of trials and subjects confirmed the accuracy of the algorithm. In our experience and for our data collection setup, this is a fairly stringent set of criteria. Thus, there may be additional partial errors that went undetected because they have small amplitude or noisy baseline EMG activity. However, these very small partial error are thus included as pure correct trials; thus if anything, a conservative threshold increases the probability of Type II and not Type I errors (see the Limitations section in the Discussion for more on this point). The advantage of an algorithmic approach is that it has no biases that change over time, conditions, or subjects. Furthermore, for the very small partial errors that might be missed by the algorithm, it is not clear that a human would be able to reliably identify such very small partial errors from noise. We selected trials that fell into one of three conditions: Pure correct, mixed correct (the correct button was pressed and there was a partial error on the hand not used, and only one button was pressed), and full error (an error was made). In all cases, trials included in the analyses followed correct trials without partial errors. Some subjects had a small number of “partial correct” responses (the wrong button was pressed but the hand corresponding to the correct response twitched); these

Identification of partial errors was done via an algorithm and confirmed by visual inspection for random subsets of trials and subjects (see Fig. 1A). The derivative of the EMG signal from each hand was first Z-transformed (across the entire trial period) and rectified. This allowed us to eliminate hand- and subject-specific differences in impedance and signal amplitude, and base partial error identification on time-varying

EMG amplitude

Pure correct trial incorrect response Button press correct response

B1 Number of subjects

Identification of partial errors

A1

505

8 6 4 2 0

0

10

20

30

Percent partial errors

B2

EMG amplitude

Mixed correct trial correction time

correct response

partial error

stimulus

400

% correct trials with PE

A2

800

30 20 10 0

D

RT (ms)

Matched RT trials

600

600

500

500

400

400

PC

FE

MC

PC

500

700

FE

MC

EMG ERPs EMG signal (μV)

C All trials RTs

300

mean RT bin (ms)

Time (ms)

MC FE PC

160 120 80 40 0 −200

0

200

400

600

Time (ms) Fig. 1. Panels A1-2 show two example trials that illustrate a pure correct trial with only one EMG response (A1), and a mixed correct trial that contains a partial error (A2; subthreshold EMG response on the incorrect hand). Panel B1 shows a histogram of the average percent of correct trials with partial errors across all subjects, and panel B2 shows the distribution of partial errors as a function of RT bin (error bars indicate standard errors of the means along both X- and Y-axes). Panel C shows average reaction times (RTs) from pure correct (PC), mixed correct (MC; correct trials with partial errors), and full error (FE) trials. The bars in the left show results from all trials, and the bars in the right show trials used in the EEG data analyses (selected to match RT and trial count within-subjects). (D) Time-domain rectified (at the single-trial level before averaging) EMG trial averages for each condition, timelocked to EMG onset.

506

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

trials were identified and removed from analyses. Any trials with no button presses, with more than one button press, that followed full errors or partial errors, or with responses faster than 200 ms were removed. Of the 78 subjects across all four studies, 14 had fewer than 25 partial errors or had EMG data that were too noisy to confidently identify partial errors, and were not included in group-level analyses. Thus, the present paper includes analyses from 64 subjects. For the EEG analyses, the same number of RT-matched trials (based on button press times) for all conditions was selected. This was done to eliminate the possibility that our results were due to differences in average RT or time-on-task across conditions (Carp et al., 2011), and to help equalize signal-to-noise ratio across conditions. On the one hand, this involves potentially selecting from slightly different regions of the RT distribution in different conditions. On the other hand, differences in trial count may lead to differences in signal-to-noise ratio across conditions, which can produce spurious condition differences, particularly for phase-based analyses (Cohen, 2014). Furthermore, large condition differences in average RT may also be a confound that prevents interpretation of condition differences. Selecting trials to trial- and RT-match conditions minimizes these two confounds. After removing these trials, on average, 18.5% of total task trials (standard error 1.35%) were retained for each subject for EEG analyses (this includes partial errors, pure corrects, and full errors). Although this may seem like a low number, it includes both trial rejection and trial selection based on the number of mixed correct trials. EEG time-frequency decomposition All analyses were performed in Matlab. Single-trial data were first decomposed into their time-frequency representation by fast convolution, which involves multiplying the power spectrum of the EEG (obtained from the fast-Fourier-transform) by the power spectrum of 2 2 complex Morlet wavelets (ei2πtf e−t =ð2σ Þ , where t is time, f is frequency, which increased from 1 to 80 Hz in 40 logarithmically spaced steps, and σ defines the width of each frequency band, set according to n/(2πf), and n increases logarithmically from 4 to 12 as a function of frequency bin), and then taking the inverse fast-Fourier-transform. From the resulting complex signal, an estimate of frequency-band-specific power at each time point was defined as the squared magnitude of the result of the convolution Z (real[z(t)]2 + imag[z(t)]2), and an estimate of frequencyband-specific phase at each time point was taken as the angle of the convolution result. Power was normalized using a decibel (dB) transform (dB power = 10*log10[power/baseline]), where the baseline activity was taken as the power at each frequency band, averaged across conditions, from −300 to −100 ms pre-stimulus. Conversion to a dB scale ensures that data across all frequencies, time points, electrodes, conditions, and subjects are in the same scale and thus visually and statistically comparable.

