We thank Sabrina Döring and Philipp Euskirchen for assistance with data collection and analysis, Bertram Huth for graphical assistance, and Jan Derrfuss, Joe King, and three anonymous reviewers for helpful comments on earlier versions of this article.
Modulation of the error-related negativity by response conflict
Article first published online: 1 JUL 2009
Copyright © 2009 Society for Psychophysiological Research
Volume 46, Issue 6, pages 1288–1298, November 2009
How to Cite
Danielmeier, C., Wessel, J. R., Steinhauser, M. and Ullsperger, M. (2009), Modulation of the error-related negativity by response conflict. Psychophysiology, 46: 1288–1298. doi: 10.1111/j.1469-8986.2009.00860.x
- Issue published online: 12 OCT 2009
- Article first published online: 1 JUL 2009
- (Received March 27, 2008; Accepted January 22, 2009)
- Cognitive control;
- Conflict monitoring;
- Performance monitoring
An arrow version of the Eriksen flanker task was employed to investigate the influence of conflict on the error-related negativity (ERN). The degree of conflict was modulated by varying the distance between flankers and the target arrow (CLOSE and FAR conditions). Error rates and reaction time data from a behavioral experiment were used to adapt a connectionist model of this task. This model was based on the conflict monitoring theory and simulated behavioral and event-related potential data. The computational model predicted an increased ERN amplitude in FAR incompatible (the low-conflict condition) compared to CLOSE incompatible errors (the high-conflict condition). A subsequent ERP experiment confirmed the model predictions. The computational model explains this finding with larger postresponse conflict in far trials. In addition, data and model predictions of the N2 and the LRP support the conflict interpretation of the ERN.
The ability to monitor ongoing behavior and make appropriate adjustments when errors are committed is essential for successful human behavior. The error-related negativity (ERN) is an event-related brain potential associated with errors in forced-choice reaction time tasks (Falkenstein, Hohnsbein, Hoormann, & Blanke, 1990; Gehring, Goss, Coles, Meyer, & Donchin, 1993). The ERN reaches its peak within 100 ms after the erroneous response and is maximal over frontocentral electrode sites. It is likely to be generated in the rostral cingulate zone (RCZ) on the posterior frontomedian wall (Debener et al., 2005; Ridderinkhof, Ullsperger, Crone, & Nieuwenhuis, 2004). On a within-subject level, the amplitude of the ERN predicts post-error slowing (Debener et al., 2005; Gehring et al., 1993; but see Gehring & Fencsik, 2001), supporting the notion that the RCZ plays a major role in signaling the need for adjustments when errors have occurred.
However, the exact nature of the processes underlying the ERN and the models best describing these processes are still under debate. A recent model suggested that activity in the posterior mesial frontal cortex is driven by conflict during response selection (response conflict, i.e., the simultaneous activation of incompatible actions; Botvinick, Braver, Barch, Carter, & Cohen, 2001; Carter et al., 1998). Within this framework, the ERN has been proposed to reflect postresponse conflict, that is, the conflict between the executed, erroneous response tendency and the still-evolving correct response tendency (Yeung & Cohen, 2006; Yeung, Cohen, & Botvinick, 2004). Connectionist models of conflict monitoring have been used to make testable predictions about the properties of the ERN (e.g., Yeung et al., 2004), and a number of findings are in line with these predictions (Rodriguez-Fornells, Kurzbuch, & Munte, 2002; Ullsperger & von Cramon, 2006; Yeung et al., 2004). However, some studies confirmed the predictions only in part (Fiehler, Ullsperger, & von Cramon, 2005), and others have directly challenged the conflict monitoring model (Carbonnell & Falkenstein, 2006; Maier, Steinhauser, & Hübner, 2008; Masaki, Falkenstein, Sturmer, Pinkpank, & Sommer, 2007; Masaki & Segalowitz, 2004).
The present study sought to test the predictions of the conflict monitoring model and its ability to explain the ERP results in a modified arrow version of the Eriksen flanker task, where the degree of conflict was modulated between task conditions. In this task, participants are asked to respond to a central target arrow that is flanked by distracting arrows. The version employed here was designed to modulate the degree of response conflict by varying the distance between target arrow and flanking arrows. Flankers were presented either close to (CLOSE condition) or far from the target arrow (FAR condition). For the letter version of the flanker task, earlier studies have shown that the flanker interference effect decreases with increasing distance between flankers and target (Egeth, 1977; Eriksen & Eriksen, 1974; Gathercole & Broadbent, 1987; Miller, 1991). Consistent with these studies, we expected an increase in interference for incompatible trials in the CLOSE condition. This increase in interference was assumed to be reflected in slower reaction times as well as higher error rates in CLOSE incompatible trials relative to FAR incompatible trials. In Experiment 1, this was tested behaviorally. In a second step, we adapted the connectionist model described by Yeung and colleagues (2004) to our task and simulated the behavioral findings of the first experiment. By extracting activities from the output layer of the model we made predictions about the behavior of the ERN, the lateralized readiness potential (LRP), and the N2 based on the conflict monitoring model. According to the conflict monitoring model, the N2 reflects preresponse conflict in correct trials, the ERN reflects postresponse conflict in error trials, and the response-locked LRP is a measure of correct and wrong response tendencies and therefore could provide additional evidence for the presence of response conflict. From the simulation the prediction was derived that the ERN amplitude is larger after FAR incompatible errors than after CLOSE incompatible errors. According to the model, this is due to the fact that postresponse conflict is larger in FAR than in CLOSE errors. Note that the situation is different for preresponse conflict: In correct trials preresponse conflict is larger in CLOSE than in FAR trials, which is reflected in the N2 amplitude. In Experiment 2, we recorded ERPs in a different set of participants and tested the predictions regarding the ERN, N2, and LRP derived from the simulation.
