Abstract
Errors can elicit post-error adjustments that serve to optimize performance by preventing further errors. An essential but unsolved issue is that whether post-error adjustments are domain-general or domain-specific, which was investigated in the present study through eliciting different types of errors. Behavioral and electrophysiological data were recorded when male and female subjects performed the Eriksen flanker task. For this study, we examined the aforementioned issue by combining event-related potential and multivariate pattern analysis. The results indicated that post-error slowing, error-related negativity, and error positivity were comparable between congruent and incongruent errors, indicating that errors triggered domain-general interference mechanisms, whereas post-error accuracy and late positive potential elicited by incongruent errors were larger than those elicited by congruent errors, exhibiting domain-specific control adjustment mechanisms. Importantly, no successful decoding soon after errors was found between congruent and incongruent errors, but above-chance decoding was observed between these two types of errors with increasing time, which further support that domain-general adjustments occurred in the early stage, whereas domain-specific adjustments appeared in the late stage. Furthermore, brain-behavior correlation results suggested that the late post-error adjustments predicted subsequent behavior performance. Together, this study revealed that early domain-general interference adjustments induced by errors are reflected in error detection and error awareness, which are independent of error types; on the contrary, late domain-specific control adjustments are reflected in attentional adjustments, which are modulated by error types.
SIGNIFICANCE STATEMENT To date, clear evidence on the specificity of post-error adjustments is lacking. The present study provides neurophysiological evidence that post-error adjustments simultaneously rely on both domain-general and domain-specific mechanisms. Event-related potential results indicated that domain-general adjustments were accompanied by the interference of error detection and error awareness. In contrast, domain-specific adjustments were associated with attentional adjustments. Multivariate pattern analysis further decoded the two features of post-error adjustments in the early stage matching the time patterns of error-related negativity and error positivity and in the late stage corresponding to the late positive potential. Temporal generalization analysis showed that domain-specific processing appeared stably in late post-error adjustments. Hence, we propose that post-error different stages may determine the specificity of post-error adjustments.
Introduction
Error commission usually induces the necessary adjustments to avoid repeating errors, which are termed “post-error adjustments.” Recent studies have hypothesized that post-error adjustments include general adjustments reflected in reduced sensitivity (i.e., the decrease of perceptual sensitivity to task-relevant sensory information) and increased decision boundary (i.e., the raise of response threshold or motor inhibition process) universally after errors, and specific adjustments manifested in increased task-related selective attention with time (Ullsperger and Danielmeier, 2016). Obviously, these suggest that both domain-general and domain-specific adjustments may occur, but the available evidence is rather weak. This study aimed to test these hypotheses by exploiting behavioral and neural approaches.
Several studies have attempted to clarify the nature of error-induced adjustments; however, the limitations of existing evidence need to be considered. Purcell and Kiani (2016) proposed that errors elicited general orienting adjustments with short response–stimulus intervals (RSIs) but specific control adjustments at long RSIs. Steinhauser et al. (2017) designed a dual-task paradigm in which the color task and tone task separated by a variable stimulus onset asynchrony (SOA), and reported that errors in the color task caused post-error slowing (PES) both in the tone task of the same trial and the color task of the next trial. The former effect disappeared with an increasing SOA, reflecting that errors induced transient, task-unspecific adjustments; whereas the latter effect was independent of SOA and lasted for several trials, suggesting more long-lasting, task-specific adjustments. Clearly, several limitations should be concerned. First, the length of RSIs or SOA determines the specificity of adjustments after errors but affects its integrality because of the different available times after errors. Second, in the study conducted by Steinhauser et al. (2017), error sources in the two tasks may be confused by additional factors (e.g., response criterion and task-related attention states), thus disturbing subsequent adjustments. Third, the abovementioned studies only infer the features of adjustments after errors based on behavioral evidence, rather than providing more direct neural evidence.
Considering these limitations, we manipulated two types of errors in the same flanker task by Eriksen and Eriksen (1974), namely, the errors on congruent trials (congruent errors) and those on incongruent trials (incongruent errors). This design was adopted for three reasons. First, it could eliminate the confounding effect of additional factors by adopting two types of errors elicited under the same RSIs in the same task. Second, in error-related paradigms, they have broad representation as the most basic and typical types of errors (Beatty et al., 2021). Third, they are induced by two types of distinct and stable sources. Congruent errors are mainly caused by insufficient attention to task-related information, such as speed pressure or unspecific noise, whereas incongruent errors are mainly because of interferences evoked by task-irrelevant information.