trials), INT is the intercept, b1–2 are regression coefficients to be estimated, E is unexplained variance, and EMG_cortime is a trial vector of each subject's partial error correction time (bigger numbers mean a longer time period between partial error EMG onset and correct response EMG onset) and EMG_ratio is a trial vector of partial error vs. correct EMG energy ratio on each trial (bigger numbers mean a larger EMG energy for the partial error compared to the correct response). “Energy” of the EMG response was defined as the averaged rectified signal from onset to offset. (Note that because power values were not baselinenormalized, the intercept largely accounts for power law scaling across frequencies and therefore is not of interest here). Robust regression was used to minimize the contribution of potential outliers via iterative reweighted least squares that minimizes the impact of outliers with large leverage (O'Leary, 1990). Ultimately, this procedure resulted in a time X frequency X electrodes X condition matrix of b values for each subject. Before averaging across subjects, b values were standardized by scaling the coefficients by their standard deviations; this ensured that the coefficients were in the same scale and thus directly comparable across time, frequency, electrodes, and subjects. Furthermore, these values are normally distributed under the null hypothesis, and therefore can be subjected to parametric statistical analyses, such as t-tests. Trial-to-trial correlations with RT We examined brain-RT relationships by correlating the timefrequency power estimates over trials with the RTs over trials. This produces a time-frequency map of correlation coefficients for each subject. Correlation coefficients at each time-frequency pixel were then tested against zero at the group level. Spearman correlations were used because power is non-normally distributed. Correlation coefficients were Fisher-Z transformed prior to group-level analyses. Correlations were used instead of regression because there was only one independent variable; the use of Spearman's correlation minimizes the impact of potential outliers with large leverage. EEG statistics Statistics were performed by t-tests, and multiple comparisons were corrected using cluster-based permutation testing (Maris and Oostenveld, 2007; Nichols and Holmes, 2002), in which the assignment of condition to each data point was randomly shuffled, and statistics were re-computed. After thresholding each permutation map (p b 0.05), the number of pixels in the largest supra-threshold cluster was stored. This was repeated 1000 times, generating a distribution of maximum cluster sizes under the null hypothesis. Any clusters in the real data that were at least as large as the 95% of the distribution of null hypothesis cluster sizes were considered statistically significant. Results

Inter-site phase clustering

Behavioral results

Frequency band-specific phase synchronization is a putative measure of electrophysiological functional connectivity, and was computed n according to 1n ∑ ei½ϕjt −ϕkt  , where n is the number of trials, and ϕj and

On average, 13.4% (SD = 6.3%) of correct trials contained a partial error (Fig. 1A). However, this proportion varied considerably across subjects (Fig. 1B1). Across the 64 subjects, the number of mixed trials and full error trials were not significantly correlated (Spearman's rho = −0.096, p = 0.45). Distribution analyses show that the proportion of mixed correct trials increased with RT (Fig. 1B2) (Burle et al., 2002). When considering mean RT from each condition, mixed correct trials are slow, as others have reported (Allain et al., 2009; Coles et al., 1995; Szucs et al., 2009) (main effect of condition in repeatedmeasures ANOVA; F1,62 = 51.74, p b 0.001), and follow-up t-tests comparing condition pairs were significant (all p's b 0.05) (Fig. 1C). EMG onset times (relative to stimulus onset) showed that the earliest response activations were for partial errors, then pure correct trials, and then full errors (mean EMG onsets from stimulus onset in ms: 289,

t¼1

ϕk are the phase angles of electrodes j and k. This is an index at each time-frequency point of the consistency of phase angle differences between two electrodes over trials. Single-trial robust regression analyses A robust regression was computed that estimated parameters at each time-frequency-space point (Cohen and Cavanagh, 2011) for the following linear equation: Y = INT + b1EMG_cortime + b2EMG_ratio + E. Y is the data vector (power values at each time-frequency point across

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

372, and 430; one-factor repeated-measures ANOVA on condition differences in onset time: F1,126 = 261, p b 0.001). The average correction time was 141 ms (time difference between EMG onset of the incorrect response and EMG onset of the correct response in mixed trials, see Fig. S3). Fig. 1D displays the rectified time-domain EMG potentials, time-locked to EMG onset, which shows that the pure correct trials had the largest EMG deflection, followed by full errors and partial errors on mixed correct trials. Although not the focus of the present paper, we also investigated post-error and post-mixed correct trial slowing. We observed the typical post-error slowing effect, such that trials following full errors had longer RTs compared to those following pure correct trials and mixed correct trials (F1,62 = 20.66, p b 0.001; data not shown). Subjects were not significantly slower on trials following mixed correct compared to trials following pure correct trials (respective means: 480.9 and 473.9 ms; t63 = 1.04, p = 0.30). Time-frequency EEG power In general, the EMG-onset-locked time-frequency dynamics revealed a strong increase in midfrontal theta/delta activity and a suppression of beta-band activity (Fig. 2A). Averaged across conditions at electrode FCz

507

the time-frequency peak was 4.49 Hz and 184 ms after EMG onset (SEM: 0.13 Hz and 8.6 ms), and the beta-band suppression peaked at 20.1 Hz and 163 ms after EMG onset (SEM: 0.91 Hz and 12.1 ms). As predicted based on several previous studies (Cavanagh et al., 2009; Cohen, 2011b; Cohen and Cavanagh, 2011), we found enhanced delta/theta-band activity localized to midfrontal electrode FCz during full errors compared to pure correct responses (Fig. 2B). Note that the error-related theta increase here appears later in time compared to other studies (Trujillo and Allen, 2007). This is because EEG data are typically time-locked to the response button press, whereas here we timelocked the data to the EMG onset. The comparison between mixed correct trials and pure correct trials revealed a focused theta-band activity increase (Fig. 2B, middle panel). The direct comparison between full errors and mixed correct trials revealed that full errors had stronger delta-band activity than mixed correct trials. There was also a non-significant trend towards more theta-band power for mixed correct vs. full errors. Power results from each of the four experiments separately are shown in Fig. S2. The consistency of the results across the four experiments demonstrates that the theta dynamics accompanying mixed correct trials reflect neurocognitive mechanisms that are present beyond any possible task-specific features (see also Discussion).

A) Time-frequency power at electrode FCz Full error

Mixed correct

3

40

Power (dB)

Frequency (Hz)

Pure correct 19 9 4 2

−3 −400

0

400

800

−400

0

400

800

−400

0

400

800

Time (ms)

B) Condition differences at electrode FCz MC−PC

FE−MC

40

3

19 9 4 2

−3 −400

0

400

800

−400

0

400

800

−400

0

400

Power (dB)

Frequency (Hz)

FE−PC

800

Time (ms)

C) Trial-to-trial multiple regression at electrode FCz EMG energy ratio 1.5 40 19

beta

Frequency (Hz)

Correction time

9 4 2

−1.5 −400

0

400

800

−400

0

400

800

Time (ms) Fig. 2. Time-frequency power plots from midfrontal electrode FCz for each condition. Black contours outline time-frequency regions significant at p b 0.05, corrected for multiple comparisons using cluster-based permutation testing. (A) Time-frequency power dynamics relative to the pre-stimulus baseline period, plotted separately for each condition. (B) Condition differences; topographical maps show scalp distribution of the theta-band effects. Topographical maps were made by averaging power in windows of 0–600 ms and 4–9 Hz for the theta-band (lefthand and middle plots), and 0–800 ms and 1.5–3 Hz for the delta-band (right-hand plot). All time-frequency plots have the same color scale (+/−3), and the color scale range for topographical plots is +/−2 for the FE-PC and MC-PC results, and +/−1 for the FE-MC result. (C) Results from single-trial regression analyses show a negative relationship between midfrontal theta and the time lag between partial error and correct EMG onsets (“correction time”), such that the closer the partial error and correct EMG onsets, the greater the post-EMG onset theta power. The ratio of the two EMG responses did not correlate with any EEG dynamics (right plot).

