Neural signatures of adaptive post-error adjustments in visual search

NeuroImage 150 (2017) 270–278

Contents lists available at ScienceDirect

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

Neural signatures of adaptive post-error adjustments in visual search

MARK

⁎

Robert Steinhauser , Martin E. Maier, Marco Steinhauser Catholic University of Eichstätt-Ingolstadt, Germany

A R T I C L E I N F O

A BS T RAC T

Keywords: Post-error adjustments Selective attention Performance monitoring EEG Multivariate pattern analysis

Errors in speeded choice tasks can lead to post-error adjustments both on the behavioral and on the neural level. There is an ongoing debate whether such adjustments result from adaptive processes that serve to optimize performance or whether they reflect interference from error monitoring or attentional orientation. The present study aimed at identifying adaptive adjustments in a two-stage visual search task, in which participants had to select and subsequently identify a target stimulus presented to the left or right visual hemifield. Target selection and identification can be measured by two distinct event-related potentials, the N2pc and the SPCN. Using a decoder analysis based on multivariate pattern analysis, we were able to isolate the processing stages related to error sources and post-error adjustments. Whereas errors were linked to deviations in the N2pc and the SPCN, only for the N2pc we identified a post-error adjustment, which exhibits key features of source-specific adaptivity. While errors were associated with an increased N2pc, post-error adjustments consisted in an N2pc decrease. We interpret this as an adaptive adjustment of target selection to prevent errors due to disproportionate processing of the task-irrelevant target location. Our study thus provides evidence for adaptive posterror adjustments in visual search.

Introduction Human behavior is fallible. Even in the simplest cognitive tasks, errors can occur due to attentional lapses (Weissman et al., 2006), speeding, or failures of cognitive control (Steinhauser et al., 2012). In recent years, errors like these have been investigated to elucidate how the human brain can detect and learn from these errors. Although this research has identified an error monitoring system in the medial frontal cortex that rapidly detects and evaluates errors (Ridderinkhof et al., 2004), the cognitive and behavioral consequences of error monitoring are still unclear (Danielmeier and Ullsperger, 2011). Some studies demonstrated that errors lead to adaptive adjustments that aim to prevent further errors (e.g., Dutilh et al., 2011; Maier et al., 2011), whereas others suggested that errors primarily elicit nonadaptive adjustments that impair performance even further (e.g., Notebaert et al., 2009; Van der Borght et al., 2014). In the present study, we applied a visual search task that allows for distinguishing between two stages of selective attention – target selection and target identification. By measuring event-related potentials (ERPs) associated with each stage, we aimed to identify the specific processes at which errors and post-error adjustments occur, and to describe whether and how adjustments are related to the source of the error. Post-error adjustments are thought to be adaptive if they serve to improve performance by preventing further errors (Ridderinkhof et al.,

⁎

2004). Generally speaking, any adjustment that prevents further errors can be considered adaptive, and there are indeed studies that find evidence for adaptive post-error adjustments that improve performance independently of the source of the error (e.g., by compensating error-induced detriments through a general increase of cautiousness; (Purcell and Kiani, 2016)). However, most studies on adaptivity report adjustments that specifically seek to counteract the source of the error (Dutilh et al., 2012; Jentzsch and Leuthold, 2006; King et al., 2010; Maier et al., 2011; Steinhauser and Kiesel, 2011; Danielmeier et al., 2011). Such source-specific adaptation requires that the type of adjustment is directly linked to the error source and attempts to counteract the deviations in cognitive processes that lead to the error in the first place. A well-known example for source-specific adaptation is the variation in the response criterion. Errors in speeded choice tasks are often preceded by decreased response times (pre-error speeding) but followed by increased response times (post-error slowing, PES). Whereas pre-error speeding has been attributed to a low response criterion favoring speed over accuracy (Jentzsch and Leuthold, 2006; Brewer and Smith, 1989; Danielmeier et al., 2011), post-error slowing has been ascribed to a response criterion shift towards a more cautious response strategy in order to reduce this error source (Dutilh et al., 2012; Botvinick et al., 2001). Further examples are post-error adjustments of selective attention. Studies considering hemodynamic corre-

Correspondence to: Catholic University of Eichstätt-Ingolstadt, Germany, Fachbereich Psychologie, Ostenstraße 25, D-85072 Eichstätt, Germany. E-mail address: [email protected] (R. Steinhauser).

http://dx.doi.org/10.1016/j.neuroimage.2017.02.059 Received 22 September 2016; Accepted 21 February 2017 Available online 22 February 2017 1053-8119/ © 2017 Elsevier Inc. All rights reserved.