Event-related potential (ERP) methods provide important evidence on post-error adjustments in which error-related negativity (ERN) and error positivity (Pe) have been consistently reported (Ullsperger et al., 2014). ERN may be associated with error detection, whereas Pe is likely related to error awareness. Moreover, effortful attentional allocation processing affects the late positive potential (LPP), which is associated with attentional adjustments after errors (Hajcak et al., 2010). Multivariate pattern analysis (MVPA) is relatively sensitive in decoding the temporal course of the specificity evoked by errors (Fahrenfort et al., 2018), and supports for examining the temporal stability of decoding performance (King and Dehaene, 2014). Therefore, ERP and MVPA methods may provide valuable data for clarifying the features of error-induced adjustments.
In the present investigation, post-congruent and post-incongruent error behaviors and EEG responses were compared in a flanker task. Our central hypothesis was that post-error adjustments are domain-general in the early stage but domain-specific in the late stage. Accordingly, comparable PES, ERN, Pe, and around-chance decoding between congruent and incongruent errors should be observed because of interference universally following errors. In contrast, we should observe divergent post-error accuracy (PEA), LPP, and above-chance decoding between them because of task-related control adjustments.
Materials and Methods
Subjects
A total of 32 healthy volunteers (18 females; mean age: 20.44 ± 1.48 years, age range: 18-25 years) participated in the present study. They were right-handed, had normal or corrected-to-normal vision, and had no history of neurologic or psychiatric disease. They signed informed consent before the experiment. All subjects received ¥70 (∼$10.5) for compensation at the end of the experiment, no matter how they performed in the experiment. This study was conducted in accordance with the principles of the Declaration of Helsinki and its later amendments and was approved by the local Human Ethics Committee for Human Research.
Procedure and task design
The subjects were instructed to complete the experiment in a soundproof room. Post-error adjustments between different types of errors were examined using the Eriksen flanker task (Eriksen and Eriksen, 1974). The experimental task was performed using a program designed with E-Prime 2.0 (Psychology Software Tools). The experimental program was run on a 19 inch Dell monitor with a refresh rate of 85 Hz and a resolution of 1024 × 768. The experimental stimuli were presented on the central screen.
At the beginning of the flanker task, an instruction screen was presented, in which subjects were instructed to remember the stimulus–response mapping rules and respond to the targets as quickly and accurately as possible. Each trial began with the appearance of a white fixation point for 150 ms, followed by five white arrows on a black background for 100 ms (Fig. 1). The direction of the middle arrow was “<“ or “>,” corresponding to the response button “F” with the left index finger or “J” with the right index finger. Subjects had to respond to the direction of the middle arrow and ignore the flankers on each side before a 1500 ms response deadline. The congruent trial included flankers facing the same direction as the middle arrow (e.g., < < < < < or > > > > >), whereas the incongruent trial included flankers facing the opposite direction as the middle arrow (e.g., < < > < < or > > < > >). Once a response was given, the next trial began after an intertrial interval of 800 ms. Each block contained 96 trials, with 50% congruent trials and 50% incongruent trials presented randomly in the block. The feedback of response accuracy was provided on the screen after each block during practice session; however, no feedback was given during the formal experiment. The subjects first performed two practice blocks of 96 trials. They were allowed to participate in the formal experiment when the average error rate was <15%. The formal experiment consisted of 20 blocks with a total of 1920 trials. The entire experiment lasted for ∼1 h.
Task design. The sequence of a typical trial in the flanker task.
Experimental design and statistical analyses
Behavioral analysis
For all analyses, trials in which response times (RTs) for correct responses in congruent trials (congruent corrects), correct responses in incongruent trials (incongruent corrects), congruent errors, and incongruent errors that deviated >3 SDs away from the condition mean were removed. The RT and accuracy were computed for each condition. PES was calculated as the difference between the RT on correct trials following errors and that on correct trials following correct responses. PEA was calculated as the difference between the accuracy of trials following errors and that of trials following correct responses (Rabbitt, 1966; Danielmeier and Ullsperger, 2011; Wang et al., 2015; Li et al., 2020). The RT and accuracy were analyzed using repeated-measures ANOVA, with within-subject factors of previous congruency (congruent, incongruent) and previous response type (correct response, error response), respectively. The unit for the RT was milliseconds.
EEG recording and preprocessing
EEG data were collected using standard 64 in-cap Ag/AgCl electrodes following the extended international 10–20 system (Brain Products), and two additional electrodes were placed over the left and right mastoids. Additionally, vertical and horizontal electro-oculograms were recorded from below the left eye and over the outer canthus of the right eye. During data acquisition, all electrodes were referenced to the electrode FCz. EEG data were continuously recorded at a sampling rate of 500 Hz and online filtered with a 0.1-100 Hz bandpass filter. The impedance of all electrodes was maintained <5 kΩ throughout the recording process.