508

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

correlations reported in our previous studies (e.g., Cohen and Cavanagh, 2011).

Trial-to-trial EEG-EMG regression To examine more closely the relationship between muscle response dynamics and neural activity, we performed brain-behavior regression analyses. Based on intuition and a dominant model of conflict monitoring (Yeung et al., 2004), we predicted that when there is more overlap between the partial error and the correct response, the subject experienced more conflict, and this would be reflected in stronger frontal theta power (thus, we predicted a negative correlation between correction time or EMG energy ratio and theta power). In fact, the opposite result was reported several years ago using binned ERPs (Burle et al., 2008), which we discuss more in a later section. The regression model predicted, at each time-frequency-electrode point, the extent to which EEG timefrequency power predicted (1) the correction time, and (2) the ratio of the signal energies of the two EMG responses (see Methods). We included both the correction time and the signal energy ratio because it was not clear a priori which would be a better manifestation of conflict. Note that because single-trial regression utilizes variance from continuous variables, it has higher sensitivity for uncovering brain-behavioral relationships than discretizing and binning continuous data. We found highly significant negative regression coefficients selectively in the theta band, maximal over FCz, from the partial error EMG onset until about 900 ms thereafter (Fig. 2C). These negative coefficients indicate that there was relatively stronger midfrontal theta power on trials in which the partial error and correct response EMG onsets were closer in time to each other (thus, shorter correction times). No significant effects were observed for the EMG amplitude ratio coefficients, indicating that midfrontal cortical electrical activity is sensitive to the timing of initiated motor commands rather than to the relative strength of the muscular response. Although we focus on the mid-frontal theta effect because of a priori hypotheses, we note that there was also a statistically significant positive effect in the betaband over sensory-motor electrodes (data not shown).

Frontal–partial synchronization-based networks In addition to theta power, errors have also been linked to increases in oscillatory inter-site phase clustering (a measure of electrophysiological functional connectivity; Lachaux et al., 1999; Womelsdorf et al., 2007) between midfrontal and lateral frontal sites (Cavanagh et al., 2009; Hanslmayr et al., 2008). Based on previous literature and the power results, we used FCz as a “seed” electrode, and examined ISPC between FCz and all other electrodes in the time-frequency window that was based on the results of the power analyses (full error vs. mixed correct; see Fig. 2B): 1.5–3 Hz and 0–800 ms. Note that data were scalpLaplacian transformed, which attenuates volume conduction and therefore renders the data more appropriate for inter-regional synchronization analyses (see Methods). As in previous studies (Cavanagh et al., 2009; Cohen and Cavanagh, 2011; Nigbur et al., 2011), ISPC between FCz and a network of frontalparietal electrodes increased after full errors compared to pure correct and mixed correct trials in the delta band (Fig. 4A). We then selected the significant frontal/parietal electrodes for subsequent analyses. The time-frequency ISPC maps in Fig. 4B confirm the significantly stronger delta-band ISPC for full errors compared to pure correct and mixed correct trials. In this frontal–parietal network, mixed correct trials also exhibited stronger theta-band ISPC compared to pure correct trials, even though the electrodes were selected based on delta-band effects (Fig. 4B2, middle panel, and Fig. 4C, right panel). These findings demonstrate that the delta-band network is preferentially related to error processing, whereas the theta-band network is preferentially related to conflict processing. We next repeated this analysis procedure for the theta band. Again, we selected a time-frequency window based on the power results from MC-PC in Fig. 2B (electrode FCz, 0–400 ms, 4–9 Hz). ISPC was enhanced for mixed correct trials compared to pure correct trials selectively between FCz and lateral prefrontal electrodes (Fig. 4D, middle panel). Comparing full errors to pure correct trials revealed additional ISPC differences in posterior parietal electrodes. Time-frequency plots of ISPC in the frontal–parietal network (Fig. 4E1) revealed stronger theta-band ISPC for mixed correct compared to both pure correct and full error trials (Fig. 4E2).

Trial-to-trial EEG-RT correlations We next linked the EEG data to conflict processing by means of within-subject cross-trial correlations with RT. Here the idea is that if conflict-related midfrontal activity reflects neural mechanisms of detecting and resolving conflict, then trial-to-trial fluctuations in midfrontal activity should predict trial-to-trial fluctuations in task performance as measured by RT. In these analyses, a positive correlation at a given time-frequency point indicates that trials with longer RTs generally have more power at that time-frequency point. As seen in Fig. 3, no statistically significant correlations were observed in pure correct trials. For full error trials, we observed a pre-EMG-onset significant positive correlation in the theta band, and a post-EMG-onset significant negative correlation in the delta/theta band from around 0 to 600 ms. During mixed correct trials, significant positive RT-power correlations were observed selectively in the theta band from around 200–800 ms post-EMG-onset, which is consistent with peri-response

Event-related potentials (ERPs) In our final set of analyses, we examined the ERPs—the time-domain EMG onset-locked EEG potential. This was done mainly to replicate previous findings concerning the relationship between the ERP and partial errors. As expected, we observed a typical error-related negativity (ERN) (van Veen and Carter, 2006) and a subsequent error-related positivity (Pe) (Falkenstein et al., 1991; Nieuwenhuis et al., 2001; Overbeek et al., 2005) following full errors (see Figs. 5A–B for ERPs, topographical

TF plots of RT-power correlations Full error

Mixed correct .15

19 9 4 2

−.15 −400

0

400

800

−400

0

400

800

−400

0

400

Correl. coeff.

