Distinguishing between response conflict and error expectancy in inhibitory error processing: the role of the presupplementary motor cortex

Abstract

Error monitoring is a critical cognitive function that enables the detection of deviations from intended goals and the initiation of corrective actions. Two influential theoretical frameworks propose distinct mechanisms underlying this process: conflict detection and reinforcement learning. The conflict detection account emphasizes the recognition of incompatible response tendencies, while reinforcement learning models focus on predicting error likelihood and updating expectations based on outcomes. Disentangling the contributions of these mechanisms remains challenging, as errors frequently involve both heightened response conflict and unexpected results. The present study aimed to differentiate these mechanisms using functional magnetic resonance imaging (fMRI) in the stop-signal task. Forty-four participants completed the task, during which response conflict intensity (reflected in stop-signal delay, SSD) and error expectancy (indexed by stop-response interval, SRI) were assessed. fMRI data were analyzed to investigate how these measures relate to neural activity associated with error processing. The results revealed that both SSD and SRI influenced post-error slowing. However, only SSD—reflecting response conflict—was significantly associated with error-related brain activity, particularly in the pre-supplementary motor area and superior frontal gyrus. These findings support the conflict detection theory, emphasizing the central role of response conflict in the neural mechanisms underlying inhibitory control failures.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-42784-6.

Introduction

Error monitoring is a core cognitive function that detects and evaluates behaviors that deviate from intended goals and subsequently implements appropriate remedial actions^1,2. The role of this cognitive process is crucial, as making an error can have negative effects on adaptation and survival. One of the most influential cognitive theories of error monitoring refers to the process of conflict detection, which involves identifying co-activation between incompatible responses and triggering control processes to resolve discrepancies³. Other seminal models, rooted in reinforcement learning theory, focus on error likelihood prediction, which involves recognizing that the outcome is worse than expected^4,5 or the prediction of response outcome (the PRO model), which entails detecting that the outcome differs from what was expected, regardless of whether it is better or worse^6,7. This discrepancy calls for an update in expectations regarding the results of actions.

In most experimental studies, disentangling the effects of conflict detection and error expectancy is challenging, as errors are inherently both conflict-evoking and unexpected events. Consequently, adjudicating between these influential theoretical accounts has proven difficult⁸. One approach to distinguishing these two effects is to use a task in which both conflict intensity and error expectancy are continuous, separately measured variables, and examining which of them exhibits a stronger relationship with error-related brain activity (for further consideration, see⁸.

Such an approach is particularly feasible for inhibitory errors in the stop-signal task⁹, where participants rapidly respond to go stimuli but must cancel their response when a stop signal appears after a delay. The stop-signal delay (SSD) is controlled by an adaptive tracking algorithm: it increases after successful inhibition and decreases after failed inhibition, thereby creating substantial trial-by-trial variability in the difficulty of stopping. A crucial characteristic of this task is that conflict arises not between two competing motor responses (as in the Stroop or flanker tasks), but between processes controlling a single response – execution versus cancellation. Go processing begins with the go stimulus, whereas stop processing begins only when the stop signal appears. Errors occur when the go process reaches the execution threshold before inhibition becomes effective.

Stahl and Gibbons¹⁰ adapted conflict-monitoring theory specifically to the stop-signal task. According to their model, response conflict emerges shortly after the stop signal and increases dynamically as long as both go- and stop-related activations are co-active. Crucially, longer SSDs produce stronger conflict, because the go process has more time to accumulate activation before the stop process starts. As a result, SSD can be interpreted as a continuous, mechanistically grounded index of conflict intensity on failed-stop trials.

In contrast to conflict-based accounts, models rooted in reinforcement learning emphasize error expectancy as a key determinant of error-monitoring intensity^4–7. In the stop-signal task, error expectancy is associated with the amount of time available for the stop process to counteract the go process. Electromyographic and transcranial magnetic stimulation studies show that inhibition becomes effective only ~ 130–175 ms after stop-signal presentation^11–14, suggesting that the “point of no return” occurs roughly 100–200 ms before movement completion. Accordingly, the stop-response interval (SRI), i.e., the interval between stop-signal presentation and response execution, may serve as a continuous index of error expectancy. Errors occurring soon after the stop signal (short intervals) are highly expected, because inhibition has no time to act; errors with longer intervals are less expected, as inhibition could in principle succeed. Longer intervals therefore correspond to lower predicted error likelihood and larger prediction errors, which should amplify medial frontal error-monitoring activity⁷.

Error monitoring also manifests behaviorally through post-error adjustments, most prominently post-error slowing, an increase in response time following an error^15,16; for a review, see^17,18. Post-error slowing has been interpreted either as adaptive caution^19,20 or as an automatic orienting response driven by unexpected events²¹. Because both conflict intensity and prediction-error magnitude may influence post-error slowing^4,22, its relationship to SSD and SRI may help disentangle these accounts.