After data acquisition, EEG data preprocessing was performed using EEGLAB version 14.1.1 and MATLAB R2014b. Offline data were referenced to the mean of the left and right mastoids. Data were filtered with 0.1-30 Hz bandpass filter using a basic finite impulse response filter. Continuous EEG data were segmented from −200 to 1000 ms using epochs locked on the response markers. The segmented data were baseline-corrected using a −200 to −100 ms baseline window. For the removal of electro-oculograms artifacts, an eye-movement correction program based on the linear regression method was run using the Automatic Artifact Removal toolbox (Gratton et al., 1983). Automated rejection of epochs was performed whenever the voltage exceeded 100 μV. The results indicated an average of 893 congruent correct trials, 784 incongruent correct trials, 54 congruent error trials, and 155 incongruent error trials for the statistical analyses of EEG data.
ERP analysis
In light of previous studies and grand mean mapping, response-locked ERN was calculated using a time window of 80 ms (−20 to 60 ms) at fronto-central electrodes (Fz, FCz, F1, F2). Response-locked Pe was scored as the mean amplitude at centro-parietal electrodes (CPz, Pz, CP1, CP2, P1, P2) during a time window of 250 ms (150-400 ms). Response-locked LPP was computed using a time window of 300 ms (700-1000 ms) at parieto-occipital electrodes (Pz, POz, Oz, PO4, O2). These ERP components were analyzed using repeated-measures ANOVA, with within-subject factors of current congruency (congruent, incongruent) and current response type (correct response, error response). The unit for ERP amplitude was µV.
MVPA
Considering the higher sensitivity of multivariate analyses in decoding higher-order cognitive processes, MVPA was applied to the preprocessed EEG data. Trials were classified into four trial types according to current congruency (i.e., congruent and incongruent) and current response type (i.e., correct response and error response). MVPA involved a backward decoding classification algorithm (linear discriminant analysis), with two of the four trial types as classes (i.e., incongruent errors and congruent errors, incongruent errors and incongruent corrects, congruent errors and congruent corrects, incongruent corrects and congruent corrects) and all electrodes as features. The analysis aimed to examine whether the classifier could learn from divergent EEG patterns following each single-trial response to distinguish between different types of errors.
The Amsterdam Decoding and Modeling toolbox (Fahrenfort et al., 2018) was used for the preprocessed EEG data. Before MVPA, the EEG data were undersampled to 50 Hz owing to the time-consuming decoding algorithms. A classifier was trained and tested at each time point by using a five fold cross-validation procedure; that is, the trials were sequentially randomized and split into five equal-sized folds. Subsequently, a leave-one-out procedure was performed (i.e., the classifier was trained on four folds and tested on the remaining fold). This procedure was repeated 5 times until all data were tested. Classifier performance was determined using the mean of test folds. As a measure of classification accuracy, the area under the curve (AUC) per time point was used. Larger AUC values indexed higher classification accuracy (Fahrenfort et al., 2018). To avoid the subtle biases of the next trial correctness on the classification performance, all classification analyses were limited to the trials in which the subsequent trial was correct. Within-class balancing using undersampling was applied for the next trial congruency (i.e., the trials were randomly selected from the conditions with more trials to balance the conditions with fewer trials), such that the count of subsequent congruent and subsequent incongruent trials was the same within each class. This procedure ensured that the training of the classifier was not biased by the next trial congruency. Considering the potential different estimations that may result from the unbalanced design (asymmetric trial counts), between-class balancing using undersampling (i.e., a random selection of trials from the conditions with more trials) was applied to ensure that the classifier would not develop a bias for the overrepresented class during training. Moreover, to rule out the subtle effects of the next trial RT on the classification performance, for incongruent errors and congruent errors, and incongruent corrects and congruent corrects, the procedure randomly selected trials in the conditions with more trials to match the next trial RT in the conditions with fewer trials (i.e., paired sample t tests between the two classes showed no significant differences in the next trial RT, p > 0.05). As a result, the following trial counts entered the analyses for each classification (mean given): incongruent errors and congruent errors (39 trials per condition, 78 trials in total), incongruent errors and incongruent corrects (132 trials per condition, 264 trials in total), congruent errors and congruent corrects (39 trials per condition, 78 trials in total), and incongruent corrects and congruent corrects (657 trials per condition, 1314 trials in total). Furthermore, a temporal generalization analysis using classification across time was performed (King and Dehaene, 2014). This analysis was used to evaluate the stability of neural activity patterns with underlying classification performance by testing the classifier at the same time point on which it was trained. After that, temporal generalization matrices for time-by-time classification accuracy were obtained. As a result, the classification accuracy outside the diagonal axis was the above-chance level, suggesting stable neural activity. AUC values were calculated at each time point with double-sided t tests across subjects against a 50% chance level. Cluster-based permutation tests (p < 0.05, 1000 iterations) were then used to perform multiple-comparison correction for these t tests over time. According to the observed cluster size, the null distribution of the cluster size under random permutation was determined, and the p values for the clusters were calculated using this comparison (Fahrenfort et al., 2018).