Frequency (Hz)

Pure correct 40

800

Time (ms) Fig. 3. Results of within-subject cross-trial EEG-RT correlation analyses. The color at each time-frequency pixel indicates the Spearman correlation coefficient (Fisher-Z transformed) between cross-trial RTs and cross-trial time-frequency power (computed within-subjects, and then the coefficients are averaged across subjects). Black contours outline regions that are statistically significant at p b 0.05, corrected for multiple comparisons using cluster-based permutation testing. Color scaling is the same for all time-frequency plots and topographical maps.

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

FE-PC

MC-PC

FE-MC

Pure correct

Pure error

Mixed correct

40 19 9 4 2

65

−65 −400

0

400 800 −400

0

400 800 −400

0

Percent change

(All topoplots: 1.5-3 Hz, 0-800 ms)

B1) TF plots of connectivity: main effects Frequency (Hz)

A) FCz-seeded connectivity in delta

509

400 800

Time (ms)

80

60

60

40

40

20

20

0

0

PC FE MC

B2) TF plots of connectivity: difference maps Theta-band

PC FE MC

Frequency (Hz)

80

Delta-band

FE-PC

MC-PC

FE-MC

0

0

0

40 19 9 4 2

65

−65 −400

400 800 −400

400 800 −400

Percent change

Percent change

C) Barplots of connectivity

400 800

Time (ms)

FE-PC

MC-PC

FE-MC

Pure correct

Pure error

Mixed correct

40 19 9 4 2

65

−65 −400

0

400 800 −400

0

400 800 −400

0

Percent change

(All topoplots: 4-9 Hz, 0-400 ms)

E1) TF plots of connectivity: main effects Frequency (Hz)

D) FCz-seeded connectivity in theta

400 800

Time (ms)

80

60

60

40

40

20

20

0

PC FE MC

0

E2) TF plots of connectivity: difference maps Theta-band

PC FE MC

Frequency (Hz)

80

Delta-band

FE-PC

MC-PC

FE-MC

0

0

0

40 19 9 4 2

65

−65 −400

400 800 −400

400 800 −400

Percent change

Percent change

F) Barplots of connectivity

400 800

Time (ms) Fig. 4. Inter-site-phase-clustering (ISPC, a measure of frequency band-specific functional connectivity) seeded from midfrontal electrode FCz shows network-level dynamics of error and conflict processing. (A) Topographical maps of ISPC condition differences in the delta band. Black dots indicate electrodes with significant differences at p b 0.001; green dots at p b 0.01. (B) Time-frequency plots of synchronization between FCz and a group of significant frontal and parietal electrodes (FPz, AF7, AF3, F3, FC4, P6, PO4, PO3, P3) for individual conditions (B1) and for condition differences (B2) (see significant electrodes in panel A, middle and right-hand panels). (C) Bar plots of connectivity further highlight the relationship between frequency band-specific connectivity and conflict vs. error processing. (D–F) Same as panels A–C, except that connectivity was defined by a theta-band window from the power analyses. Electrodes used in the time-frequency plots are P7, P5, P8, O2, AF8, F5 (see significant electrodes in panel D, middle and right-hand panels).

maps, and additional statistics; note that EEG data were scalp-Laplacian transformed). Partial error-locked data showed a narrow and largeamplitude ERN followed by a brief positivity that does not resemble a typical Pe (the Pe tends to be longer-lasting; Overbeek et al., 2005), but does have similar temporal characteristics as in previously reported studies (Burle et al., 2008; Masaki et al., 2012; Seymour and Schumacher, 2009). Although the ERN was larger after partial errors compared to full errors, the literature as a whole does not consistently show larger or smaller ERNs on partial errors compared to full errors (Burle et al., 2008; Coles et al., 1985; Masaki et al., 2012; Roger et al., 2010; Vidal et al., 2000, 2003). The condition-averaged negative ERP peak ocurred at 90 ms, and the condition-averaged subsequent positive ERP peak occurred at 218 ms. We next performed several analyses to examine the relationship between the ERP and conflict processing. The first analysis was similar to the within-subjects regression analysis shown in Fig. 2C; the only difference was that here we used the time-domain signal (that is, the single-trial ERP) instead of time-frequency power. Consistent with the time-frequency results, we found a significant (p b 0.001) relationship between correction time and ERP amplitude (peak time at 160 ms

after partial error EMG onset; note that this is 70 ms after the peak of the ERN), and no significant relationship between EMG energy ratio. The significant beta coefficient is positive, indicating that longer correction times are associated with larger single-trial EEG amplitude (at the single-trial level, this could mean larger positive amplitude or less negative amplitude). In our next set of analyses, we attempted to replicate a finding reported by Burle et al., 2008. The ERP in our dataset with 64 subjects resembled the overall waveform characteristics of that reported in Burle et al., 2008 (compare the timing and amplitude of the negative/positive ERP peaks in our Fig. 5D with their Figs. 2 and 4). Using a similar analysis approach as taken by Burle et al., we observed no consistent linear or monotonic relationship between the EMG onset to button press latency and the ERP peak timing or magnitude for the early negative peak (the ERN) (Fig. 5D). Although there was a significant main effect of EMG-to-RT bin (F1,189 = 5.26, p = 0.002) and a marginally significant linear effect (F1,63 = 3.1, p = 0.083), this was not driven by a monotonic linear increase in ERN magnitude as a function of bin. For example, a t-test between the ERN amplitudes on the first and fourth bins was not significant (t63 = 0.62, p = 0.53) (see bar plots in Fig. 5D, left-hand

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

A) ERPs from electrode FCz

D) MC ERPs in EMG-to-button-press time bins

Mixed correct (MC) Full error (FE) Pure correct (PC) −.1

-0.4

0 0.2 0.4 100

200

300

400

500

Time (ms)

0

Significant at p<0.01

.1 MC-PC

FE-PC FE−MC

−400

0

400

800

Time (ms)

0

Average of 50-150 ms

Average of 150-300 ms 0.2

-0.1 -0.2

0.1

-0.3 -0.4

1

2

3

0

4

1

Bin

Mixed correct

Full error

Pure correct

C) Regression between FCz and EMG

2

3

4

Bin

E) MC ERPs in EMG-to-EMG time bins Voltage (μV/cm2)

B) Topographical maps (70-110 ms)