The association between inhibitory error monitoring and stop-signal delay and stop response interval in the stop-signal task has already been studied in electroencephalographic (EEG) reseach focusing on error-related negativity (ERN;²³), also known as error negativity (Ne;²⁴). ERN is a negative deflection in the EEG signal observed over mediofrontal scalp regions, typically within 100 ms after an incorrect motor response. It is most likely generated in the anterior middle cingulate cortex (aMCC)^25–28, and/or the supplementary motor^29–31, which is located within the superior frontal gyrus (SFG) and includes the supplementary motor area (SMA), the pre-supplementary motor area (pre-SMA) and the supplementary eye field¹. Stahl and Gibbons¹⁰ reported that longer SSDs, interpreted as trials with stronger response conflict, were associated with larger ERN amplitudes, supporting predictions derived from conflict-monitoring theory. In contrast, Senderecka and Szewczyk³² observed that longer SRIs, reflecting lower error expectancy, predicted larger (more negative) ERNs and increased post-error slowing, while no relationship was found between ERN amplitude and SSD. Together, these findings suggest that SSD and SRI, although correlated, may contribute to different aspects of the error-monitoring process. As a result, the existing EEG evidence remains mixed, underscoring the need for further research capable of jointly examining conflict detection and error expectancy.

Given these mixed EEG findings, a complementary approach is needed to clarify the neural mechanisms underlying inhibitory error monitoring. Functional magnetic resonance imaging (fMRI) offers a key advantage over EEG: it allows dissociating the specific anatomical loci contributing to error-related signals that, at the scalp level, reflect mixtures of overlapping neural sources and therefore cannot be spatially separated with sufficient precision³³. The stop-signal task has been widely used in fMRI research to investigate neural correlates of error processing. e.g.,^34–39. Of particular relevance is the study by Hughes et al.⁴⁰ who observed that uninhibited motor responses in the stop-signal task evoke larger activity within the MCC and pre-SMA than inhibited responses. They also found that the probability of successful inhibition correlates negatively with the activity in these regions. Importantly, they analyzed both successful and failed inhibition trials, capturing processes more strongly associated with inhibitory control than with error monitoring. In our study, we took it a step further and explored activity specifically correlated with error monitoring itself. We aimed to determine which brain area, if any, exhibits activity associated with the variance in response conflict intensity and/or error expectancy during error monitoring in the stop-signal task. Our overarching goal was to reveal which process – conflict detection or reinforcement learning – is more involved in the monitoring of inhibitory errors.

Results

Behavioral results

Table 1 presents basic statistics derived from data collected from 43 participants. Data from three participants consisted of only one run due to technical problems with data registration. The majority of individuals completed go trials with near-perfect accuracy. The percentage of correctly inhibited responses after the stop signal was around 50%. The mean reaction time for correct go trials following an incorrect stop trial was 552.49 ms (SD = 145.65), while the mean reaction time for correct trial preceding the incorrect stop trial was 524.7 ms (SD = 160.29). A significant difference was found between these two values, indicating behavioral post-error slowing (t(969) = 4.71; p < 0.001; CI = [0.02, 0.04]). Post-error slowing was positively correlated with overall stop accuracy (r = 0.250, p = 0.023), indicating that participants who inhibited more successfully overall also tended to slow down more following errors.

Table 1.

Summary of basic statistics for 43 participants.

	Correct go [%]	Correct stop [%]	Correct go rt [ms]	Stop-signal reaction time [ms]	Stop-signal delay [ms]	Stop-response interval [ms]	Post-error slowing [ms]
Mean	97.00	53.16	535.64	286.84	258.32	200.08	29.36
SD	2.77	5.18	81.86	157.31	91.90	115.24	49.53

Following a basic statistical analysis of general task performance, we examined the relationships between the target variables. First, we checked the correlation between the two predictors, i.e., our measures of response conflict intensity (SSD) and error expectancy (SRI), calculated for incorrect (uninhibited) stop trials (for further details, see the Methods section). The analyses revealed that SRI and SSD were negatively correlated, r = − 0.529, p < 0.001. This relationship is expected, as longer SSDs reduce the temporal window available for inhibition. However, although the predictors share moderate overlap, they are not collinear, indicating that each contributes unique variance to the models. Second, using linear mixed-effects models (implemented with the ‘statsmodels’ package in Python⁴¹), we explored the relationship between two predictors and one dependent variable, i.e., post-error reaction times in correct go trials. We found that longer SRIs in incorrect stop trials were associated with longer reaction times in post-error correct go trials (β = 0.251, z = 5.521, p < 0.001). Similarly, longer SSDs were associated with longer response times in the correct go trials following an error (β = 0.592, z = 8.141, p < 0.001). Thus, the degree of behavioral post-error slowing was related to both the intensity of response conflict and error expectancy. Detailed analysis results are presented in Table 2.

Table 2.

Results of the linear mixed-effects model regression, with stop-response interval and stop signal delay as predictors of post-error slowing (n = 970), grouped by subject.