Eleven healthy volunteers with normal or corrected-to-normal vision (5 women; mean age: 32.3 years, range: 19–59) participated in this experiment after signing informed consent. According to the Edinburgh Handedness Inventory (Oldfield, 1971), 7 participants were right-handed, 4 were left-handed. The experiment was carried out in accordance with the Declaration of Helsinki.
A speeded modified flanker task was employed (cf. Fiehler et al., 2005). Participants had to respond as quickly and accurately as possible to a target arrow presented in the center of the screen for 34 ms. The target could point either to the right or the left side, requiring a corresponding response with the right or left thumb. The onset of the target arrow was delayed by 80 ms from the onset of four vertically arranged flanker arrows (size of target and flankers: 1.9°× 1.3° of visual angle). Two of the irrelevant flankers were displayed above the screen center and two below it. Importantly, the distance between the target and the flankers varied, thereby creating a “flankers far” (FAR) and a “flankers close” (CLOSE) condition. In the far condition the four flanker positions were 6.5° and 4° above and below the center of the screen; in the CLOSE condition the flankers were presented 3.5° and 1.75° above and below the center.
A thousand trials were equally distributed across four conditions: (a) far compatible (FAR-COM), in which flankers were presented far away from the target and pointed in the same direction as the target (compatible trials), (b) far incompatible (FAR-INCOM), in which flankers were presented far away and pointed in the opposite direction as the target (incompatible trials), (c) close compatible (CLOSE-COM), in which flankers were presented close to the target, pointing in the same direction, and (d) close incompatible (CLOSE-INCOM), in which flankers were presented close to the target, pointing in the opposite direction as the target.
An individually adapting response deadline was initially set to 400 ms and then was adjusted continuously during the experiment depending on the error rate in FAR conditions based on all preceding trials. If the participant's error rate was less than 10%, the response deadline was shortened by 10 ms down to a minimum of 250 ms. When the error rate was higher than 20%, the response deadline was prolonged by 10 ms up to a maximum of 1000 ms. At the end of each trial, there was a response–stimulus interval of 1000 ms. Thus, trial length varied between 1364 ms and 2914 ms.
Participants were instructed to press a separate “error signaling” button whenever they noticed an error. The signaling button was located in the middle between the two target response buttons, so participants could choose their left or right thumb for error signaling. When the signaling button was pressed, the intertrial interval was prolonged by 800 ms. If participants responded to the target only after the response deadline, the response was counted as late response (but not as an error, unless the late response was incorrect as well).
Every 100 trials, participants received feedback about their total number of correct responses and errors and whether they should try to speed up their reactions (criterion for speed-up message: more than 10 late responses within the last 100 trials). Trials were presented in a pseudorandomized order. Stimuli were presented using Presentation 10.3 (Neurobehavioral Systems, San Francisco, CA).
A repeated measures analysis of variance (ANOVA) revealed that participants committed more errors in the close condition than in the FAR condition, F(1,10)=38.58, p<.001, and, as expected, they also committed more errors in incompatible trials than in compatible trials, F(1,10)=52.23, p<.001. In addition, there was a significant interaction between flanker distance and compatibility, F(1,10)=63.36, p<.001. In accordance with our hypotheses, the most error-prone condition was CLOSE-INCOM, the least error-prone condition was CLOSE-COM (Figure 1a).
Regarding the reaction times (RT; Figure 1b), there was a main effect for compatibility, F(1,10)=272.20, p<.001, and a significant interaction between flanker distance and compatibility, F(1,10)=46.05, p<.001. There was no main effect for flanker distance, F(1,10)=0.35, p=.57. Subsequent t-tests showed that RTs were slower in the CLOSE-INCOM condition than in the FAR-INCOM condition, t(10)=4.87, p=.001, and the opposite result for compatible conditions (FAR-COM slower than CLOSE-COM), t(10)=4.98, p=.001.
In every condition, mean RTs were faster in error trials than in correct trials: FAR-COM: 262 ms (vs. 311 ms in correct trials; p=.019), FAR-INCOM: 271 ms (vs. 385 ms; p<.001), CLOSE-COM: 268 ms (vs. 299 ms; p=.17), CLOSE-INCOM: 258 ms (vs. 399 ms, p<.001).
In the analysis of the error-signaling data, 4 participants had to be excluded from the data set because they did not signal any errors. They reported that they felt they had not enough time to signal their errors and then be in time for the answer of the next trial, although the stimulation program was set up in a way that the intertrial interval was extended by 800 ms whenever participants signaled an error. The mean error signaling rate across conditions was 50.45%. There was no difference in signaling rates between conditions (p>.20). Mean signaling time (time between erroneous response and signaling button press) across conditions was 472 ms. Again, there was no difference between conditions (p>.18). Although not explicitly encouraged to do so, participants corrected 27.13% of their errors spontaneously. Correction rate did not differ between conditions (p>.13).