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

mulated, visual attention is allocated towards the target (target selection). If the task requires to subsequently classify the target, a further stage is involved in which relevant target features are processed in working memory (target identification). Although there is behavioral and neural evidence for the distinctness of these stages, they do not necessarily proceed in a strictly serial manner and partial overlapping in time is likely (Eimer, 2014a; Wolfe, 2007). Such a two-stage visual search task has two crucial advantages for our purpose: First, errors as well as post-error adjustments can emerge on either of these stages. This allows us to investigate whether post-error adjustments occur on the same stage that caused the error, i.e., whether error source and adjustment are directly linked. Second, the stage at which errors and post-error adjustments occur can easily be identified using ERPs because each stage is associated with a characteristic ERP component that we describe in the following. A neural correlate of target selection in visual search is the N2pc, a negativity emerging about 200 to 250 ms after stimulus onset on posterior electrodes contralateral to the hemifield at which the target is presented. Whereas it is generally believed that the N2pc represents the allocation of attention towards the target, it is up to debate whether it reflects the suppression of distractors (Luck and Hillyard, 1994) or the enhancement of the target (Mazza et al., 2009b). In visual search tasks that involve target classification, a later sustained posterior contralateral negativity (SPCN) is thought to represent the target identification stage (Eimer, 2014a; Mazza et al., 2007). Depending on the paradigm and the duration of the presented stimulus, it can be found in a time period of 300 ms to 800 ms post-stimulus. As this component is sensitive to working memory load (Jolicœur et al., 2008), it is argued that in fact the SPCN represents the storage of selective features of the target items (Woodman and Vogel, 2008). By considering these components in a paradigm that has previously been used to investigate ERP correlates of visual search (Mazza et al., 2009b; Mazza et al., 2007), we aimed to identify the exact time course of the employment of post-error adjustments as well as their relationship to the error source. In the paradigm at hand, participants viewed displays with twenty items (one red target and 19 green distractors) each being a diamond with a missing corner on the left or right side, respectively (Fig. 1). The red target was presented in the left or right hemifield. The participants’ task was to indicate whether the missing corner of the target was on the left or right side. Because participants first had to select the target and then analyze the target features, this task involves two stages of selective attention, target selection and target identification, which, on the neural level, are represented by the N2pc and the SPCN (Eimer, 2014a; Mazza et al., 2007). Another important feature of this paradigm is that the required response category (left/right) can be compatible or incompatible to the target location (left/right hemifield), thus leading to two conditions of stimulus-response compatibility (Simon and Rudell, 1967). As error sources and adjustments might differ across these conditions, we conducted all analyses separately for compatible and incompatible