	Coeff	SE	CI	Z-value	p
Dependent variable: response time in the go trial following an error
Intercept	0.343	0.026	[0.291, 0.395]	12.968	< 0.001
Stop-response interval	0.251	0.045	[0.162, 0.340]	5.521	< 0.001
Stop signal delay	4.513	1.390	[1.789, 7.237]	3.247	0.001

Neuroimaging results

After the behavioral analysis, we examined the two predictors in relation to changes in the blood oxygen level-dependent (BOLD) signal to identify regions exhibiting greater activity during error processing compared to correct response processing. To this end, we performed a general linear model (GLM) analysis using a subtraction contrast (incorrect stop > correct go). We identified several clusters with significant activity after family-wise error (FWE) correction, including the right insula, right supramarginal gyrus, left insula, right lingual gyrus, pre-SMA/SFG, left supramarginal gyrus (two clusters) (see Fig. 1; Table 3). These activations closely resemble those previously reported in the literature^42–44. The highest activity was observed within the right insula (675 voxels), which also formed a cluster with the right inferior frontal gyrus (675 voxels). The second most significant cluster was located within the right supramarginal gyrus and right angular gyrus (785 voxels), while the third cluster exhibited the highest peak within the left anterior insula and left frontal operculum (404 voxels). Following this, a cluster of 199 voxels was detected within the lingual gyrus bilaterally and the calcarine cortex. Activity was also observed within the right superior frontal gyrus, medial segment, extending towards the supplementary motor cortex (473 voxels). This cluster also encompassed the pre-SMA, a region frequently associated with error-processing^45,46. Moreover, substantial two clusters were found consisting of the left supramarginal gyrus (102 voxels) and left supramarginal gyrus, left superior parietal lobule and left angular gyrus (364 voxels). Based on the cluster size and z-values, we selected these seven clusters for further analysis.

Fig. 1.

Visualization of the contrast incorrect stop > correct go in the GLM analysis, highlighting clusters with statistically significant FWE-corrected activity, in sagittal, coronal, and axial slices. MNI coordinates = 4,30,44.

Table 3.

Results of the contrast incorrect stop > correct go in the standard GLM model of group-level data.

Region	X	Y	Z	Z-value	Number of voxels
r anterior insula	38	18	− 4	7.69	675
	30	20	− 14	6.85
	34	24	4	6.12
r supramarginal gyrus	54	− 38	50	7.08	785
	56	− 40	46	6.82
r angular gyrus	52	− 46	38	7.07	785
	40	− 56	50	6.53
	34	− 52	44	6.14
	44	− 58	44	5.75
	32	− 66	50	5.75
	46	− 48	52	5.49
	44	− 52	40	5.35
l anterior insula	− 32	20	− 8	6.99	404
	− 36	18	8	6.62
	− 38	16	− 12	6.39
	− 30	18	− 14	6.28
	− 42	8	− 8	5.31
r lingual gyrus	4	− 82	− 4	6.85	199
r superior frontal gyrus, medial segment	4	38	44	6.75	473
l frontal operculum	− 44	14	− 2	6.71	404
r supplementary motor cortex	6	18	62	6.49	473
	6	10	64	5.41
l calcarine cortex	− 8	− 94	0	6.3	199
	− 4	− 90	− 4	6.05
r superior parietal lobule	26	− 66	42	6.23	785
	32	− 62	46	5.96
l supramarginal gyrus	− 60	− 50	28	6.12	102
	− 54	− 50	32	5.88
	− 54	− 44	28	5.6
	− 52	− 44	46	6.58	364
	− 60	− 44	42	6.06
r superior frontal gyrus	12	16	64	6.09	473
	14	10	64	6.06
	10	22	60	5.96
l angular gyrus	− 36	− 60	46	5.95	364
	− 36	− 54	48	5.81
	− 38	− 60	52	5.39	364
r inferior frontal gyrus, opercular part	56	14	0	5.87	675
	54	28	2	5.46
r calcarine cortex	10	− 88	0	5.74	199
l superior parietal lobule	− 30	− 52	40	5.68	364

Time series analysis and relationship between SRI, SSD and BOLD signal amplitude within selected clusters

To test the hypotheses regarding response conflict intensity and error expectancy, time series from seven selected clusters were extracted and averaged. These clusters included: cluster 1— right anterior insula, cluster 2—right parietal cortex, cluster 3—left anterior insula, cluster 4—pre-SMA/SFG, cluster 5—right lingual gyrus, and clusters 6 and 7—left parietal cortex (as shown in Fig. 1). To preview the findings, a significant relationship was observed between time series amplitude and SSD, primarily in cluster 4 (pre-SMA/SFG), as indicated by the red circles in Fig. 1.

Averaging the BOLD signal waves from incorrect stop trials revealed a peak around the 4th second after presentation of the stop signal (see Fig. 2A). To minimize the influence of the inhibitory processes on our results, we subtracted the correct stop waves from the incorrect stop waves, matched by the same SSD (see Fig. 2C). We also tested the difference between the two waves using a repeated-measures t-test. In the resulting BOLD difference wave, the maximum value was observed around the 6th second after the stop signal, which corresponds to the peak of the canonical HRF. The most pronounced difference between the averaged incorrect and correct stop waves was observed between the 5th and 7th seconds within the pre-SMA/SFG cluster (Fig. 2B). This effect was not statistically significant in other clusters, except for the lingual gyrus, where the amplitude exhibited more negative values (see Table 4).