The twofold purpose of Experiment 1 was (a) to show on a behavioral level that our experimental manipulation is able to induce a conflict modulation within the incompatible condition and (b) to generate RTs and error rates that could be used to simulate this task within a connectionist model. With the present manipulation, we were able to replicate earlier results (cf. Miller, 1991), showing that the flanker-target distance can influence error rates and RTs in an arrow version of the Eriksen flanker task. Similar results have already been demonstrated for the letter version of the flanker task (Egeth, 1977; Eriksen & Eriksen, 1974; Gathercole & Broadbent, 1987; Miller, 1991). As expected, the flanker interference effect was reduced in the FAR condition compared to the CLOSE condition. In the incompatible condition, the flankers were located less centrally in the focus of attention and thus had a weaker effect on the development of a response activation competing with the target-driven response tendency. This is also reflected in longer RTs for incompatible errors in the FAR condition than in the CLOSE condition.
The present experiment replicated the common finding that RTs are shorter in error trials than in correct trials (e.g., Falkenstein, 2004; Masaki et al., 2007). This suggests that errors are due to premature responding.
Error signaling data were acquired to check how many errors were consciously perceived by the participants. However, some of the participants did not signal any errors because they thought that, due to signaling, they would not be in time for the next trial. Moreover, possibly because of the same reason, the remaining participants showed much lower signaling rates than what we obtained in an earlier study (Ullsperger & von Cramon, 2006). Therefore, we slightly changed the instruction for Experiment 2 regarding the signaling (see below).
Next we explored the effects of flanker distance on postresponse conflict using a computational model. Our goal was to derive predictions as to how flanker distance affects the ERN, the N2, and the LRP.
We modified the connectionist model of the Eriksen flanker task used by Yeung et al. (2004) to apply to our task (see Figure 2). A stimulus layer contained one unit for each possible combination of a stimulus (<, >) and a spatial location (above target far (AF), above target close (AC) target, below target close (BC), below target far (BF)). In this way, each possible stimulus array was represented as a unique pattern of activated stimulus units. The response layer contained one unit for each possible response (left, right). Each stimulus unit was associated with its corresponding response unit. To ensure that the units representing the target had the strongest influence on the response units, the stimulus layer received modulatory input from an attention layer containing one unit for each possible stimulus location. The pattern of activation in the attention layer represented the allocation of attention to the respective stimulus locations.
The model was simulated as described by Yeung et al. (2004) with some minor changes. A trial started by activating stimulus and attention units. To implement the stimulus onset asynchrony (SOA) between flanker and target, the flanker units in the stimulus layer were activated five cycles prior to the target unit.1 Stimuli in the FAR condition and the CLOSE condition were implemented by activating only the outer or inner flanker units, respectively. To simulate an attention gradient across flanker positions, activation in the attention layer was higher for the inner than for the outer flanker positions. A response was selected when activation of a response unit exceeded a response criterion. Simulated response time was calculated as the number of cycles until a response was selected. A simulation consisted of 10 runs of 1,000 trials per condition (cf. Yeung & Cohen, 2006).
Response conflict was calculated as the product of response activations in each cycle. Accordingly, simulated response conflict was high whenever both responses were concurrently activated, which occurs when target and flankers activate different response units. The ERN was assumed to reflect the increased postresponse conflict on error trials relative to trials with correct responses. This increased postresponse conflict following errors resulted from the tendency of the model to build up a correct response. Basically, errors in this model reflect premature responses due to noise. Following such an error, the stimulus continued to activate the correct response. This resulted in a postresponse conflict because both responses were concurrently activated after the error. The simulated ERN was computed as the difference in response-locked response conflict between errors and correct trials (Yeung & Cohen, 2006; Yeung et al., 2004).
Following Yeung et al. (2004) and Yeung and Cohen (2006), we calculated a simulated N2 as the difference in stimulus-locked response conflict between correct incongruent trials and correct congruent trials. The resulting waveform typically peaks prior to the response because incongruent stimuli induce a preresponse conflict which, however, is resolved when a response unit's activation exceeds a threshold. This waveform can be used as a predictor of the same difference wave on fronto-central electrodes, which mainly reflects difference voltages in the time range of the N2. Finally, the simulated LRP was calculated as the difference in response unit activation between the correct and the incorrect response unit. This simplified procedure was chosen because no differences between response side can emerge in the model (which is normally corrected by means of double subtraction; Coles, 1989).
In a first step, the model parameters were adjusted to fit the behavioral data obtained in Experiment 1. It was not our goal to determine an optimal fit by means of exhaustive parameter search (cf. Steinhauser, Maier, & Hübner, 2008). Rather, we wanted to adjust as few parameters as possible to maintain comparability to earlier models. Only two parameter changes were necessary to achieve this goal. First, the external input to the attention units had to be adjusted to simulate the reduced attention allocation to outer flankers as compared to the inner flankers. In the original model, external input to the flanker units, extflanker, was computed as (3−exttarget) × 0.5. In our model, the external input to the outer and inner flanker corresponded to (3−exttarget) × 0.03 and (3−exttarget) × 0.47, respectively. Second, the response criterion K was increased from 0.18 to 0.20 to compensate for the increased susceptibility to errors due to the introduction of an SOA2 and to adapt the model to the empirical data. All other parameters were the same as in the simulations by Yeung et al. (2004). As can be seen in Figure 1, these moderate changes led to a remarkably good correspondence between model results and empirical data.