Brain-behavior correlation analysis
Single-trial ERP-behavior correlation analysis
To test whether error-related processing (ERN, Pe, LPP) could predict post-error behavior (PES, PEA) on a single-trial level, similar to Buzzell et al. (2017), we performed generalized linear mixed-effects analyses using the R statistical software, version 4.1.3 (R Core Team, 2016), combining with the lme4 package, version 1.1-29 (Bates et al., 2015) and the lmerTest package, version 3.1-3 (Kuznetsova et al., 2016). These analyses were not planned a priori and should therefore be considered exploratory. Since statistical analyses were run to explore the specificity of error-induced adjustments based on different types of errors, we conducted the single-trial correlation analyses only on congruent errors and incongruent errors.
Before carrying out each analysis, models were built such that the continuous variables (ERN amplitude, Pe amplitude, LPP amplitude, next trial RT) were centered and scaled to have a mean of 0 and SD of 1 across the dataset, and the category variables (current congruency, next trial accuracy) were entered using sum and contrast. The continuous variables were fitted using linear mixed-effects analysis using the lmer function, with restricted maximum likelihood estimation. And the categorical variables were fitted using generalized linear mixed-effects models using the glmer function with logit link with maximum likelihood estimation.
Within each model, the variation in the within-subjects intercept was treated as a random effect, and the effects of interest and their interactions (plus an intercept) were treated as fixed effects. We adopted the lmerTest, using the Satterthwaite's approximation to denominator degrees of freedom, to calculate the statistical significance for each fixed effect (Kuznetsova et al., 2016). We defined the mixed-effects models with the following formula:
Here, Y represents the response variable, X is the fixed effect design matrix, β is the fixed effect coefficient, Z indicates the random effect design matrix, γ is the random effect coefficient, and ε represents the error term.
We used the syntax of the R Package lme4 to construct the mixed-effects models as follows:
This syntax represents a model with a fixed effect on the whole model intercept (the initial '1'), fixed effects on all independent variables of interest and their interactions, and a random effect on the variation in intercept per participant ('1|Subject').
To examine the effect of error-related ERN, Pe, or LPP and current congruency (congruent vs incongruent) on subsequent behavior, we constructed separate models for next trial RT and next trial accuracy. Considering our interest in post-error adjustments, we limited the types of trials in all models. When predicting the next trial RT, we limited the pairs of trials in which the subsequent trial was correct. When predicting the next trial accuracy, we limited the trials in which the current trial was error.
Across-subject AUC-behavior correlation analysis
To further investigate the behavioral relevance of significant classification of error types, we tested how classification performance correlated to PES and PEA on the response to the two types of errors. These correlation analyses were not planned a priori. Therefore, these results should be considered exploratory to better characterize the robust results derived from the classification analyses, and should enrich our understanding of the data. First, we locked the time window of significant above-chance classifications between congruent errors and incongruent errors (i.e., the time window in which AUC was significant at p < 0.05), and calculated the mean AUC values for each subject within this time window. Next, two across-subject Pearson's correlation analyses were run between the individual AUC values and PES, and between individual AUC values and PEA. Through this procedure, we thus linked significant classification of error types from neural activity to subsequent behavioral performance.
Availability of materials and data
The datasets analyzed during the current study are available from the corresponding author on reasonable request.
Results
Behavioral results
With respect to behavioral data, the trials in which the RT deviated >3 SDs away from the condition mean were excluded. A total of 12.56% of trials were excluded. The overall accuracy was 88.44% (SE = 0.73%), and the mean RT was 401.31 (SE = 4.93). The accuracy was 93.96% (SE = 0.61%) on congruent trials, and was 82.92% (SE = 1.02%) on incongruent trials.