lates of brain activity found that post-error trials were associated with decreased activity in brain regions linked to task-irrelevant stimulus features, if these regions had shown increased activity on the error trial and thus had formed a potential error source (King et al., 2010; Danielmeier et al., 2011). Moreover, it was shown that when different error types could occur within the same task, then post-error adjustments varied depending on the error type. Maier, Yeung, and Steinhauser (Maier et al., 2011) used a variant of the Eriksen flanker task (Eriksen and Eriksen, 1974) in which errors could occur because participants erroneously responded to the flanking distractors or because of speeding alone. Whereas both error types elicited posterror slowing, only erroneous responses to the distractor led to reduced distractor processing on the subsequent trial, which was interpreted as an adaptive adjustment to counteract the error source. Steinhauser and Kiesel (2011) distinguished between errors caused by the participant (internally-caused) and errors caused by presumed technical failures (externally-caused). Whereas internally-caused errors led to post-error slowing, externally-caused errors were followed by a decrease of selective attention presumably reflecting an adaptive disengagement from the task that served to save resources in the face of uncontrollable action outcomes. Whereas these studies demonstrated source-specific adaptive posterror adjustments, other studies provided evidence for non-adaptive adjustments, that is, post-error adjustments that do not prevent further errors but even further impair performance. An alternative explanation for post-error performance decrements is the idea of a resourceconsuming response monitoring process that interferes with subsequent processing (Jentzsch and Dudschig, 2009; Dudschig and Jentzsch, 2009). Furthermore, Notebaert and colleagues (Houtman and Notebaert, 2013; Notebaert et al., 2009) proposed that errors, like any infrequent event, elicit an orienting response. Both ideas could explain why errors are followed not only by post-error slowing but often also by a decrease of post-error accuracy. Specific evidence for an orienting response was provided by Notebaert et al. (2009) showing that post-error slowing turns into post-correct speeding when correct responses become less frequent than errors, and by Van der Borght et al. (2014), who report reduced attentional selectivity following errors in a flanker task. Adaptive and non-adaptive adjustments have frequently been viewed as two alternative accounts. However, recent studies provided evidence that both types of adjustments can co-occur. Purcell and Kiani (2016) combined intracranial recordings in primates with model-based analysis of post-error slowing to show that errors are followed by a non-adaptive reduction of the sensitivity of stimulus processing and an adaptive increase of the response criterion. They argued that the adaptive criterion increase serves to compensate for the reduced sensitivity (in a source-unspecific way). Steinhauser, Ernst, and Ibald (in press) analyzed behavioral data in a dual-task paradigm to demonstrate that the same error can elicit adaptive and non-adaptive adjustments. Whereas non-adaptive adjustments were task-unspecific and decayed within a second, adaptive adjustments spanned several trials and affected only the same task in which the error had occurred. Our brief review of research on post-error adjustments shows that valid conclusions about the nature of post-error adjustments may in many cases require that the exact processes that cause errors are taken into account. This is even more important given that it can be impossible to distinguish non-adaptive from adaptive adjustments if both types manifest in a similar way, such as response slowing. In the present study, we investigated the relationship between error sources and post-error adjustments by combining a conflict task with a visual search paradigm. Visual search tasks typically require participants to indicate whether a target stimulus is present among a set of distractors, with target and distractors differing in one or more feature dimensions (e.g., color, shape). Selective attention in such a task follows a succession of distinct stages (Eimer, 2014a; Eimer, 2014b; Ghorashi et al., 2010). After information about task-relevant features is accu-

Fig. 1. Visual search paradigm that was used in the present study, adapted from Mazza et al. (2009b). The depicted trial is SR incompatible (target diamond on the left hemifield but clipped-off corner on the right side).

271

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

the missing corner of the red target diamond by pressing a left or right button with the index or middle finger of their right hand. With regard to stimulus-response compatibility, this resulted in two different conditions: compatible trials, in which the response matched the hemifield in which the target was presented, and incompatible trials, in which the response and the target hemifield differed. After two practice blocks à 30 trials, participants worked through eight test blocks à 100 trials, resulting in 800 test trials. No trial-wise feedback about response correctness was provided but verbal instruction to respond faster was given after blocks with error rates below 5%.

trials. In two analysis stages, we first identified signatures of error sources by considering whether and how the N2pc and SPCN on error trials deviate from correct trials. We then compared post-error trials with pre-error trials to identify post-error adjustments in these components. If these post-error adjustments were adaptive in a source-specific way, i.e., aiming to counteract the error source, they should be observed in the same component that was previously associated with the error source, and the adjustment should counteract the deviation on the error trial. For instance, if errors occur during target identification, we would expect a reduced SPCN on error trials but an enhanced SPCN on post-error trials. Alternatively, if post-error adjustments were nonadaptive, activity on post-error trials should be impaired irrespective of whether the same impairment is obtained on error trials. These predictions are not mutually exclusive. More complex patterns are also conceivable, indicating, for instance, that adaptive and non-adaptive adjustments coexist, or that adaptive adjustments are unrelated to the error-source, thus being source-unspecific. To increase sensitivity to detect error-related deviations in processing, we applied a decoding approach based on multivariate pattern analysis (MVPA). In standard ERP analysis, signal-to-noise ratio is increased by averaging across a large number of trials. Here, a problem with error trials emerges because errors are rather infrequent and heterogeneous given that correct responding can fail due to a multitude of reasons. In contrast, MVPA can be used to extract robust estimates of specific ERP components by enhancing signal-to noise ratios on a single-trial level and by simultaneously compensating for interindividual differences in the spatial distribution of activity (Parra et al., 2002; Parra et al., 2005; Steinhauser and Yeung, 2010). In its present application, we trained classifiers to decode lateralized ERP components such as N2pc and SPCN using the large dataset of all available trials. These classifiers then served as robust decoders for the respective components and were used to reveal deviations of these components in the much smaller number of error and post-error trials.