In a second step, we computed the simulated ERN for the CLOSE-INCOM and the FAR-INCOM conditions. The model predicted an increased amplitude of the simulated ERN in the FAR-INCOM condition (Figure 3a). As the simulated ERN is simply the difference between the response conflict in error trials and correct trials, this result is a direct consequence of the increased postresponse conflict following errors in the FAR-INCOM condition (Figure 3b). Why was the postresponse conflict in FAR-INCOM error trials increased relative to CLOSE INCOM error trials? Remember that response conflict is the result of the multiplication of the correct response activation times the incorrect activation. Given that the two response units inhibit each other, this product (i.e., the response conflict) will be maximized when the two responses are activated to a similar degree. Inspection of Figure 3c,d reveals that the two response tendencies were more similar to each other after FAR-INCOM errors, thus leading to a higher degree of response conflict after these errors. Taking this line of reasoning a step further, one might then ask why the correct response activation was higher after FAR-INCOM error trials. The reason for this is that, after an error has occurred due to noise in the system, the response layer receives further input from the continued processing of the stimulus. The input for the correct response will be higher in FAR trials because the flankers in FAR trials receive less attentional weight and thus will be less able to interfere with the buildup of activation in the correct response unit.
In a third step, we computed simulated LRPs for incorrect trials of the CLOSE-INCOM and the FAR-INCOM condition (see Figure 5, below). The model predicts a strong effect on the LRPs. At first, the time course of the simulated LRP is almost identical until the response has occurred, but then the offset of the LRP occurs earlier in the FAR-INCOM condition. This earlier offset is due to the stronger correct response tendency in these trials, which leads to a stronger corrective activity after an error.
Finally, we computed the simulated N2 for the CLOSE and the FAR condition (see Figure 6). Here, the model predicts a larger difference wave between incongruent and congruent trials for the CLOSE condition than for the FAR condition. This reflects the fact that incongruent flankers activated the incorrect response more strongly in the CLOSE than in the FAR condition.
Taken together, the simulation confirmed our expectation that conflict monitoring theory predicts an increased ERN for stimuli with an increased flanker distance. Moreover, the model predicts an increased effect of congruency on the N2 for the CLOSE than for the FAR condition as well as stronger corrective activity following an error for the simulated LRP of the FAR-INCOM condition.
The predictions for the ERPs based on the different theories and made by the simulation were tested in Experiment 2, in which a different group of participants, who were unfamiliar with the paradigm, performed the flanker task, which was identical to that in Experiment 1, except that we slightly changed the instruction regarding the error signaling procedure: It was emphasized that participants should always signal their errors, and it was explicitly clarified that the intertrial interval would be prolonged after error signaling, so that participants would not be “late” for the next trial. Within the same session, participants performed an additional task that will not be discussed here. The order of the two tasks was counterbalanced across participants.
Twenty-one neurologically and psychiatrically healthy volunteers (11 male, 10 female; mean age: 23.4 years, range: 18–33; 19 right-handed) participated in the electroencephalogram (EEG) experiment after signing informed consent. One male participant had to be excluded from data analysis because he did not follow the experimental instructions.
ERP Data Collection
Participants were seated in a dimly lit, electrically and acoustically shielded chamber. EEG activity was recorded with Ag/AgCl sintered electrodes mounted in an elastic cap (Easycap, Herrsching, Germany) from 60 scalp sites of the extended 10–20 system. The ground electrode was positioned at F2. The vertical electrooculogram (EOG) was recorded from electrodes located above and below the left eye. The horizontal EOG was collected from electrodes positioned at the outer canthus of each eye. Electrode impedance was kept below 5 kΩ. Potentials were online referenced on electrode CPz and re-referenced off-line by subtracting the average of all electrodes from each individual electrode signal. The EEG and EOG were recorded continuously and were A-D converted with a 16-bit resolution at a sampling rate of 500 Hz using BrainAmp MR plus amplifiers (Brain Products, Gilching, Germany).
ERP Data Analysis
All EEG data processing was done using EEGLAB 6.0b (Delorme & Makeig, 2004) and custom routines in Matlab 7.01. The signal was band-pass filtered between 0.5 Hz and 40 Hz and split into stimulus-locked epochs of 1700 ms that spanned a time window from 200 ms prestimulus to 1500 ms poststimulus. Epochs that contained more than one button press as a response (e.g., spontaneous corrections) were excluded from further analysis. Two artifact rejection methods were applied: (a) calculation of the joint data probability to detect epochs containing “improbable data,” that is epochs with data deviating more than 5 SD from the epochs mean probability distribution were excluded, and then (b) a temporal infomax independent component analysis (ICA) was computed for isolating artifacts in the EEG signal (Delorme, Sejnowski, & Makeig, 2007). Components representing eyeblinks, muscle artifacts, or other types of noise, were removed from the signal. A baseline correction was applied, with the 120-ms preflanker signal of each epoch being defined as baseline. These preprocessed and baseline-corrected data were re-epoched in order to get response-locked epochs. The new response-locked epochs spanned a time window from 200 ms preresponse to 400 ms postresponse.
ERN amplitudes were measured from the pre-ERN trough to the peak of the ERN (search window for peak: from response to 150 ms postresponse; search window for trough: from 100 ms preresponse to ERN peak) for each subject individually at FCz. ERNs were only computed for incompatible error trials, because of an insufficient number of error trials in the compatible conditions. The amplitude for incompatible correct trials (correct-related negativity, CRN) was determined with the same procedure as the ERN.