-0.4

bin 3 bin 4

bin 1 bin 2

-0.2 0 0.2 0.4 0

100

200

300

400

500

Time (ms)

1

Correction time EMG energy ratio

0 p<.001

−400

0

400

800

Time (ms)

Voltage (μV/cm2)

Regression β coefficient

bin 3 bin 4

bin 1 bin 2

-0.2

0

Voltage (μV/cm2)

EEG amplitude (µV/cm2)

−.2

Voltage (μV/cm2)

510

0

Average of 50-150 ms

0.4

Average of 150-300 ms

0.3 -0.1

0.2 0.1

-0.2 1

2

3

Bin

4

0

1

2

3

4

Bin

Fig. 5. Event-related potentials (scalp-Laplacian transformed, see Methods), time-locked to EMG onset. (A) All ERPs from electrode FCz, and (B) topographical maps from each condition. (C) Results from a within-subject regression analysis, which was similar to the analysis shown in Fig. 2C but using the time-domain EEG instead of the time-frequency data. (D) The attempt to replicate a linear increase in ERN amplitude (the negative peak at around 80 ms) and the time lag between partial error EMG onset and correct button press. The X-axis limits are different from our other plots to facilitate direct comparison with the figures in Burle et al. (2008). (E) Results from a similar analysis, but based on binning the data according to EMG-toEMG onsets (rather than EMG onset to button press).

panel). The subsequent positivity (which was not examined in Burle et al., 2008) also showed a main effect of bin (F1,189 = 10.48, p b 0.001) and a linear effect (F1,63 = 24.13, p b 0.001), although this linear effect was also not driven by a monotonic change as a function of bin (see bar plots in Fig. 5D, right-hand panel). Finally, we tried several alternative ways of binning the data, such as basing the bins on the correction time (that is, EMG to EMG time lag rather than on the EMG to button-press time lag) (Fig. 5E), and using each subject's RT quartiles rather than specified time lags (data not shown). In none of these approaches did we observe a clear increase in ERN amplitude with increasing bin (the smallest p-value for a linear effect was p = 0.105). Discussion Theta-band vs. delta-band dynamics for conflict vs. error processing We found that subthreshold muscle twitches of the incorrect response hand (partial errors) are relatively frequent during correct trials in cognitive tasks, and that these partial errors elicited strong midfrontal theta dynamics, even though the overt response was correct. Local power dynamics as well as large-scale fronto-parietal networks grouped by delta phase synchronization (ISPC) further dissociated full errors from partial errors: Full errors had increased delta-band

dynamics (in terms of power, ISPC, and cross-trial correlations with RTs) compared to mixed correct trials, while mixed correct trials had increased theta-band dynamics (in terms of ISPC and correlations with RTs) compared to pure correct trials. These findings help clarify the functional roles of theta- and delta-band local and large-scale dynamics during conflict processing vs. error detection, and suggest that errors are uniquely associated with enhanced delta-band activity (Yordanova et al., 2004) whereas response conflict (as measured on both mixed correct and full error trials) are associated with enhanced theta-band activity. That both theta-band and delta-band dynamics are often coobserved in other studies may be related to mixing between conflict and errors: Errors generally occur when there is conflict, and errors themselves may incite conflict. Indeed, one may speculatively interpret the middle panel of Fig. 3 to indicate that conflict is first experienced (reflected as a positive theta-RT correlation), and then error processing occurs (reflected as a negative delta-RT correlation). These findings contribute to a growing literature highlighting the relevance of frontal theta-band dynamics for cognitive control and adapting behavior after response conflict and negative performance feedback (Cavanagh et al., 2009; Cohen, 2011b; van de Vijver et al., 2011). The present findings also add an important and novel piece of the puzzle: Theta-band activity localized to medial frontal electrodes reflects mainly conflict, whereas delta-band activity seems to be related

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

to additional error-related processing. Recent work suggests that midfrontal regions are a hub in a larger network (Cohen, 2011b) that utilizes delta/theta-band synchronization to coordinate controlled processing after conflicts and errors. The present findings add to this literature by showing a similar dissociation between theta-band synchronization during conflict (mixed correct and full error trials) versus delta-band synchronization only during “pure” error processing (full error trials only), suggesting that partial errors and full errors may recruit at least partly distinct neural mechanisms. This is further supported by behavioral measures: Across subjects, there was no significant correlation between the proportion of partial errors and full errors. Furthermore, full errors elicited significant post-error slowing, whereas partial errors did not, although others have shown that there is some behavioral slowing after both partial errors and full errors (Allain et al., 2009). We did not measure subjective awareness of partial errors; further research would be required to determine the relationship between consciousness and partial errors. Within-subject, cross-trial analyses clarify brain-conflict relationships The cross-trial robust regression and correlation analyses further highlighted the link between MFC theta activity and partial error dynamics in several ways. First, regression analyses revealed that the closer in time the partial error was to the correct response (the correction time), the larger was the frontal theta power increase (Fig. 2C). No effects were observed for the ratio of the partial error and correct response EMG activities, suggesting that midfrontal theta is sensitive to the timing of response activations, but not to the strength of the response activations. Interestingly, the average correction time was also in the theta band (see Figure S3). Second, correlation analyses revealed selective theta-band correlations between theta power and RT during mixed correct trials, but not during pure correct trials. This replicates previously reported findings that showed similar patterns of theta-RT correlations during incongruent compared to congruent trials (Cohen and Cavanagh, 2011; Cohen and Nigbur, 2013). However, this seems to contrast with a conflict model that predicts that RT would index conflict even on correct trials (Yeung et al., 2011). Assuming that midfrontal theta can be interpreted as an index of response conflict (Cavanagh et al., 2009; Cohen, 2011b; Cohen and Cavanagh, 2011; Hanslmayr et al., 2008; Trujillo and Allen, 2007), this finding indicates that long RTs on pure correct trials may be driven by factors other than conflict per se, such as attention lapses (van Driel et al., 2012; Weissman et al., 2006). Note that a better test of response overlap might be the temporal overlap between the partial error and the correct response EMG activations. However, very few trials contained such overlap, as illustrated in Fig. 1A. Third, the correlation analyses revealed a negative relationship between midfrontal delta power and RT during pure error trials (Fig. 3, middle panel), which is a frequency band qualitative dissociation from the brain-behavior relationship observed in the theta band. This finding indicates that error trials with faster RTs contain more delta power. To our knowledge, this has not been reported previously. Note that this result is inconsistent with recent suggestions that increased midfrontal activity during conflict tasks reflects only “time-on-task” (Cohen and Nigbur, 2013; Scherbaum and Dshemuchadse, 2013). Event-related potentials Some of the ERP findings shown in Fig. 5 (e.g., panel A) are replications of previous findings. However, we did not replicate a relationship between the ERN amplitude and the time lag between partial error EMG onset and button press (Burle et al., 2008), although the waveform shape and timing of the ERPs in our dataset was similar to that of Burle et al. This apparent discrepancy between the ERPs and the timefrequency responses in terms of their correlations with trial-varying behavior is not surprising, considering our ongoing work demonstrating