Data acquisition The electroencephalogram (EEG) was recorded from 64 electrodes using a BIOSEMI Active-Two system (BioSemi, Amsterdam, The Netherlands; channels Fp1, AF7, AF3, F1, F3, F5, F7, FT7, FC5, FC3, FC1, C1, C3, C5, T7, TP7, CP5, CP3, CP1, P1, P3, P5, P7, P9, PO7, PO3, O1, Iz, Oz, POz, Pz, CPz, Fpz, Fp2, AF8, AF4, AFz, Fz, F2, F4, F6, F8, FT8, FC6, FC4, FC2, FCz, Cz, C2, C4, C6, T8, TP8, CP6, CP4, CP2, P2, P4, P6, P8, P10, PO8, PO4, O2 as well as the left and right mastoid, relative to common mode sense (CMS) and driven right leg (DRL) electrodes). Vertical and horizontal electrooculogram (EOG) was recorded from electrodes above and below the right eye and on the outer canthi of both eyes. All electrodes were off-line re-referenced to linked mastoids. EEG and EOG data were continuously recorded at a sampling rate of 512 Hz. Data analysis Behavioral data Trials with response times (RTs) exceeding two standard deviations from the mean for each participant and condition were excluded from all RT analyses. Error rates were arcsine-square-root-transformed for significance testing (Winer et al., 1991). Conventional measures of post-error adjustments (post-error trials vs. post-correct trials) have been shown to be strongly biased by artefactual variance due to global performance shifts, that is, by changes of RTs and error rates across the experiment (Dutilh et al., 2012). For this reason, we calculated posterror slowing as the difference in RTs of pre-error and post-error trials. Only post-error changes of accuracy were calculated by comparing post-correct trials with post-error trials because the pre-post method is not applicable here. In these and all following analyses, repeated measures ANOVAs and t-tests for paired samples were used.

Method Participants Twenty-two participants (20 female, 19 right-handed) with mean age of 22.0 years participated in the study for payment or course credit. They were recruited at the Catholic University of Eichstätt-Ingolstadt, had normal or corrected to normal vision and had no history of neurological or psychiatric disease. The study was conducted in accordance with the Declaration of Helsinki and informed consent was acquired from all participants. Two participants were excluded from further analysis because their error rates (1.8% and 2.6%) fell below a threshold of two standard deviations below the mean error rate.

Event-related potentials We conducted all analyses using custom-made MATLAB v8.2 (The Mathworks, Natic, MA, USA) scripts together with EEGLAB v12.0 (Delorme and Makeig, 2004) functions. EEG data were band-pass filtered to exclude frequencies below 0.5 Hz and above 40 Hz, divided into epochs from 500 ms before to 1000 ms after stimulus onset and baseline-corrected to the interval of 100 ms before stimulus onset. Following this, a channel was interpolated using spherical spline interpolation if it met the joint probability criterion (threshold 5) as well as the kurtosis criterion (threshold 10) in EEGLAB's channel rejection routine (pop_rejchan.m). This lead to an interpolation of M=1.45 electrodes per subject. Epochs were excluded based on four criteria: a) epochs with activity deviating more than 300 µV from the baseline in any electrodes except Fp1, Fpz, Fp2, Af7, and Af8 (to prevent exclusion of blink artifacts which were corrected in a later stage), b) epochs whose joint probability deviated more than 5 standard deviations from the epoch mean, c) epochs that contained eye blinks in the period of −100 ms to +200 ms around stimulus onset; d) epochs that had response times slower than 1000 ms. In a next step, an infomax-based independent component analysis (Bell and Sejnowski, 1995) implemented in EEGLAB (Delorme and Makeig, 2004) was conducted and after visual inspection, out of the resulting 64