Fig. 2.

Averaged BOLD time series within the preSMA/SFG from 4 s before to 14 s after stop-signal presentation (panels on the left), and scatter plots with regression lines illustrating relationships between BOLD signal amplitude in the preSMA/SFG and two predictors: stop-signal delay (SSD, in ms) and stop-response interval (SRI, in ms) (panels on the right). (A) Averaged signal from all incorrect stop trials (n = 1494). (B) Difference in signal between incorrect (uninhibited) and correct (inhibited) stop trials, matched by SSD (n = 865). (C) Averaged signals from incorrect (blue) and correct (orange) stop trials, matched by SSD (n = 865). (D) Relationship between BOLD signal amplitude and SSD. (E) Relationship between BOLD signal amplitude and SRI.

Table 4.

Results of a paired-samples t-test comparing the averaged BOLD signal between incorrect and correct stop trials (matched by SSD value) in the 5–7 s window following the stop signal (N = 865).

	Right insula	Right supramarginal gyrus	Left insula	preSMA/SFG	Right lingual gyrus	Left supramarginal, cluster 1	Left supramarginal, cluster 2
t-stat	0.7223	0.0023	1.3929	2.0375	− 2.9615	1.148	0.4618
p	0.4703	0.9982	0.1640	0.0419	0.0031	0.2513	0.6443

After extracting the time series for the difference between the correct and incorrect stop trials, we performed two linear mixed-effects model regressions with SRI and SSD, respectively, as predictors of BOLD signal amplitude. No significant effect of SRI on BOLD amplitude was observed in any of the examined regions (see Fig. 2E for the results specific to the pre-SMA/SFG; further details of the analysis are presented in Table 5). In contrast, a significant positive association was found between SSDs and BOLD signal amplitudes in the pre-SMA/SFG (see Fig. 2D) and right supramarginal gyrus (comprehensive results are provided in Table 5). In other words, longer SSDs, which correspond to stronger response conflict, were associated with more pronounced activity in these regions. However, it should be noted that in the right supramarginal gyrus, the difference between the waveforms for incorrect and correct stop trials was not significant. Importantly, the associations between SSDs and BOLD signal amplitudes were not significant in the remaining five clusters, indicating that activity in only two regions was sensitive to changes in the intensity of response conflict. Plots and results for the remaining clusters are presented in the Supplementary Materials.

Table 5.

Results of the mixed-effects linear regression models with stop response interval (n = 863) and stop signal delay (N = 861) as predictors of BOLD signal amplitude (grouped by subject) measured as the difference between incorrect and correct stop trials, matched by SSD value. Results are shown for the presupplementary motor area/superior frontal gyrus (preSMA/SFG) and right supramarginal gyrus clusters.

	Coeff	SE	CI	Z-value	p
Dependent variable: BOLD signal amplitude in the preSMA/SFG
Intercept	5.953	3.066	[− 0.056, 11.962]	1.942	0.052
Stop-response interval	− 1.513	1.184	[− 3.833, 0.808]	− 1.278	0.201
Intercept	− 9.610	4.017	[− 17.484, − 1.737]	− 2.392	0.017
Stop signal delay	4.513	1.390	[1.789, 7.237]	3.247	0.001
Dependent variable: BOLD signal amplitude in the right supramarginal gyrus
Intercept	3.409	2.353	[− 1.203, 8.022]	1.449	0.147
Stop-response interval	− 1.465	0.899	[− 3.228, 0.298]	− 1.629	0.103
Intercept	− 7.006	3.190	[− 13.259, − 0.754]	− 2.196	0.028
Stop signal delay	2.668	1.104	[0.504, 4.831]	2.417	0.016

Discussion

The aim of the present study was to contribute to the ongoing theoretical debate regarding the mechanisms underlying error monitoring (for a review, see^1,8). Specifically, we sought to determine whether response conflict intensity, as described in conflict detection theory³, or error expectancy, grounded in reinforcement learning theory and the PRO model^4–7, plays a primary role in the neural processing of inhibitory errors. To disentangle these two effects, we employed the stop-signal task^47,48, in which both conflict intensity and error expectancy are measured as continuous, separate variables: longer SSDs reflect stronger response conflict¹⁰, whereas longer SRIs index more unexpected errors, thus indicating a stronger outcome prediction error³². Using linear mixed-effects models, we examined trial-wise estimates of SSD and SRI as predictors of both error-related neural activity and behavioral post-error slowing^17,18. The behavioral results revealed that both SSD and SRI significantly influence post-error adjustments, making it difficult to distinguish between the two competing theoretical accounts. However, the positive association between BOLD activity in the pre-SMA/SFG with response conflict intensity primarily supports the conflict detection perspective.