To compute response-locked LRPs, the difference of the signal at positions C3 and C4 was calculated (procedure as described by Coles, 1989) for error trials. A clear wrong response tendency was observed in the LRP data, leading to erroneous responses. In order to quantify the earlier offset of the LRP in FAR-INCOM than in CLOSE-INCOM error trials, which was present in the simulation data, the offsets of the LRPs were defined as the points in time where the activation of the LRP dropped below 20% of the peak-to-baseline amplitude (20% criterion). To cope with the low signal-to-noise ratio of the LRP, which leads to validity problems when determining on- and offset differences of LRPs, a jackknife procedure (Miller, Patterson, & Ulrich, 1998; Ulrich & Miller, 2001) was employed. The latencies of the offset for FAR and CLOSE LRPs were then compared using a t test. Following the recommendations of Miller et al. and Ulrich and Miller for the jackknife procedure, correctional formulas were applied to these t statistics to account for the changes in variance due to the jackknifing procedure. For visualization purposes, the LRPs were smoothed via a moving average window of 30 sampling points.
To characterize the N2, difference waves for FAR and CLOSE correct trials were calculated based on stimulus-locked epochs by comparing the signal for FAR-INCOM versus FAR-COM and CLOSE-INCOM versus CLOSE-COM, respectively, analogous to the N2 calculation in the connectionist model. N2 amplitudes were then measured from pre-N2 trough to N2 peak (search window for peak and preceding trough: from 50 ms to 600 ms poststimulus) for each participant individually at FCz.
As in Experiment 1, participants committed more errors in the CLOSE condition than in the FAR condition, F(1,19)=95.07, p<.001 (Figure 1e). Also, we found a main effect for compatibility, F(1,19)=128.42, p<.001, with more errors being committed in incompatible trials. Again, there was a significant interaction between flanker distance and compatibility, F(1,19)=119.35, p<.001. Note that although the error rate in the FAR-INCOM condition is rather high, this is not due to guessing behavior of the participants: The false alarm rates (i.e., signaling an error when the answer was actually correct) were extremely low for all conditions (<0.01%), and, in addition, the error rate is rather low in the CLOSE-COM condition, indicating that the high error rate is specific for incompatible trials.
The RT effects also replicated the results of Experiment 1 (Figure 1f). Participants responded faster in compatible trials than in incompatible trials, F(1,19)=296.19, p<.001. There was a significant interaction between flanker distance and compatibility, F(1,19)=48.36, p<.001, but no main effect for flanker distance, p=.90. Thus, the behavioral data of Experiment 1 were replicated in this experiment.
In both incompatible conditions, 83% of all errors were consciously perceived, that is, errors that were signaled by the participant. There was no significant difference in the number of perceived errors between FAR-INCOM and CLOSE-INCOM conditions, t(19)=−0.49, p=.62. Therefore, we decided to include only the perceived errors in the ERP analysis. The number of nonsignaled errors was insufficient for separate statistical analyses. Signaled and nonsignaled errors were not collapsed in the analysis because signaled errors were followed by an additional motor response and nonsignaled errors were not. By excluding nonsignaled errors, the number of motor responses per trial was kept constant for all error-related ERPs. Mean signaling time across conditions was 456 ms. Signaling times did not differ between conditions (p>.3). Some errors were corrected spontaneously, although participants were not instructed to do so. Errors in the FAR-INCOM condition were more often corrected spontaneously (11%) than errors in the CLOSE-INCOM condition (8%), t(19)=2.19, p=.04. These errors were also excluded from the ERP analysis due to a variable number of motor responses following these errors.
Error-related negativity. A two-way repeated measures ANOVA with the factors response correctness (erroneous vs. correct) and flanker distance (close vs. far) revealed significant main effects of both error/correct, F(1,19)=48.31, p<.0001, and distance, F(1,19)=5.84, p=.026, and a significant error/correct response × distance interaction, F(1,19)=15.67, p=.001. In error trials, the ERN amplitude was higher in far than in close trials, but there was no difference between FAR and CLOSE in correct trials. FAR-INCOM and CLOSE-INCOM conditions were compared with a paired t test contrast. The ERN amplitudes showed a significant effect of flanker distance, with a mean amplitude of 9.39 μV (SD=3.97) in the FAR-INCOM condition and 7.37 μV (SD=3.78) in the CLOSE-INCOM condition, t(19)=3.71, p=.001 (Figure 4). When ERN and CRN amplitudes were determined as mean amplitudes (instead of trough-to-peak measures) in the time window from response to 150 ms postresponse, analyses revealed the same results as above.
Lateralized readiness potential. The simulation data suggested that the LRP for FAR-INCOM error trials shows an earlier offset of the wrong-response tendency than the LRP for CLOSE-INCOM trials (Figure 5b). According to the model, this is due to a stronger correct response tendency after the error in FAR than in CLOSE trials. However, the comparison of the offset of these two LRPs in Experiment 2 revealed only a tendency toward an earlier offset in FAR error trials, tc(19)=1.26, p=.056, one-tailed (Figure 5a). As predicted by the simulation, there was neither a difference in LRP amplitude between FAR and CLOSE errors, tc(19)=.779, p=.223, nor a difference in peak latency of the LRPs, tc(19)=−0.604, p=.724.
N2. The comparison of the N2 difference wave for FAR and CLOSE trials (INCOM-COM for each condition) revealed a significantly larger amplitude in CLOSE than in FAR trials, t(19)=−3.14, p=.005. The mean trough-to-peak difference wave amplitude in CLOSE trials was 5.48 μV (SD=2.57); the mean amplitude in FAR trials was 4.28 μV (SD=1.79). There was no difference in latency between these two waves, t(19)=−.77, p=.45 (Figure 6).