Task and procedure For this study, we adopted the visual search paradigm used by Mazza et al. (2009b) and implemented the condition that elicited the largest N2pc in their study (Experiment 1, “constant – 20 elements”, see Fig. 1). Visual search arrays consisted of 1 red and 19 green diamonds (.6°×.8°, each having a .4° corner clipped off on the left or right side) randomly distributed within an 8×8 matrix (8.6°×8.6°) with an equal number in each hemifield. A small white square (.1°×.1°) in the middle of this array served as fixation point. The red target diamond was constrained to the outermost two columns of each hemifield and appeared with equal probability on the left or right side. After a random and exponentially distributed response-cue interval ranging between 1000 ms and 1350 ms, the fixation square appeared and, after 1000 ms, was complemented by the visual search array, which was visible for 150 ms. Participants had to indicate the side of 272

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

aging across participants. Attempting to optimally distinguish the event e from the complementary event, for every time window t this pattern analysis assigns a classification output value ye (t ) to every single trial. Because ye (t) represents the predicted value of a logistic regression, and thus corresponds to the probability of the trial to belong to the event e (averaged discrimination activity, see Parra et al. (2002), it ranges from 0 to 1. Due to the specific contrast used for training, i.e. separating RHtrials and LH-trials, this value initially indicated the degree of the respective trial to resemble a prototypical RH-trial, and thus, the strength of visual attention to the right hemifield. As we were interested in an estimate of the strength of visual attention independent of the actual side of the target, we subtracted the output values of all LH-trials from 1 so that afterwards the output value indicated the strength of visual attention to the target, independent of hemifield. As a result, we obtained a decoder of lateralized visual attention to the target for each time window, with output values close to 1 indicating a large degree of visual attention to the target hemifield. To visualize the spatial distribution of each decoder, we computed the coupling coefficient vector, which represents the activity at each electrode site that correlates with the respective discriminating component, and thus can be thought of as the “sensor projection” of that component (Parra et al., 2002, 2005). Being interested in decoder-based measures of the N2pc and the SPCN (subsequently termed robust N2pc and robust SPCN), we considered only classifier time windows within the a priori regions of interest of these ERP components for significance testing. Hence, to test for the robust N2pc, classifiers with centers at 200, 220 ms, 240 ms, 260 ms, and 280 ms were considered (covering ERP data from 180–300 ms) and as for the robust SPCN, classifiers with centers at 300 ms, 320 ms, 340 ms, 360 ms, 380 ms, and 400 ms were analyzed (covering ERP data from 280–420 ms). Within these groups of classifiers, all statistical tests were conducted separately for each time point, and results are reported after correcting for false discovery rate (FDR) using Benjamini-Hochberg correction (Benjamini and Hochberg, 1995).

independent components (ICs) those were removed that represented eye blinks and muscular artifacts. As MVPA extracts information from the whole scalp topography and markers of selective attention might correlate with horizontal eye movements, it is of particular importance that the data set is cleaned thoroughly from EOG artifacts. To achieve this, we used a two-step approach of EOG artifact correction. First, the time courses of the remaining ICs were correlated with the difference of the two horizontal EOG eye channels for each subject.1 Those components were removed that both correlated with the horizontal EOG time course with r > .5 and showed an IC topography with bipolar activity peaking at the most frontal electrodes. Second, an EOG correction method based on least mean squares regression (Gómez-Herrero, 2008) was applied to the ICA-corrected dataset. A combination of these two approaches yielded superior results compared to a separate application of each of these methods, as ICA-based correction alone was not effective enough and the regression-based approach alone would have led to flawed correction and overcompensation. After artifact correction, an average of 96.2% of the trials was kept for further analysis, from which a mean of 44.4 compatible and 57.6 incompatible error trials and comparable numbers of the respective pre-error and post-error trials emerged. In a next step, we considered posterior cortical activity from electrodes PO7 and PO8 from correct compatible and incompatible trials to investigate whether a typical N2pc and SPCN can be found in our data. We first averaged waveforms separately for the electrode contra- and ipsilateral to the target hemifield. We then created the difference wave by subtracting the ipsilateral waveform from the contralateral waveform (Luck and Hillyard, 1994). N2pc and SPCN were quantified by computing mean amplitudes within time intervals of 180–300 ms for the N2pc, and 300–400 ms for the SPCN (Mazza et al., 2009a, 2009b; Mazza et al., 2007). Activity beyond 400 ms was not further analyzed due to overlap with response-locked activity (RTerror=427 ms).