The behavioral results revealed that both SSD and SRI significantly influenced post-error slowing. Specifically, longer SRIs in incorrect stop trials were associated with greater post-error slowing, suggesting that error expectancy plays a role in behavioral adjustments following an error. This finding aligns with reinforcement learning models of error monitoring, which emphasize that greater prediction errors lead to stronger learning signals and subsequent behavioral modifications^4–7. Similarly, longer SSDs were also associated with greater post-error slowing, supporting the conflict detection theory of error monitoring³. This suggests that both error expectancy and the intensity of response conflict influence the extent of post-error behavioral adjustment. Moreover, the discrepancy between behavioral and neural findings indicates that, while post-error slowing shares some similarities with error processing based on conflict detection, it also involves more general learning processes related to the prediction of response outcomes.

Regarding neuroimaging data, we first compared activity between incorrect (uninhibited) stop trials and correctly executed go trials to identify the brain network involved in processing erroneous responses, independent of motor and go stimulus-related processes. Then, within regions of the identified network, we subtracted the activity observed during correct (inhibited) stop trials from that of incorrect (uninhibited) stop trials matched by the same SSD. This approach allowed us to eliminate the influence of stop-processing, revealing the activity specifically associated with error processing. We identified seven clusters, including two clusters located in the insular cortices bilaterally (including the frontal operculum and right inferior frontal gyrus), parietal regions with foci in the right supramarginal gyrus (including the right angular gyrus), the pre-SMA/SFG region (containing the SMA and the left medial part of the SFG), a cluster peaking in the right lingual gyrus (including the calcarine cortex), and two clusters within the left supramarginal gyrus (including the left angular gyrus and left superior parietal lobule). Previous fMRI studies have consistently linked these regions to performance monitoring^43,44. The cluster encompassing the pre-SMA/SFG plays a specific role in error processing, as it is implicated as a neural generator of the ERN, as suggested by simultaneous EEG–fMRI³¹ and intracranial^29,30 recordings. Regarding the other main clusters, the insula is typically associated with error awareness and the emotional processing of errors^49,50. In turn, activity in the parietal cortex is usually attributed to attentional processes and working memory⁵¹.

Most importantly for our research question, when examining the relationship between SSD, SRI, and BOLD signal amplitude across the seven previously identified clusters, we found no significant effect of SRI in any brain region. In contrast, SSD was positively associated with BOLD activity in the pre-SMA/SFG and the right supramarginal gyrus, indicating that these regions are sensitive to variations in response conflict intensity. This aligns with previous studies implicating the pre-SMA/SFG in conflict detection and resolution^45,46. On the other hand, the association between BOLD changes and SSD within the right supramarginal gyrus does not appear to be directly related to error monitoring, as we found no difference between stop incorrect and stop correct (inhibitory) activity following the stop signal within this region. Thus, we suggest that the right supramarginal gyrus may assist the processes managed by the pre-SMA/SFG by controlling the difficulty of withholding the initiated response, irrespective of the performed action. Taking into account previous studies, the role of the right supramarginal gyrus might be related to attentional reorienting⁵² and proactive inhibition⁵³. Overall, our findings reinforce the conflict detection theoretical account, which views errors as a special case of response conflict^3,22.

Beyond this primary interpretation, several additional factors warrant consideration. First, the posterior medial frontal cortex (pMFC) demonstrates functional specialization, with subregions associated with error processing⁴³ and post-error adjustments^17,30. Zarr & Brown⁵⁴ proposed that more abstract behavioral rules are represented in the anterior pMFC. In our study, we identified a peak in the anterior SFG (MNI: x = 4, y = 38, z = 44), extending into the pre-SMA. The hemodynamic response along the anterior-posterior axis suggests a broad role in performance monitoring; however, this specialization may exclude error expectancy processing. Additionally, although the aMCC is considered the ERN generator^25–28, we did not observe significant error-related activity in this region. This suggests that error expectancy processing may be carried out in other areas, underscoring the need for further studies to refine the parcellation of pMFC subregions and clarify their roles in error monitoring and behavioral adjustments^55,56.

Second, there is an inherent difference between SSD and SRI: SSD is a task-defined, externally controlled variable, whereas SRI depends on participants’ response times in incorrect stop trials, introducing greater individual variability. As a result, detecting SRI-related effects may be more challenging also due to the limited temporal resolution of 3T MRI. Consequently, brain responses may be more strongly linked to SSD, while SRI-related effects could be obscured. Additionally, we cannot rule out the possibility that SRI-related processes generate weaker neural signals than those associated with SSD.

Third, we used an unconventional approach to fMRI BOLD signal analysis to align with a previous EEG study³², which may have influenced our findings. To address the preSMA/SFG region’s potential link to inhibitory control in the stop-signal^57–59, we subtracted activity during correct stop trials from incorrect ones to more accurately capture error-related activity. Additionally, we calculated BOLD signal amplitude by comparing averaged activity within a selected time window after the stop signal to a general baseline (average activity across the entire run). We chose this approach to better control the BOLD time series trajectory compared to standard models. However, future research should also consider other methodological and analytical approaches.