511

that most of the conflict-related signal in the EEG is non-phaselocked, and that ERP signatures of conflict (such as the ERN and N2) are statistically poor predictors of both single-trial and trial-averaged conflict task performance relative to the non-phase-locked theta power (Cohen and Donner, in press). The N2, for example, has relatively low statistical power and therefore requires many subjects to obtain reasonable statistical power (N100, based on the effect sizes in Cohen and Donner, in press). Along a similar vein, a null result was previously reported between (time-domain) ERN magnitudes according to RT median splits (Falkenstein et al., 2000), which contrasts with our statistically robust finding of a continuous relationship between RT and delta power during error trials. Both the increase in signal-to-noise ratio that is observed in time-frequency power compared to ERPs, and the very large number of subjects included in the present analyses, increase statistical confidence in our results (including the null results). The only single-trial ERP-behavior analysis that revealed a significant effect was the brief correlation between ERP amplitude and correction time (Fig. 5C). The peak of this correlation, and the time during which it was significant, occurred after the peak of the ERN (90 ms vs. 160 ms; compare the timing of Figs. 5A and 5C), suggesting that this correlation may have a limited or unclear functional relationship with the ERN itself. Inspection of the discretization analyses in Fig. 5E (the discretization variable–EMG-to-EMG onset times–is the same as the correlating variable in Fig. 5C) suggests that trials with longer correction times may have slower-decaying or later-occurring potentials, rather than the peak of the ERN being modulated. Although one may be tempted to conceptualize the theta response as the frequency-domain version of the ERN or N2, growing evidence suggests that this is not the case. As examples: ERP and theta-band signatures of conflict and error processing are often weakly or non-significantly correlated (Cavanagh et al., 2012), most of the conflict modulation of theta activity is in the non-phase-locked part of the signal (Nigbur et al., 2011), and it seems that ERP and theta-power measures of conflict are statistically independent of each other (Cohen and Donner, in press). Limitations Here we pooled data across four different experiments. Different tasks might elicit different kinds of errors (indeed, even different variants of the same task may elicit different kinds of errors; Scheffers and Coles, 2000; van Driel et al., 2012). Thus, it is possible that the different tasks pooled here involved somewhat different error monitoring processes. For example, some experiments may involve more perceptual errors whereas other experiments may involve more motor-based errors or attention-based errors. Some recent studies show that errors that are due to perceptual failures trigger smaller error-related neural activity than motor errors, and reduced error awareness typically decreases ERN amplitude (Charles et al., 2013; Cohen et al., 2009; Pavone et al., 2009). Nonetheless, errors in many different kinds of tasks elicit similar patterns of midfrontal theta dynamics (e.g., Cavanagh et al., 2012), and several dominant theories of error processing and action monitoring propose that the anterior cingulate and surrounding medial frontal cortex houses a task-general monitoring mechanism (Ridderinkhof et al., 2004a; Ridderinkhof et al., 2004b). Furthermore, a recent study showed high reliability of the ERN within and across tasks (Riesel et al., 2013). Indeed, Figure S2 highlights that the different tasks had similar patterns of time-frequency power dynamics with similar timing. However, in addition to the general mechanisms that are likely to be similar across tasks that we focused on here, there may be additional subtle partial error dynamics that are task-dependent. Those task-specific error/conflict dynamics were not the focus of this paper. We used an automated algorithm to identify partial errors. Although this may have resulted in failing to identify small or unusually shaped partial errors, this is unlikely to have resulted in any significant bias in the results, for the following reason. If our results were biased because our partial error-detecting algorithm selectively missed small partial