Decoder analysis The linear integration method by Parra et al., (2002) has previously been used successfully to compute neural single-trial representations of cognitive processes (Boldt and Yeung, 2015; Maier et al., 2011; Philiastides et al., 2006; Steinhauser and Yeung, 2010). This method utilizes penalized logistic regression (Parra, 2005) to compute a weight vector that discriminates optimally between trials in two conditions – in our case between trials with the target stimulus in the left hemifield (LH-trials) and trials with the target stimulus in the right hemifield (RH-trials). For each participant, a weight vector was calculated for each of 31 partly overlapping time windows of 40 ms width, separated by 20 ms in the time range from 0 ms to 600 ms post-stimulus. Training set was an equal, randomly drawn number of LH-trials and RH-trials. The area under the Receiver Operating Characteristic curve (Az score) served to quantify the sensitivity of the resulting classifiers. Az=0.5 would indicate classification at chance level, while Az=1.0 mirrors perfect discrimination. To prevent overfitting, leave-one-out (LOO) cross-validation was used so that eventually, every weight vector used for further analysis was the mean of N weight vectors trained with T*(N-1) samples of N-1 trials to predict the T samples of the remaining trial. Az scores were finally tested with regard to significant classifier sensitivity by means of a permutation test. For each participant, a test distribution under the null hypothesis (discrimination at chance level) was generated by recomputing Az scores 1000 times with random assignment of the RH/LH categories. The resulting test distribution was used to determine critical Az values associated with different significance levels. Overall critical Az values were computed by aver-

Results Behavioral data Mean RTs and mean error rates are presented in Table 1. Error rates were lower for compatible than for incompatible trials, t(19) =3.07, p=.006. Analyzing mean RTs with an ANOVA with the variables response type (correct, error) and compatibility (compatible, incompatible) yielded significant main effects of response type, F(1,19)=126.9, p < .001, and compatibility, (F(1,19)=34.8, p < .001, as well as a significant interaction effect, F(1, 19)=14.8, p=.001. Mean RTs were slower for compatible error trials than for incompatible error trials, t(19)=4.84, p < .001, whereas no such difference was found for correct trials, t(19)=1.06, p=.30. We obtained significant post-error slowing, t(19)=7.71, p < .001, which was larger after errors on incompatible trials than after errors on compatible trials, t(19)=3.58, p=.002. Furthermore, we obtained a significant post-error increase of accuracy, t(19)=4.28, p < .001, which was independent of compatibility. EEG data We first analyzed stimulus-locked ERPs of correct trials to investigate whether N2pc and SPCN components can be found in our data. Fig. 2 shows waveforms from parieto-occipital electrode sites (PO7/ PO8) contralateral and ipsilateral to the target, as well as the difference wave between both. We obtained a strong contralateral negativity peaking around 200 ms post-stimulus, the N2pc, t(19)=2.74, p=.013, and an additional, though somewhat smaller negativity from 300 ms to 400 ms, the SPCN, t(19)=2.09, p=.050. Both components were similar

1 Four subjects did not show typical eye movement artifacts in the EOG electrodes at the outer canthi, most likely due to incorrect application of the electrodes. For this reason, we used the difference of channels AF7 and AF8 as substitutes in these cases.