Taken together, our findings suggest that while both conflict intensity and error expectancy contribute to behavioral post-error slowing, only conflict intensity is reflected in error-related neural activity in the stop-signal task. This dissociation supports the notion that these two aspects of performance monitoring involve distinct neural mechanisms. While reinforcement learning models posit that the brain updates expectations based on the magnitude of prediction errors, our findings suggest that this process may not be directly reflected in BOLD responses within the regions examined here. Instead, our results emphasize the role of response conflict in driving neural activity following an inhibitory error, reinforcing the notion that conflict detection plays a primary role in inhibitory error monitoring at the neural level.

Methods

Participants

The present study is based on a dataset collected as part of a larger multi-task fMRI project involving fifty young, healthy adults. All participants were right-handed and reported no history of psychiatric or neurological disorders. Of these, 44 participants (24 female; mean age = 22.05; standard deviation = 1.52) completed the stop signal task. One participant was excluded from behavioral and fMRI analyses due to low accuracy on go trials (below 10%).

Procedure

The study was approved by the Research Ethics Committee at the Institute of Psychology of Jagiellonian University in Kraków, Poland, and followed the standards of the Declaration of Helsinki. Participants were recruited through social networking websites to take part in a project on self-control, which involved two experimental sessions: one comprising a series of behavioral tasks and questionnaires, and the other involving fMRI scanning. The latter session included three runs of the n-back task, interleaved with three types of structural scans, followed by two runs of the stop signal task. For the purposes of the present study, we focused exclusively on the fMRI data from the stop-signal task and the T1-weighted structural images, as the other MRI data were not relevant to our research question. All participants signed an informed consent form and received compensation equivalent to 25 Euros for their time spent in the laboratory.

Stop-signal task

Participants completed two consecutive runs of the stop-signal task, each consisting of 160 trials. The go stimulus was the letter “X” or “O”. Participants were instructed to press a button on the left if the presented letter was “X” and a button on the right if the letter was “O”. Each trial lasted 1.5 s. In 40 trials, a stop signal (indicated by a red frame around the go stimulus) was presented with a variable delay following the onset of the go stimulus. Participants were instructed to withhold their response upon perceiving the stop signal. The SSD for the first stop trial was set to 200 ms. In subsequent stop trials, the SSD was adjusted based on individual performance: following successful stop trials, the SSD for the next stop trial was increased by 90%, while following failed stop trials, the SSD was decreased by the same rate. The minimum SSD value was 17 ms, with a minimum change of 20 ms. The inter-trial interval was jittered between 3 and 5 s. This duration was chosen to ensure adequate separation of the hemodynamic responses to consecutive events in the event-related fMRI design and to reduce collinearity between regressors. Within the task, it was possible for two or three trials with stop signals to be presented consecutively. Importantly, consecutive stop trials were infrequent: within each run (160 trials), two or three stop trials in a row appeared at most five times, corresponding to approximately 3% of all trials per run. Such a low rate of consecutive stop trials is unlikely to compromise the unexpectedness of the stop signal, which is a key requirement of the task. In addition, participants were not provided with any explicit information about the likelihood or structure of stop trials, including the possibility of consecutive stop trials. Participants completed a short practice session inside the scanner, consisting of 20 trials (8 of which included a stop signal). An overview of the task is provided in Fig. 3A.

Fig. 3.

General schema of the stop signal task (panel A) and illustration of inhibitory errors varying in the degree of their expectancy (panel B). (A) The go stimulus, either a letter “X” or “O”, indicated that the participant should make a left-hand or right-hand response. The go stimulus remained on the screen for 1.5 s. In 25% of all trials a red frame was presented around the go stimulus with a certain delay. The frame acted as a stop signal and indicated that the participant should inhibit the planned response. The stop-signal delay between go and stop signals was adaptively adjusted. The remaining time between trials occupied a range of 3–5 s. The left subpanel shows a go trial without a stop signal; the middle subpanel shows a successfully inhibited stop trial; the right subpanel illustrates a failed stop trial. (B) Incorrect responses that occur a longer time after the stop-signal presentation are more unexpected, compared with those that occur a shorter time after the stop-signal appearance.

Behavioral analysis

We calculated standard behavioral measures for the stop-signal task, including the percent of correct go trials, percent of incorrect go trials, percent of correct stop trials, mean reaction times in go trials, mean SSD, and stop-signal reaction time (SSRT), computed using the integration method^47,48. Importantly, SSRT was calculated for descriptive purposes only and was not used as an outcome measure, given that the study targeted error monitoring rather than inhibitory control. Consequently, participants were not excluded on the basis of stop accuracy, allowing us to maintain the full sample for the fMRI analyses. We also calculated mean post-error slowing, defined as the difference between the mean reaction time of correct go trials following incorrect stop trials and the mean reaction time of correct go trials preceding incorrect stop trials.