As for the RT, ANOVA with previous congruency × previous response type revealed the main effect of previous response type (F(1,31) = 15.70, p < 0.001, η2 p = 0.34), in which the RTs on correct trials following errors (412.36, SE = 6.25) were significantly slower than those on correct trials following correct responses (399.68, SE = 5.00). However, the main effect of previous congruency (F(1,31) = 0.59, p = 0.448) and the two-way interaction (F(1,31) = 0.01, p = 0.944) did not reach significance (Fig. 2A,C).
Behavioral results in the flanker task. The overall reaction time (A) and accuracy (B) on trials following correct responses and errors. Individual PES (C) and PEA (D) for congruent and incongruent trials. Error bars indicate SEM. *p < 0.05. **p < 0.01. ***p < 0.001.
Regarding accuracy, ANOVA showed the main effect of previous congruency (F(1,31) = 25.10, p < 0.001, η2 p = 0.45) but failed to reveal the main effect of previous response type (F(1,31) = 0.30, p = 0.586). Notably, the interaction between previous congruency and previous response type was significant (F(1,31) = 7.72, p = 0.009, η2 p = 0.20). Post hoc tests showed that, for congruent trials, the accuracy did not differ between trials following errors (86.57%, SE = 1.35%) and trials following correct responses (87.50%, SE = 0.85%) (p = 0.478). In contrast, for incongruent trials, the accuracy was significantly higher for trials following errors (90.85%, SE = 0.64%) than for trials following correct responses (88.93%, SE = 0.84%) (p = 0.012), indicating that a significant post-error improvement in accuracy occurred only when the previous trial was incongruent (Fig. 2B,D).
ERP results
ERN
For the response-locked ERN, ANOVA with current congruency × current response type showed the main effect of current response type (F(1,31) = 18.03, p < 0.001, η2p = 0.37) but failed to reveal the main effect of current congruency (F(1,31) = 1.26, p = 0.271). Importantly, the interaction between current congruency and current response type was significant (F(1,31) = 9.60, p = 0.004, η2p = 0.24). Post hoc tests showed that, for errors, the ERN amplitude did not differ between congruent (−1.71, SE = 0.52) and incongruent trials (−1.97, SE = 0.55) (p = 0.149). On the other hand, for correct responses, the ERN amplitude evoked by congruent trials (0.35, SE = 0.43) was greater (more negative) than that evoked by incongruent trials (0.83, SE = 0.47) (p = 0.001), as shown in Figure 3A, B.
ERP results in the flanker task. The waveform, topography, and amplitude of ERN (A,B), Pe (C,D), and LPP (E,F) for correct responses and errors on congruent and incongruent trials. Shaded regions represent the defined time windows for ERN, Pe, and LPP, respectively. The topography distributions represent the average amplitude in time windows. Error bars indicate SEM. **p < 0.01. ***p < 0.001.
Pe
For the response-locked Pe, ANOVA revealed the main effect of current response type (F(1,31) = 14.37, p = 0.001, η2p = 0.32), in which the Pe amplitude for errors (1.90, SE = 0.51) was significantly greater than that for correct responses (0.17, SE = 0.41). However, neither the main effect of current congruency (F(1,31) = 1.48, p = 0.232) nor the two-way interaction (F(1,31) = 1.52, p = 0.227) was significant (Fig. 3C,D).
LPP
For the response-locked LPP, ANOVA showed the main effect of current response type (F(1,31) = 8.93, p = 0.005, η2p = 0.22) but failed to reveal the main effect of current congruency (F(1,31) = 0.06, p = 0.805). Importantly, the interaction between current congruency and current response type was significant (F(1,31) = 18.39, p < 0.001, η2p = 0.37). Post hoc tests showed that, for congruent trials, the LPP amplitude did not differ between errors (1.08, SE = 0.40) and correct responses (0.83, SE = 0.31) (p = 0.194). On the contrary, for incongruent trials, the LPP amplitude evoked by errors (1.41, SE = 0.39) was significantly greater than that evoked by correct responses (0.55, SE = 0.30) (p < 0.001). Moreover, for errors, the LPP amplitude induced by incongruent trials was greater than that induced by congruent trials (p = 0.028); however, for correct responses, the LPP amplitude evoked by incongruent trials was smaller than that evoked by congruent trials (p < 0.001) (Fig. 3E,F).