The influence of different degrees of conflict on the ERN was investigated in order to test predictions made by the conflict monitoring theory and its implementation in a formal connectionist model. The conflict modulation was achieved by varying the flanker–target distance in an arrow version of the Eriksen flanker task. The behavioral data of Experiments 1 and 2 showed that the conflict manipulation was successful, which was reflected in higher error rates and slower RTs in CLOSE-INCOM trials than in FAR-INCOM trials. In the following, the predictions of the conflict monitoring model for the ERN, the LRP, and the N2 will be discussed in turn. Then our results will be compared with predictions from alternative ERN theories.
The conflict monitoring model predicted that the ERN would be larger after FAR error trials. Our ERP experiment confirmed this modeling result. The model predicted this effect because of the higher postresponse conflict in FAR error trials. Our results are qualitatively similar to Simulation 2 in the study by Yeung et al. (2004), who showed that postresponse conflict and, as a consequence, the ERN amplitude were higher after congruent errors than after incongruent errors. Both results may appear counterintuitive at first sight because congruent trials in the simulation of Yeung et al. and FAR-INCOM trials in our simulation were associated with lower response conflict on correct trials. This might lead one to assume that response conflict following errors should also be lower after these trials. However, the simulations showed that this was not the case. As explained in the Simulation section, the reason for this is that, after an error has occurred, the incorrect response receives less activation from the flanker in the FAR condition (because the FAR flanker receives less attentional weight) and thus the incorrect response will interfere less with the correct response. In the simulation of Yeung et al., a similar situation arises after congruent errors because the incorrect response unit does not receive activating input from the stimulus unit and thus will also interfere less with the buildup of activation in the correct response unit. Thus, the increased ERN after FAR errors is a direct consequence of the network architecture and its dynamics after adapting the network to the behavioral data from Experiment 1.
Although our results are in agreement with the predictions of the conflict monitoring model, other recent studies have questioned the model's ability to explain the observed ERN effects (Burle, Allain, Vidal, & Hasbroucq, 2005; Carbonnell & Falkenstein, 2006; Masaki et al., 2007). These studies have in common that the degree of response conflict was defined by peripheral measures such as electromyography (EMG; Burle et al., 2005; Masaki et al., 2007) or button press force (Carbonnell & Falkenstein, 2006).
For example, Carbonnell and Falkenstein (2006) investigated ERP differences between “full errors” and “partial errors.”Full errors were defined as trials in which an incorrect response tendency exceeded an arbitrarily defined force threshold, partial errors were defined as trials in which the correct response was preceded by an incorrect response tendency, which, however, did not reach the threshold. The onset for the ERP analysis was time-locked to the incorrect response in full error trials and to the onset of the incorrect response tendency in partial error trials. On the basis of the button press force, Carbonnell and Falkenstein argued that response conflict was higher in full error trials. As they found no significant ERP amplitude differences between full and partial errors, they concluded that the conflict monitoring theory cannot explain their results.
However, the question arises of whether partial error trials are true error trials. One could argue that these trials are, in fact, correct trials with a weak but measurable incorrect response tendency at the beginning of the trial. Interpreted this way, one would not expect to find an ERN in partial error trials but instead an N2. Thus, Carbonnell and Falkenstein (2006) might have compared two different ERP components. In support of this interpretation, also Yeung et al. (2004) associated partial errors with the N2, for which the conflict monitoring model makes different predictions (Yeung & Cohen, 2006; Yeung et al., 2004).
In a similar vein, Burle et al. (2005) reported ERPs only for correct trials that showed an incorrect response tendency in the EMG before the final response. The reported correct response “ERN” was locked to the onset of the wrong response tendency. However, the correct response tendency, finally leading to the correct response, started only about 45 ms after the “ERN” peak. Thus, the analyzed ERP signal occurred clearly before the button press.
Note that the approach of quantifying response conflict via peripheral measures relies on the assumption that task-induced conflict can actually be detected in distal muscles. However, it has been shown that incorrect response activations can occur without any EMG activations (cf. Coles, 1989).
In the simulation, the LRP was computed as the difference between the activation of the correct response (Figure 3c) and the activation of the incorrect response (Figure 3d). The simulation predicted that this difference would decrease earlier following far errors (Figure 5). This prediction resulted from the higher correct response tendency combined with the lower incorrect response tendency after FAR errors. In Experiment 2, the error-related LRP showed a tendency toward an earlier decrease after FAR errors. However, this result should be interpreted cautiously because the error responses analyzed in Experiment 2 were always followed by a second motor response, the pressing of the signaling button. Thus, superimposed on the post-error LRPs is the development of an additional LRP associated with error signaling. As participants were instructed to signal their errors, errors that were not followed by a signaling button press occurred so infrequently that they could not be analyzed separately, thus leaving open the question of whether more pronounced LRPs would have resulted after errors that were not followed by a subsequent button press.
It has been proposed that the N2 and the ERN reflect very similar processes and also have a common origin in the anterior cingulate cortex (for reviews, see Folstein & Van Petten, 2008; van Veen & Carter, 2002). Based on a flanker task, Yeung et al. (2004) suggested that the N2 occurs before the response in correct trials and after the response (as ERN) in error trials. Therefore, we investigated whether the experimental FAR–CLOSE manipulation that affects the ERN amplitude also has an impact on the N2 amplitude.