273

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

Table 1 Reaction times, error rates and post-error slowing and post-error increase of accuracy in the experimental conditions.

RT correct RT error error rate PES PEIA

compatible

incompatible

473 ms 454 ms 11.6% 17 ms 3.2%

478 ms 410 ms 15.0% 37 ms 3.2%

Notes: RT=response times, PES=post-error slowing, PEIA=post-error increase of accuracy.

Fig. 3. Training of a multivariate pattern classifier to optimally distinguish between LHtrials and RH-trials. The red line indicates classifier accuracy (Az-Scores) after leave-oneout cross-validation (dotted line: significance threshold for p < .01 based on permutation test). Bottom: scalp topography of the discriminating component represented by coupling coefficients.

significantly above chance level in a time window starting at around 160 ms after stimulus onset. Matching the time pattern of conventional ERP analysis, the classifiers reached their peak of discrimination accuracy around 200 ms with an Az-Score of .76 and a discriminating topography that is very similar to the spatial distribution of the N2pc. The classifiers then showed a plateau of accuracy (Az ≈ .67) in the time range of the SPCN (300–400 ms) after which accuracy gradually decreased back to chance level. These results demonstrate that our classifiers capture the time course and spatial distribution of the N2pc and the SPCN, and thus, are valid decoders of these components. By considering the averaged output values ye (t ) of these decoders in the time range of each of these components, we obtained the robust N2pc and the robust SPCN. We first compared these robust components for correct trials and errors, separately for compatible and incompatible trials, in order to identify neural signatures of error sources. To this end, we computed two-way ANOVAs with the variables response type (correct, error) and compatibility (compatible, incompatible) for each time point in the intervals associated with the N2pc and SPCN, respectively. Ranges of F-values and FWE-corrected p-values are reported. Fig. 4 reveals two effects: First, we obtained a significant reduction of the robust SPCN in errors in the time windows from 320 ms to 400 ms, all Fs(1,19) > 7.09, all ps < .018. No significant effects involving compatibility were observed for the SPCN. Second, we observed an increased robust N2pc on error trials but this effect was restricted to incompatible trials. Significant interaction effects were obtained in the two time windows centered at 220 ms and 240 ms, all Fs(1,19) > 7.23, all ps < .036. Planned contrasts revealed an increased robust N2pc on errors only for incompatible trials in these time windows, all ts(19) > 2.44, all ps < .025 (Fig 4B). These results point at two distinct sources for performance errors in this visual search paradigm: One source is reflected by a reduction of the robust SPCN and contributes to all errors irrespective of compatibility. Another source is characterized by an increased robust N2pc and contributes to errors in incompatible trials only. In our next step of analysis, we wanted to find out about possible post-error adjustments in these components. To this end, we analyzed classifier outputs from correct trials following errors and those preceding errors with two-way ANOVAs with the variables response type (pre-error, post-error) and compatibility (compatible error, incompatible error) for each time point in the intervals of the N2pc and SPCN. The classifier outputs depicted in Fig. 5A and B reveal a reduction of the robust N2pc after compatible and incompatible errors. The late classifier time window within the time area of the robust N2pc, centered at 240 ms, shows a significant main effect of response type, F(1,19) =8.84, p=.039. Additionally, the time window centered at 220

Fig. 2. Stimulus-locked neural activity on correct compatible (A) and incompatible (B) trials ipsilateral and contralateral to the hemifield of the target diamond, recorded at posterior electrodes PO7 and PO8. Gray areas on difference waves (C) indicate the apriori time windows for quantifying mean N2pc and SPCN amplitudes.

for compatible and incompatible trials (ps > .26). In a next stage, we trained MVPA classifiers on consecutive, partially overlapping time windows to create decoders whose output can be used as a robust measure of these components. As illustrated in Fig. 3, the classifiers were able to distinguish LH trials from RH trials

274

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

Fig. 4. Classifier output values ye (t ) of the decoder analysis for correct vs. error trials on congruent (A) and incongruent (B) trials, with larger values indicating higher levels of selective attention. White bars below the respective chart indicate the a-priori time areas of N2pc and SPCN. Black bars indicate classifier windows with significant (FDR-corrected) effects within those time areas (Interaction=Interaction effect of response type and compatibility; ME=main effect).