MVPA results
To determine the temporally extended pattern of neural activity after an error occurred, linear discriminant classifiers were trained at each time point to perform the following: (1) distinguish between incongruent errors and congruent errors; (2) distinguish between incongruent errors and incongruent corrects; (3) distinguish between congruent errors and congruent corrects; and (4) distinguish between incongruent corrects and congruent corrects (Fig. 4A). With respect to incongruent and congruent errors, MVPA revealed a significant above-chance difference between classes of incongruent and congruent errors from 700 to 1500 ms after errors occurred (the late stage after errors) (p < 0.05, cluster-corrected). Regarding incongruent errors and incongruent corrects as well as congruent errors and congruent corrects, MVPA showed a significant above-chance difference between the two classes from response initiation to 1500 ms after responses (the whole decoding stage after responses) (p < 0.05, cluster-corrected). As for incongruent and congruent corrects, MVPA revealed a significant above-chance difference between the two classes from response initiation to 680 ms after responses (the early stage after correct responses) (p < 0.05, cluster-corrected). These results suggest that, for correct responses, the two types of correct responses could be distinguished only in the early stage after responses. On the other hand, for error responses, the two types of errors could not be decoded in the early stage after errors but could be distinguished in the late stage after errors.
MVPA results in the flanker task. A, Classification accuracy for the four datasets. Thicker lines indicate the time windows of significant classification performance (p < 0.05, cluster-corrected). B, Time generalization matrix of classifier performance. Saturated colors represent significant samples. C, Maps of forward transformation weights averaged over the time windows of significant classification performance.
Because successful classification was observed in all four datasets, in the next step, the temporal generalization matrices were calculated to test the stability of neural activity patterns with underlying significant classification performance. The time-by-time generalization results are presented in Figure 4B. For incongruent and congruent errors, a cluster was detected at the above-chance level from 200 to 1500 ms after responses. That is, the time generalization matrix indicated that the neural activity from 200 to 1500 ms after responses could be decoded by classifiers trained during the time window from 700 to 1500 ms, suggesting that the differences between incongruent and congruent errors were maintained stably over time. Additionally, for incongruent errors and incongruent corrects as well as congruent errors and congruent corrects, the above-chance activities were detected from 0 to 1500 ms. As for incongruent and congruent corrects, significant above-chance activity showed stable neural activity pattern from 0 to 680 ms after responses.
During the time windows of significant classification performance, the topographical patterns of forward transformation weights with high spatial resolution revealed that the spatial activity patterns were different among different datasets, but there were no significant clusters of electrodes (Fig. 4C). This exhibited that the involvement of neural activity depended on a complex spatial pattern, rather than being bound to a specific brain region.
Brain-behavior correlations
Single-trial ERP-behavior correlations
The linear mixed-effects model of ERN and current congruency on the next trial RT revealed that neither the main effects (all p > 0.088) nor the interaction between ERN and current congruency was significant (p = 0.599). The model of Pe and current congruency on the next trial RT showed that all main effects (all p > 0.077) and interaction involving Pe failed to reach significance (p = 0.148). The model of LPP and current congruency on the next trial RT revealed that the main effects (all p > 0.090) and their interaction failed to reach significance (p = 0.158).
The linear mixed-effects model of ERN and current congruency on the next trial accuracy revealed an effect of current congruency (estimate = −0.045, SE = 0.009, z = −5.09, p < 0.001), in which incongruent errors induced higher next trial accuracy, with no main effect of ERN (p = 0.259) and no interaction between ERN and current congruency (p = 0.957). The model of Pe and current congruency on the next trial accuracy revealed an effect of current congruency (estimate = −0.045, SE = 0.009, z = −5.12, p < 0.001), suggesting that post-incongruent error accuracy was higher, with no significant main effect of Pe (p = 0.539) or interaction between Pe and current congruency (p = 0.771). The model of LPP and current congruency on the next trial accuracy showed a significant main effect of LPP (estimate = 0.033, SE = 0.011, z = 2.90, p = 0.004), suggesting that greater LPP was associated with a higher next trial accuracy, and an effect of current congruency (estimate = −0.045, SE = 0.009, z = −5.14, p < 0.001) such that incongruent errors were associated with a higher next trial accuracy. Critically, there was a significant interaction in which LPP amplitude predicted the next trial accuracy as a function of current congruency (estimate = −0.021, SE = 0.008, z = −2.52, p = 0.012; Fig. 5). And the interaction suggested that greater LPP amplitude predicted increased post-incongruent error accuracy (p = 0.009), but not post-congruent error accuracy (p = 0.238).
Correlations between single-trial LPP, next trial accuracy, and current congruency. Predicted values for the next trial accuracy, relative to LPP and current congruency (congruent vs incongruent). The LPP interacts with current congruency, such that the influence of LPP amplitude on the next trial accuracy differs between congruent errors and incongruent errors. Shaded region around each line represents SEM.