In the time range of the N2, the amplitude of the stimulus-locked difference between correct incompatible trials and correct compatible trials was larger for CLOSE trials than for FAR trials. This is in accordance with the simulation data, confirming the assumption of a higher preresponse conflict in CLOSE than in FAR trials when the given response is correct. Note that the effect of preresponse conflict in FAR and CLOSE correct trials is different from the effect of postresponse conflict on the two conditions in case of an error: Although preresponse conflict is larger in CLOSE trials, postresponse conflict is larger in FAR trials. According to the conflict monitoring model, preresponse conflict is higher in CLOSE trials because the flanking stimulus receives more attentional weight as it is closer to the attentional focus. Thus, the flanking stimulus activates the incorrect response more strongly in CLOSE-INCOM trials, which then leads to a higher response conflict.
Comparison with Other ERN Theories
An alternative account of the ERN is provided by the mismatch theory (Coles, Scheffers, & Holroyd, 2001; Falkenstein, Hohnsbein, Hoormann, & Blanke, 1991; Gehring et al., 1993), which hypothesizes that the ERN is elicited by the detection of errors (mismatch between executed response and required response). One could argue that the mismatch hypothesis predicts higher ERN amplitudes for CLOSE errors than for FAR errors (i.e., predicts the opposite of what we observed), as Coles et al. (2001) suggested that the mismatch is “directly related to the degree of activation of the incorrect response, and inversely related to the degree of activation of the correct response” (p. 178). Provided that our simulation accurately reflects the degree of correct and incorrect response activations, the mismatch theory would then predict a higher ERN amplitude on CLOSE trials. However, it was also argued that the ERN is influenced by the quality of the representation of the correct response (Scheffers & Coles, 2000). The quality of this representation could be argued to be lower in the more difficult CLOSE condition in which case the mismatch theory would predict the ERN results we found. However, signaling rates in our study did not differ significantly between conditions, which argues against the interpretation that the quality of the representation of the responses was generally lower in the CLOSE condition. In addition, only trials in which the error was signaled were analyzed. Taken together, the prediction of the mismatch theory regarding the ERN modulation found in the present study is somewhat ambiguous. It appears that the mismatch theory would most likely predict the opposite of what we found, but, given that the mismatch theory has not been implemented in a formal model, this is difficult to test.
A third account for explaining the ERN is the reinforcement learning theory (Holroyd & Coles, 2002). The reinforcement learning theory claims that the ERN reflects a negative reward prediction error. It could be argued that errors in the FAR condition are less frequent and could therefore result in a larger negative reward prediction error. For this reason, FAR errors could lead to a larger ERN, as was observed in our data. Clearly, this prediction relies on the assumption that the participants in our study built up different expectations about FAR and CLOSE trials.
Neither the mismatch theory nor the reinforcement learning theory make any explicit predictions about the course of the LRPs. Regarding the reinforcement learning theory, one could speculate that a worse-than-expected outcome is detected earlier in the case of FAR errors than CLOSE errors. Under this assumption, one could argue that the resulting adjustment processes, which involve a responsibility shift toward one of the motor controllers (Holroyd & Coles, 2002), occur earlier in FAR trials. This might lead to different LRPs in FAR and CLOSE error trials. However, this reasoning is speculative at present and would have to be directly tested in a future investigation.
The mismatch theory and the reinforcement learning theory are not explicitly concerned with the N2, either. The reason for this is that both theories in their original versions were introduced to explain ERP effects observed in the context of error trials, not correct trials. It is currently an open issue as to whether the mismatch theory and the reinforcement learning theory could be extended to incorporate these results.
It should be noted, though, that the conflict monitoring model and the reinforcement learning model are not mutually exclusive. As mentioned by Yeung et al. (2004), the conflict monitoring model suggests a mechanism for error detection, whereas the reinforcement learning theory proposes mechanisms that occur after an error has been detected.
In conclusion, the present version of the flanker task enabled us to modulate the degree of conflict in incompatible trials. The modulation in postresponse conflict was reflected in the ERN amplitude as predicted by the conflict monitoring model (Yeung et al., 2004). The simulation of the conflict monitoring model predicted a larger ERN after FAR errors and a larger N2 in CLOSE correct trials. Both of these effects were observed in our ERP data. The predicted effect on the LRP after errors was marginally significant. The strength of the conflict monitoring theory is, on the one hand, that it generates more testable predictions than the other two theories and, on the other hand, that it incorporates both the N2 and the ERN and explains them by the same mechanism. Given that theories are only useful to the extent that they lead to unequivocal and testable predictions, the conflict monitoring theory is currently the best available framework to explain the effects observed in the present experiments. Although the current data do not refute alternative explanations, they demonstrate that the conflict monitoring model accounts for a variety of phenomena in a comprehensive and yet parsimonious way.
1Target units in the stimulus layer were activated five cycles prior to the flanker units. This corresponds to an SOA of 80 ms using the transformation applied by Yeung et al. (2004). For simplicity, units in the attention layer were activated simultaneously to the target unit. This prevented unwanted effects emerging when the central attention unit but no target unit was activated. To counteract the resulting strong activation built up in the flanker units and the resulting susceptibility to errors, the units were activated by 75% only and the response criterion K was increased slightly (compared to the original simulation model by Yeung et al., 2004). Note that implementing an SOA between flankers and target was not responsible for any of the effects reported. The same predictions resulted when the model was applied without SOA.
2Figure 3 seems to suggest that corrective activity following an error (Figure 3c) occurs later than the corresponding postresponse conflict, which peaks at a time point for which the correct response unit is even activated negatively (Figure 3b). This results from the fact that response conflict is bound to zero whereas activation of response units can become negative. At each time point after the error, we average across trials with positive correct response activity and positive response conflict as well as across trials with negative correct response activity and zero response conflict. As a consequence, mean response conflict can be positive at time points for which mean response unit activity is still negative.