Importantly, to address our research question, we estimated three additional measures on a trial-wise basis and analyzed them using mixed-effects linear models:

(1) SSD in incorrect (uninhibited) stop trials was used as the predictor to reflect response conflict intensity, based on the conflict theory of error monitoring³. According to Stahl and Gibbons¹⁰ computational model, response conflict in the stop-signal task arises from the simultaneous processing of go and stop responses, both influencing the same action. The intensity and timing of this conflict are modulated by SSD. Specifically, longer SSDs lead to greater and more extended go-response activation compared to shorter SSDs, resulting in higher peak response conflict (due to the interaction of go and stop activations). Consequently, longer SSDs, associated with more intense response conflict, are expected to elicit stronger error monitoring brain activity. Consistent with this, Stahl and Gibbons¹⁰ observed that SSD is related to ERN amplitude, with longer SSDs corresponding to larger (more negative) ERNs.

(2) SRI in incorrect (uninhibited) stop trials was used as the predictor to reflect error expectancy, based on models rooted in reinforcement learning theory^4–7. SRI is calculated at the single-trial level as the difference between the reaction time to the go stimulus and the time of presentation of the stop signal, following the equation: SRI = RT (trial) – SSD (trial). Importantly, since SRI is shaped by moment-to-moment fluctuations in attention, motor preparation, and response speed, it captures internal variability in performance that is only partly constrained by the adaptive algorithm and SSD. Longer SRIs indicate a greater chance that the error could have been avoided (for an illustration, see Fig. 3B). The probability of avoiding an error relates directly to error likelihood prediction: the longer the SRI, the greater the discrepancy between the predicted and actual response outcome. Consequently, longer SRIs, associated with more unexpected errors, are hypothesized to elicit stronger error monitoring brain activity through the loops between the basal ganglia and medial frontal cortex region⁷. Consistent with this, Senderecka and Szewczyk³² found that longer SRIs are associated with larger (more negative) ERNs. It is worth noting that all uninhibited stop trials were retained for SRI computation, including those in which the response preceded the stop signal, since negative SRI values reflect theoretically important cases of unavoidable errors. In addition, this ensured consistency with a previous study³².

(3) Post-error slowing, a measure of the tendency to slow down after making an error, was used as the dependent variable to examine the relationship between response conflict (reflected in SSD), error expectancy (indexed by SRI), and post-error behavioral adjustment. Post-error slowing is commonly assumed to reflect either adaptive cognitive control or the automatic orientation of attention toward a significant event (for a review, see^17,18. Regardless of the exact mechanism, both more intense response conflicts and larger outcome prediction errors may hypothetically lead to more pronounced behavioral adjustments. We included this variable in the analysis as an additional test to disentangle the effects of conflict detection and error expectancy at the behavioral level. As an index of post-error slowing in our model, we used reaction times in correct go trials immediately following (1) a single incorrect stop trial or (2) a series (two or three) of consecutive incorrect stop trials. Thus, only sequences in which all stop trials were incorrect were eligible for inclusion.

The two predictors were examined in relation to error-related changes in the BOLD signal, described in the following sections.

fMRI acquisition parameters

The data were acquired using a 3T GE Discovery MR 750 MRI scanner. Functional scanning was performed with an 18-channel head/neck phased array coil using the EPI BOLD sequence with the following parameters: TR = 1800 ms, TE = 18 ms, FA = 90 degrees, bandwidth 7782 Hz/px, 64 × 64 matrix, 192 × 192 mm FOV, 3-mm slice thickness, 49 axial slices without a gap. For structural images, a T1-weighted image with an FSPGR BRAVO sequence was acquired with the following parameters: TR = 7.87 ms, TE = 3.04 ms, IT = 450 ms, FA = 12 degrees, 256 × 256 matrix, 240 × 240 mm FoV, 148 axial slices, 1 mm slice thickness, and no slice gap. The experimental setup used the Presentation software package from Neurobehavioral Systems (Berkeley, CA, USA) for managing stimulus presentation and recording participant responses. Stimuli were displayed via NeuroNordicLab (NNL) Visual System goggles (Bergen, Norway), which were additionally utilized for correcting participants’ vision. Participants provided responses using an NNL ResponseGrip response device with their right hand.

Preprocessing

Preprocessing and most of the fMRI analysis were performed using SPM12 software⁶⁰. Anatomical files were skull-stripped by segmenting structural T1 images into gray matter, white matter, and cerebrospinal fluid using tissue probability maps, and these segmented images were used for skull-stripping the anatomical scans. First, slice timing correction was performed. Then, the realignment to the mean functional image was applied using six-rigid-body-transformations. For the anatomical image, co-registration to the mean functional image was performed using normalized mutual information cost function. The co-registered anatomical image was segmented into gray matter, white matter, and cerebrospinal fluid using tissue probability maps. Additionally, this image was used to create forward deformation fields used for normalizing functional scans into MNI152 2 mm space. Following this stage, the functional data were smoothed with a 4-mm full-width at half maximum (FWHM) Gaussian kernel. A high pass filter (cutoff: 1/128 Hz) was applied to the data using the 3dBandpass function from AFNI software⁶¹.