Across-subject AUC-behavior correlations
The correlation analysis showed that the mean AUC values between congruent errors and incongruent errors over the significant interval of 700-1500 ms after responses did not correlate with PES (r = −0.192, p = 0.292; Fig. 6A). Importantly, there was a significant positive correlation between mean AUC values and PEA (r = 0.387, p = 0.028; Fig. 6B), suggesting that higher classification accuracy between incongruent errors and congruent errors predicted better subsequent behavior performance.
Correlations between across-subject classification accuracy and post-error behavior. A, Correlations between AUC values and PES. B, Correlations between AUC values and PEA. Dots represent single-subject values. Diagonal lines indicate least-square fits.
Discussion
The present study investigated the specificity of post-error adjustments by examining behavioral and neural effects across congruent and incongruent errors in the flanker task. Regarding the RT results, PES was present on both types of errors, which may be attributable to the interference triggered by a resource-consuming monitoring process (Jentzsch and Dudschig, 2009) or a distracting orienting response (Notebaert et al., 2009) immediately after errors, exhibiting domain-general mechanisms. Although PEA did not differ from post-correct accuracy on congruent trials, the former was higher than the latter on incongruent trials, which was attributable to stronger attentional adjustments after incongruent errors (Ullsperger et al., 2014; Steinhauser and Andersen, 2019), indicating domain-specific mechanisms. Thus, errors elicit both domain-general and domain-specific mechanisms. For ERP results, with regard to ERN and Pe, the contrasts between errors and correct responses followed the same pattern no matter on congruent or incongruent trials, and the two ERP amplitudes did not differ between congruent and incongruent errors. With respect to MVPA results, above-chance decoding was constantly detectable between errors and correct responses, regardless of congruent or incongruent trials; in contrast, significant decoding was not observed between congruent and incongruent errors in the early stage after responses. These EEG results indicate that different types of errors were analogous in early post-error adjustments, such as error detection (Coles et al., 2001) and error awareness (Overbeek et al., 2005), suggesting that domain-general processing was involved in the early stage. Importantly, the LPP amplitude induced by incongruent errors was greater than that induced by congruent errors. Moreover, MVPA yielded convergent results with the conventional ERP analysis and showed significant above-chance decoding between congruent and incongruent errors in the late stage after responses. Thus, different types of errors result in distinct neural responses in late post-error adjustments, indicating domain-specific processing.
ERN is believed to reflect error detection, which is related to the detection of a mismatch between the response that was made and the appropriate response (Coles et al., 2001). Pe is associated with error awareness (Nieuwenhuis et al., 2001; Overbeek et al., 2005), reflecting evidence accumulation to determine whether post-error adjustments occurred (Steinhauser and Yeung, 2010). The present results showed that the two types of errors make no difference in the basic error-related processes (error detection and error awareness). Accordingly, we propose that the comparable ERN and Pe results between congruent and incongruent errors reflect the domain-general mechanisms involved in error-induced adjustments, which should be independent of specific error types. However, not all error-related neural signatures led to this post-error adjustment. In addition to the two abovementioned ERP components, errors were also associated with the LPP, and this effect depended on the congruency of error trials. The LPP has been linked to the elaborate processing of human action in previous studies (Hajcak et al., 2010) and has been shown to represent selective attention ability in effortful attentional allocation and cognitive control (Miller et al., 2011; Mun et al., 2014). In this case, the significantly greater LPP amplitudes on incongruent errors may indicate that more attentional adjustments were implemented, which favors adaptive adjustments after incongruent errors. Accordingly, the divergent results for congruent and incongruent errors on LPP suggest that post-error adjustments induce domain-specific mechanisms, which are sensitive to error types.
Statistically, MVPA is more sensitive than the conventional ERP analysis. It combines with whole-brain activity to depict neural activity patterns over time (LaRocque et al., 2014; Stokes et al., 2015; van Ede et al., 2018; De Vries et al., 2019). The MVPA results of this study indicated that no reliable decoding was found for congruent and incongruent errors from 0 to 700 ms after responses, which is the time window corresponding to ERN and Pe. Nonetheless, above-chance decoding was observed for congruent and incongruent errors during the time window from 700 to 1500 ms, which matches the time pattern of the LPP. After successful decoding for the two types of errors based on their neural activity patterns, a further examination of neural activity is its stability over time (Takacs et al., 2020). The above-chance decoding was based on a more stable pattern of neural activity in the time window from 200 to 1500 ms than that during other windows, implying that stable difference patterns between the two types of errors appear in late post-error adjustments. These results potentially indicate that, from the time course of post-error adjustments, domain-general processing occurs in the early stage, including the two subprocesses (error detection and error awareness), whereas domain-specific processing appears stable in the late stage, involving attentional adjustments.