- 2001). Conflict monitoring and cognitive control. Psychological Review, 108, 624–652. , , , , & (
- 2005). Sequential compatibility effects and cognitive control: Does conflict really matter? Journal of Experimental Psychology. Human Perception and Performance, 31, 831–837. , , , & (
- 2006). Does the error negativity reflect the degree of response conflict? Brain Research, 1095, 124–130. , & (
- 1998). Anterior cingulate cortex, error detection, and the online monitoring of performance. Science, 280, 747–749. , , , , , & (
- 1989). Modern mind-brain reading: psychophysiology, physiology, and cognition. Psychophysiology, 26, 251–269. (
- 2001). Why is there an ERN/Ne on correct trials? Response representations, stimulus-related components, and the theory of error-processing. Biological Psychology, 56, 173–189. , , & (
- 2005). Trial-by-trial coupling of concurrent electroencephalogram and functional magnetic resonance imaging identifies the dynamics of performance monitoring. Journal of Neuroscience, 25, 11730–11737. , , , , , & (
- 2004). EEGLAB: An open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. Journal of Neuroscience Methods, 134, 9–21. , & (
- 2007). Enhanced detection of artifacts in EEG data using higher-order statistics and independent component analysis. Neuroimage, 34, 1443–1449. , , & (
- 1977). Attention and preattention. In G. H.Bower (Ed.), The psychology of learning and motivation (Vol. 11, pp. 277–320). New York: Academic Press. (
- 1974). Effects of noise letters upon the identification of a target letter in a nonsearch task. Perception & Psychophysics, 16, 143–149. , & (
- 2004). ERP correlates of erroneous performance. In M.Ullsperger & M.Falkenstein (Eds.), Errors, conflicts, and the brain. Current opinions on performance monitoring (pp. 5–14). Leipzig: MPI for Human Cognitive and Brain Sciences. (
- 1990). Effects of errors in choice reaction tasks on the ERP under focused and divided attention. Psychophysiological brain research (Vol. 1, pp. 192–195). Tilburg, Germany: Tilburg University Press. , , , & (
- 1991). Effects of crossmodal divided attention on late ERP components. II. Error processing in choice reaction tasks. Electroencephalography and Clinical Neurophysiology, 78, 447–455. , , , & (
- 2005). Electrophysiological correlates of error correction. Psychophysiology, 42, 72–82. , , & (
- 2008). Influence of cognitive control and mismatch on the N2 component of the ERP: A review. Psychophysiology, 45, 152–170. , & (
- 1987). Spatial factors in visual attention: Some compensatory effects of location and time of arrival of nontargets. Perception, 16, 433–443. , & (
- 1993). A neural system for error-detection and compensation. Psychological Science, 4, 385–390. Direct Link:, , , , & (
- 2002). The neural basis of human error processing: Reinforcement learning, dopamine, and the error-related negativity. Psychological Review, 109, 679–709. , & (
- 2008). Is the error-related negativity amplitude related to error detectability? Evidence from effects of different error types. Journal of Cognitive Neuroscience, 20, 2263–2273. , , & (
- 2007). Does the error negativity reflect response conflict strength? Evidence from a Simon task. Psychophysiology, 44, 579–585. , , , , & (
- 2004). Error negativity: A test of the response conflict versus error detection hypotheses. In M.Ullsperger & M.Falkenstein (Eds.), Errors, conflicts, and the brain. Current opinions on performance monitoring (pp. 76–83). Leipzig: MPI for Human Cognitive and Brain Sciences. , & (
- 1991). The flanker compatibility effect as a function of visual angle, attentional focus, visual transients, and perceptual load: A search for boundary conditions. Perception & Psychophysics, 49, 270–288. (
- 1998). Jackknife-based method for measuring LRP onset latency differences. Psychophysiology, 35, 99–115. , , & (
- 1971). The assessment and analysis of handedness: The Edinburgh inventory. Neuropsychologia, 9, 97–113. (
- 2004). The role of the medial frontal cortex in cognitive control. Science, 306, 443–447. , , , & (
- 2002). Time course of error detection and correction in humans: Neurophysiological evidence. Journal of Neuroscience, 22, 9990–9996. , , & (
- 2000). Performance monitoring in a confusing world: Error-related brain activity, judgments of response accuracy, and types of errors. Journal of Experimental Psychology. Human Perception and Performance, 26, 141–151. , & (
- 2008). Modeling behavioral measures of error detection in choice tasks: Response monitoring versus conflict monitoring. Journal of Experimental Psychology. Human Perception and Performance, 34, 158–176. , , & (
- 2006). The role of intact frontostriatal circuits in error processing. Journal of Cognitive Neuroscience, 18, 651–664. , & (
- 2001). Using the jackknife-based scoring method for measuring LRP onset effects in factorial designs. Psychophysiology, 38, 816–827. , & (
- 2002). The anterior cingulate as a conflict monitor: fMRI and ERP studies. Physiology & Behavior, 77, 477–482. , & (
- 2006). The impact of cognitive deficits on conflict monitoring. Predictable dissociations between the error-related negativity and N2. Psychological Science, 17, 164–171. Direct Link:, & (
- 2004). The neural basis of error detection: Conflict monitoring and the error-related negativity. Psychological Review, 111, 931–959. , , & (