512

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513

errors, then there should have been a significant relationship between partial error EMG amplitude and ERN amplitude/theta power, in the trials that we did identify. In fact, this analysis (Figs. 5C and 2C) revealed no significant results, suggesting that even if we had systematically missed small partial errors, this did not bias the results in a meaningful way. Implications for cognitive tasks involving correct vs. error trials Partial errors are undetectable without EMG or force grips. Although partial errors tended to occur in trials with slow RTs, this was not exclusively the case, and partial errors occurred even in the fastest RT bin of trials (Fig. 1B2). Thus, removing trials with slow RTs is not a successful strategy for eliminating partial errors. The present findings thus have the general implication that partial errors appear to be significant neurocognitive events that are qualitatively distinct from both pure correct responses and full errors. Partial errors are likely present in a wide variety of cognitive tasks, but without recording EMG or using force-grips, partial errors cannot be identified and are included with pure correct trials. Thus, identifying and removing mixed correct trials will lead to cleaner estimates of brain activity underlying successful and erroneous task performance in all cognitive tasks. Acknowledgments These studies were funded by a VIDI award from the Netherlands Organization for Scientific Research (NWO) to MXC and a Veni grant from NWO to SVG. We thank Marlies Vissers, Irene van de Vijver, and Claudia Arena for assistance with data collection. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.neuroimage.2013.10.033. References Allain, S., Burle, B., Hasbroucq, T., Vidal, F., 2009. Sequential adjustments before and after partial errors. Psychon. Bull. Rev. 16, 356–362. Burle, B., Possamai, C.A., Vidal, F., Bonnet, M., Hasbroucq, T., 2002. Executive control in the Simon effect: an electromyographic and distributional analysis. Psychol. Res. 66, 324–336. Burle, B., Roger, C., Allain, S., Vidal, F., Hasbroucq, T., 2008. Error negativity does not reflect conflict: a reappraisal of conflict monitoring and anterior cingulate cortex activity. J. Cogn. Neurosci. 20, 1637–1655. Carbonnell, L., Falkenstein, M., 2006. Does the error negativity reflect the degree of response conflict? Brain Res. 1095, 124–130. Carp, J., Kim, K., Taylor, S.F., Fitzgerald, K.D., Weissman, D.H., 2011. Conditional differences in mean reaction time explain effects of response congruency, but not accuracy, on posterior medial frontal cortex activity. Front. Hum. Neurosci. 4, 231. 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. J. Neurosci. 29, 98–105. Cavanagh, J.F., Zambrano-Vazquez, L., Allen, J.J., 2012. Theta lingua franca: a common midfrontal substrate for action monitoring processes. Psychophysiology 49, 220–238. Charles, L., Van Opstal, F., Marti, S., Dehaene, S., 2013. Distinct brain mechanisms for conscious versus subliminal error detection. Neuroimage 73, 80–94. Cohen, M.X, 2011a. It's about time. Front. Hum. Neurosci. 5, 2. Cohen, M.X, 2011b. Error-related medial frontal theta activity predicts cingulate-related structural connectivity. Neuroimage 55, 1373–1383. Cohen, M.X, 2014. Analyzing Neural Time Series Data: Theory and Practice. MIT Press, Cambridge. Cohen, M.X, Cavanagh, J.F., 2011. Single-trial regression elucidates the role of prefrontal theta oscillations in response conflict. Front. Psychol. 2, 30. Cohen, M.X, Donner, T.H., 2013. Midfrontal conflict-related theta-band power reflects neural oscillations that predict behavior. J. Neurophysiol. (in press). Cohen, M.X., Nigbur, R., 2013. Reply to “Higher response time increases theta energy, conflict increases response time”. Clin. Neurophysiol. 124, 1479–1481. Cohen, M.X, Ridderinkhof, K.R.R., 2013. EEG source reconstruction reveals frontal-parietal dynamics of spatial conflict processing. PLoS One 8 (2), e57293. Cohen, M.X., van Gaal, S., 2013. Dynamic interactions between large-scale brain networks predict behavioral adaptation after perceptual errors. Cereb. Cortex 23 (5), 1061–1072.

Cohen, M.X, van Gaal, S., Ridderinkhof, K.R., Lamme, V.A., 2009. Unconscious errors enhance prefrontal–occipital oscillatory synchrony. Front. Hum. Neurosci. 3, 54. Coles, M.G., Gratton, G., Bashore, T.R., Eriksen, C.W., Donchin, E., 1985. A psychophysiological investigation of the continuous flow model of human information processing. J. Exp. Psychol. Hum. Percept. Perform. 11, 529–553. Coles, M.G., Scheffers, M.K., Fournier, L., 1995. Where did you go wrong? Errors, partial errors, and the nature of human information processing. Acta Psychol. (Amst) 90, 129–144. Coles, M.G., Scheffers, M.K., Holroyd, C.B., 2001. Why is there an ERN/Ne on correct trials? Response representations, stimulus-related components, and the theory of errorprocessing. Biol. Psychol. 56, 173–189. Delorme, A., Makeig, S., 2004. EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 134, 9–21. Egner, T., 2007. Congruency sequence effects and cognitive control. Cogn. Affect. Behav. Neurosci. 7, 380–390. Endrass, T., Klawohn, J., Schuster, F., Kathmann, N., 2008. Overactive performance monitoring in obsessive–compulsive disorder: ERP evidence from correct and erroneous reactions. Neuropsychologia 46, 1877–1887. Esterman, M., Noonan, S.K., Rosenberg, M., Degutis, J., 2012. In the zone or zoning out? Tracking behavioral and neural fluctuations during sustained attention. Cereb. Cortex 23 (11), 2712–2723. Falkenstein, M., Hohnsbein, J., Hoormann, J., Blanke, L., 1991. Effects of crossmodal divided attention on late ERP components. II. Error processing in choice reaction tasks. Electroencephalogr. Clin. Neurophysiol. 78, 447–455. Falkenstein, M., Hoormann, J., Christ, S., Hohnsbein, J., 2000. ERP components on reaction errors and their functional significance: a tutorial. Biol. Psychol. 51, 87–107. Gratton, G., Coles, M.G., Sirevaag, E.J., Eriksen, C.W., Donchin, E., 1988. Pre- and poststimulus activation of response channels: a psychophysiological analysis. J. Exp. Psychol. Hum. Percept. Perform. 14, 331–344. Hanslmayr, S., Pastotter, B., Bauml, K.H., Gruber, S., Wimber, M., Klimesch, W., 2008. The electrophysiological dynamics of interference during the Stroop task. J. Cogn. Neurosci. 20, 215–225. Kayser, J., Tenke, C.E., 2006. Principal components analysis of Laplacian waveforms as a generic method for identifying ERP generator patterns: I. Evaluation with auditory oddball tasks. Clin. Neurophysiol. 117, 348–368. Lachaux, J.P., Rodriguez, E., Martinerie, J., Varela, F.J., 1999. Measuring phase synchrony in brain signals. Hum. Brain Mapp. 8, 194–208. Maris, E., Oostenveld, R., 2007. Nonparametric statistical testing of EEG- and MEG-data. J. Neurosci. Methods 164, 177–190. Masaki, H., Murphy, T.I., Desjardins, J.A., Segalowitz, S.J., 2012. The error-related negativity associated with different strength of stimulus–response interference. Clin. Neurophysiol. 123, 689–699. Mathalon, D.H., Whitfield, S.L., Ford, J.M., 2003. Anatomy of an error: ERP and fMRI. Biol. Psychol. 64, 119–141. Nachev, P., 2006. Cognition and medial frontal cortex in health and disease. Curr. Opin. Neurol. 19, 586–592. Nee, D.E., Kastner, S., Brown, J.W., 2011. Functional heterogeneity of conflict, error, taskswitching, and unexpectedness effects within medial prefrontal cortex. Neuroimage 54, 528–540. Nichols, T.E., Holmes, A.P., 2002. Nonparametric permutation tests for functional neuroimaging: a primer with examples. Hum. Brain Mapp. 15, 1–25. Nieuwenhuis, S., Ridderinkhof, K.R., Blom, J., Band, G.P., Kok, A., 2001. Error-related brain potentials are differentially related to awareness of response errors: evidence from an antisaccade task. Psychophysiology 38, 752–760. Nieuwenhuis, S., Yeung, N., van den Wildenberg, W., Ridderinkhof, K.R., 2003. Electrophysiological correlates of anterior cingulate function in a go/no-go task: effects of response conflict and trial type frequency. Cogn. Affect. Behav. Neurosci. 3, 17–26. Nigbur, R., Cohen, M.X, Ridderinkhof, K.R., Sturmer, B., 2011. Theta dynamics reveal domain-specific control over stimulus and response conflict. J. Cogn. Neurosci. 24 (5), 1264–1274. O'Leary, D.P., 1990. Robust regression computation using iteratively reweighted least squares. J. Matrix Anal. Appl. 11, 466–480. Overbeek, T.J.M., Nieuwenhuis, S., Ridderinkhof, K.R., 2005. Dissociable components of error processing: on the functional significance of the Pe vis-à-vis the ERN/Ne. J. Psychophysiol. 19, 319–329. Pavone, E.F., Marzi, C.A., Girelli, M., 2009. Does subliminal visual perception have an errormonitoring system? Eur. J. Neurosci. 30, 1424–1431. Ridderinkhof, K.R., Ullsperger, M., Crone, E.A., Nieuwenhuis, S., 2004a. The role of the medial frontal cortex in cognitive control. Science 306, 443–447. Ridderinkhof, K.R., van den Wildenberg, W.P., Segalowitz, S.J., Carter, C.S., 2004b. Neurocognitive mechanisms of cognitive control: the role of prefrontal cortex in action selection, response inhibition, performance monitoring, and reward-based learning. Brain Cogn. 56, 129–140. Riesel, A., Weinberg, A., Endrass, T., Meyer, A., Hajcak, G., 2013. The ERN is the ERN is the ERN? Convergent validity of error-related brain activity across different tasks. Biol. Psychol. 93, 377–385. Roger, C., Benar, C.G., Vidal, F., Hasbroucq, T., Burle, B., 2010. Rostral Cingulate Zone and correct response monitoring: ICA and source localization evidences for the unicity of correct- and error-negativities. Neuroimage 51, 391–403. Scheffers, M.K., Coles, M.G., 2000. Performance monitoring in a confusing world: errorrelated brain activity, judgments of response accuracy, and types of errors. J. Exp. Psychol. Hum. Percept. Perform. 26, 141–151. Scherbaum, S., Dshemuchadse, M., 2013. Higher response time increases theta energy, conflict increases response time. Clin. Neurophysiol. 124, 1477–1479. Seymour, T.L., Schumacher, E.H., 2009. Electromyographic evidence for response conflict in the exclude recognition task. Cogn. Affect. Behav. Neurosci. 9, 71–82.