Analysis overview

First, we contrasted activity between incorrect (uninhibited) stop trials and correct go trials to identify the network responsible for processing erroneous responses, independent of motor execution and go-related cognitive processes. This first-level contrast is commonly used in stop-signal research to isolate broad error-related activity while removing the contribution of motor responses (e.g.,⁶²). Next, from the clusters identified in this contrast, we extracted time series for incorrect stop trials and subtracted the activity of correct (inhibited) stop trials matched for SSD. Specifically, for each incorrect stop trial, we identified the correct stop trial(s) from the same participant with the identical SSD and averaged their time series within each ROI. This one-to-one SSD matching ensured that the temporal position of the stop signal – and therefore the amount of stop-related processing – was comparable across trial types. This logic parallels the subtraction strategy described by Rubia et al.⁶², who removed activity associated with successful inhibition to isolate the neural processes uniquely linked to failed inhibition. By subtracting correct-stop activity, we removed variance attributable to perceptual processing of the stop signal, attentional allocation to infrequent stop trials, and proactive inhibitory mechanisms, thereby isolating the activity specifically associated with error processing. This approach enabled a more accurate estimation of the hemodynamic response function (HRF) reflecting error-related processing. Finally, we examined how the amplitude of this error-related signal varied as a function of SSD and SRI. A general overview is presented in Fig. 4, and full methodological details are provided below.

Fig. 4.

General overview of the analysis workflow.

GLM analysis

After preprocessing, the data were modeled using a GLM with four event types: (1) presentation of the go signal, (2) correct button presses in go trials (go correct), (3) presentation of the stop signal in correctly inhibited trials (stop correct), and (4) presentation of the stop signal in uninhibited trials (stop incorrect). Event durations were set to 0. Reaction times were added as parametric modulators only for stop-incorrect trials. This choice reflects the fact that reaction times in uninhibited stop trials directly determine SRI, i.e., the key expectancy-related predictor in our analyses. We did not apply parametric modulation to go trials because modeling go responses at the time of the button press already incorporates the relevant reaction-time variability into the design matrix. Stop-correct trials were also not parametrically modulated, as they involve no executed response and therefore contain no meaningful reaction time-related variance. All events were convolved with a canonical hemodynamic response function (HRF) with temporal and dispersion derivatives. After estimating the coefficients, a second-level analysis was performed using the contrast of interest: stop incorrect (uninhibited) > go correct. This analysis was conducted using a t-test with FWE correction for multiple comparisons.

Time series analysis and linear mixed models

For further analysis, we extracted time series from the clusters identified in the GLM contrast of interest. This procedure followed established approaches in previous fMRI studies (e.g.,^63–65). Within each cluster, we examined voxelwise BOLD time series over a window spanning 4 s before to 14 s after the stop signal. For each incorrect stop trial, the corresponding matched correct stop trial(s) with identical SSD were identified within the same participant and run. Their time series were averaged and subtracted from the incorrect-stop waveform to remove variance associated with stop-signal processing and inhibitory control. As in Rubia et al.⁶², this subtraction isolates the unique neural contribution of failed inhibition by controlling for sensory, attentional, and inhibitory components shared across stop trials.

Then, for each participant, from the obtained wave representing error processing activity, we averaged the BOLD activity between 5th and 7th seconds after the stop signal. This time window was chosen because numerous fMRI studies have reported that the hemodynamic response to errors peaks at approximately 6 s post-stimulus^60,66,67. We extended the window symmetrically around this expected peak to accommodate individual variability in HRF latency. The group-averaged time courses (Fig. 2B–C) further confirmed that the peak error-related response in our sample fell within this interval.

Within this time window we tested the difference between waveforms for incorrect and correct stop trials using a repeated-measures t-test. We then compared this value to the average BOLD signal across the entire run and treated the resulting difference as the signal amplitude. Subsequently, we used this amplitude as the dependent variable in a linear mixed-effects model, with SSD and SRI as fixed-effect predictors and participant as a random effect. The model was estimated using the restricted maximum likelihood (REML) method. Before model estimation, we excluded observations in which SRI was longer than 900 ms and SSD was longer than 600 ms.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

KB: Conceptualization, Data curation, Formal Analysis, Methodology, Visualization, Writing – original draft, SW: Conceptualization, Funding Acquisition, Project Administration, Supervision, Writing – review & editing, EN: Conceptualization, Data Curation, Funding acquisition, Resources, Writing – review & editing, MS: Conceptualization, Data curation, Funding Acquisition, Investigation, Methodology, Project Administration, Supervision, Writing – original draft, Writing – review & editing.

Data availability

All raw data and the code needed to reproduce all results are publicly available following the link: https://osf.io/aum96/.

Declarations

Competing interests

The authors declare no competing interests.