Critically, single-trial ERP-behavior correlations revealed that LPP amplitude, but not ERN or Pe amplitude, predicted next trial accuracy as a function of current congruency. Accordingly, increased post-incongruent error accuracy was predicted by increased LPP amplitude, implying that late error-related attentional adjustments (indexed by greater LPP) were directly involved in eliciting a higher likelihood of responding to the next trial correctly. This was consistent with an across-subject AUC-PEA positive correlation: subjects with higher classification accuracy between incongruent errors and congruent errors in the late stage performed higher PEA. These results suggest a brain-behavior link, with late post-error adjustments directly promoting subsequent behavior performance.
The supply of central resources seems crucial for the specificity of post-error adjustments. According to recent studies, both ERN and Pe would lead to an impairment in sensory processing on subsequent trials (Buzzell et al., 2017; Li et al., 2021). Hence, the two basic error-related processes (error detection and error awareness) are thought to place a high demand on central resources, which may leave insufficient central resources for subsequent processing, thereby leading to impaired performance after errors. Moreover, previous studies have reported that ERN is responsible for PES, but not PEA (Debener et al., 2005; West and Travers, 2008). Accordingly, for congruent and incongruent errors, the comparable ERN and Pe as well as the early unsuccessful decoding indicate that error detection and error awareness may consume a similar amount of resources between the two types of errors, which is consistent with the comparable PES between them. In this case, early error-induced adjustments would cause a disturbance, regardless of congruent or incongruent errors, exhibiting domain-general interference mechanisms. In contrast, adequate central resources seem to improve post-error performance (Li et al., 2021). In the present study, the greater LPP evoked by incongruent errors as well as the late successful decoding for the two types of errors showed that more central resources should be recruited by incongruent errors. Whereas incongruent errors still drove PES, indicating that the supplemented central resources in the late stage would not offset the PES caused by the consumed resources in the early stage. Indeed, from the brain-behavior correlations, more central resources would be supplied for incongruent errors in the late stage, which might be used to focus on target-related information on incongruent trials and thus improve the PEA, indicating domain-specific control adjustment mechanisms.
In addition, the LPP amplitude for incongruent corrects was smaller than that for congruent corrects. From the perspective of central resource consumption, conflict-related processing may consume a large amount of central resources during the early stage after incongruent corrects, indexed by the above-chance decoding for congruent and incongruent corrects in the time window centered at 0 and 680 ms after correct responses. Thus, the supply of resources may be reduced in the late stage, which should account for the attenuated LPP after incongruent corrects relative to congruent corrects.
Theoretically, the two-stage account of error evaluation proposed by Maier et al. (2011) involves an early evaluation stage during task processing and a late evaluation stage after the execution of a response. The early stage evaluates the state of system parameters “preceding” the response, and initiates source-specific adaptive adjustments “succeeding” the response (Maier and Steinhauser, 2013). In line with the results of Maier et al. (2011), we also identified that post-error behavioral adjustments matched the error types. Whereas the difference between our investigation and that of Maier et al. (2011) is that, after the errors occurred in the present study, both source-general and source-specific adjustments were triggered; in comparison, only the latter were mentioned in the study by Maier et al. (2011). Moreover, the MVPA results even sensitively revealed the time boundary under which source-specific adjustments were chosen (∼700 ms after errors).
In conclusion, the present study provides the first piece of neurophysiological evidence regarding the specificity of post-error adjustments, which was achieved by combining ERP methods with MVPA. The results indicated that post-error adjustments involved two mechanisms: First, in the early stage, the errors induced domain-general interference mechanisms, including error detection and error awareness, indexed by around-chance decoding as well as comparable ERN and Pe between congruent and incongruent errors. Second, in the late stage, the errors evoked domain-specific control adjustment mechanisms, including task-related attentional adjustments, indexed by above-chance decoding and distinct LPP between congruent and incongruent errors, which can directly improve performance on subsequent trials. Therefore, post-error adjustments involved domain-general and domain-specific adjustments at various stages.
Footnotes
This work was supported by both National Natural Science Foundation of China Grant 32171040 and Chongqing Research and Innovation Funds for Postgraduates CYB22096.
The authors declare no competing financial interests.
- Correspondence should be addressed to Antao Chen at chenantao{at}sus.edu.cn