M.X Cohen, S. van Gaal / NeuroImage 86 (2014) 503–513 Srinivasan, R., Winter, W.R., Ding, J., Nunez, P.L., 2007. EEG and MEG coherence: measures of functional connectivity at distinct spatial scales of neocortical dynamics. J. Neurosci. Methods 166, 41–52. Swick, D., Turken, A.U., 2002. Dissociation between conflict detection and error monitoring in the human anterior cingulate cortex. Proc. Natl. Acad. Sci. U. S. A. 99, 16354–16359. Szucs, D., Soltesz, F., White, S., 2009. Motor conflict in Stroop tasks: direct evidence from single-trial electro-myography and electro-encephalography. Neuroimage 47, 1960–1973. Trujillo, L.T., Allen, J.J., 2007. Theta EEG dynamics of the error-related negativity. Clin. Neurophysiol. 118, 645–668. Ullsperger, M., von Cramon, D.Y., 2001. Subprocesses of performance monitoring: a dissociation of error processing and response competition revealed by event-related fMRI and ERPs. Neuroimage 14, 1387–1401. van de Vijver, I., Ridderinkhof, K.R., Cohen, M.X, 2011. Frontal oscillatory dynamics predict feedback learning and action adjustment. J. Cogn. Neurosci. 23, 4106–4121. van Driel, J., Ridderinkhof, K.R., Cohen, M.X, 2012. Not all errors are alike: theta and alpha EEG dynamics relate to differences in error-processing dynamics. J. Neurosci. 32, 16795–16806. van Veen, V., Carter, C.S., 2006. Error detection, correction, and prevention in the brain: a brief review of data and theories. Clin. EEG Neurosci. 37, 330–335.

513

Vidal, F., Hasbroucq, T., Grapperon, J., Bonnet, M., 2000. Is the ‘error negativity’ specific to errors? Biol. Psychol. 51, 109–128. Vidal, F., Burle, B., Bonnet, M., Grapperon, J., Hasbroucq, T., 2003. Error negativity on correct trials: a reexamination of available data. Biol. Psychol. 64, 265–282. Vissers, M., Slagter, H.A., van den Brink, R., Cohen, M.X, 2013. Networks Involved in Conflict Modulation of Visual Processing: A Combined EEG-DTI Study. Cognitive Neuroscience Society, San Francisco. Weissman, D.H., Roberts, K.C., Visscher, K.M., Woldorff, M.G., 2006. The neural bases of momentary lapses in attention. Nat. Neurosci. 9, 971–978. Winter, W.R., Nunez, P.L., Ding, J., Srinivasan, R., 2007. Comparison of the effect of volume conduction on EEG coherence with the effect of field spread on MEG coherence. Stat. Med. 26, 3946–3957. Womelsdorf, T., Schoffelen, J.M., Oostenveld, R., Singer, W., Desimone, R., Engel, A.K., Fries, P., 2007. Modulation of neuronal interactions through neuronal synchronization. Science 316, 1609–1612. Yeung, N., Botvinick, M.M., Cohen, J.D., 2004. The neural basis of error detection: conflict monitoring and the error-related negativity. Psychol. Rev. 111, 931–959. Yeung, N., Cohen, J.D., Botvinick, M.M., 2011. Errors of interpretation and modeling: a reply to Grinband et al. Neuroimage 57, 316–319. Yordanova, J., Falkenstein, M., Hohnsbein, J., Kolev, V., 2004. Parallel systems of error processing in the brain. Neuroimage 22, 590–602.