therefore conducted this analysis only for these trials. The compatibility of the error trial was not considered as a further variable, because the previous analysis did not reveal an effect of this variable and because this would have implied a rather small number of trials per cell. The respective data are presented in Fig. 5C and D. Classifier outputs from correct trials were subjected to two-way ANOVAs with the variables response type (pre-error, post-error) and compatibility of the pre/post-error trial (compatible, incompatible). This ANOVA yielded a significant interaction in a time window of the robust N2pc centered at 240 ms, F(1,19)=10.66, p=.020. Additionally, this interaction was marginally significant in adjacent time windows (220 ms: F(1,19) =5.33, p=.054; 260 ms: F(1,19)=5.72, p=.054). Planned contrasts

ms showed a marginally significant main effect, F(1,19)=6.01, p=.060. Although this post-error reduction of the robust N2pc occurs in the same time range that is associated with an increased robust N2pc on incompatible errors, the absence of an interaction effect suggests that this post-error adjustment of selective attention towards the target is implemented independently of the compatibility of the error trial. No effect in the time window of the SPCN was obtained. So far, we considered only the compatibility of the error trial but not that of the post-error trial. Because pre-error trials serve as a baseline for post-error trials, analyzing compatibility effects in posterror trials requires that only sequences are included in which the posterror trial has the same compatibility as the pre-error trial. We

Fig. 5. Classifier output values ye (t ) of the decoder analysis for pre- vs. post-error correct trials. Larger values indicate higher levels of selective attention. A and B: Classifier output as a function of the compatibility on the error trial. C and D: Classifier output as a function of the compatibility on the post-error trial. White bars below the respective chart indicate the apriori time areas of N2pc and SPCN. Black bars indicate classifier windows with significant (FDR-corrected) effects within those time areas (Interaction=Interaction effect of response time and compatibility; ME=main effect).

275

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

attention (King et al., 2010) and motion processing (Danielmeier et al., 2011). Interestingly, whether a post-error adjustment was initiated in our study did not depend on the compatibility of the error trial but on the compatibility of the post-error trial. This indicates that the posterror adjustment is not chosen based on a trial-wise evaluation of the error source. Rather, the control mechanism might have learned that errors are potentially caused by enhanced target location processing, and therefore initiates a corresponding adjustment following each error, not only on errors that were actually caused by enhanced target location processing on incompatible trials. Whether this adjustment was actually implemented on the post-error trial depended on whether the stimulus was incompatible, and thus, whether errors due to enhanced target location processing could actually occur. This implies that the adjustment was initiated during the post-error trial after incompatibility was detected. This is indeed plausible given that relevant and irrelevant features in a Simon task can activate their corresponding response side rather early (Eimer, 1998; Eimer et al., 1995), and that online adaptation of selective attention has been demonstrated in the context of other adaptive phenomena (Scherbaum et al., 2010). In a recent study by Van der Borght et al. (2016), a post-error reduction of an early marker of visual attention, the posterior N1, is interpreted as a non-adaptive effect. The authors argue that visual attention is reduced after errors because fewer resources are available for an active focusing of attention to the stimulus. Whereas this finding seems to contradict the present conclusions, it might reflect that this study differs from the present one in a crucial aspect. In their study, the task in which the error occurred (a flanker task) differed from the task in which the post-error adjustment was observed (a simple visual discrimination task). Accordingly, the results of Van der Borght et al. (2016) possibly reflect that adjustments to visual attention in a flanker task do not generalize to another task, thus preventing that any adaptive adjustments are observed. Indeed, Steinhauser et al. (2016) have recently shown that an error in one task elicits only nonadaptive adjustments in a subsequent different task (as in the study of Van der Borght et al., 2016). In contrast, adaptive adjustments are observed only when analyzing the next occurrence of the same task in which the error has occurred (as in the present study). In line with the idea that adjustments of visual attention differ between the two studies, we did not observe a corresponding N1 effect in our data.3 However, not all neural signatures of errors led to an adaptive posterror adjustment. In addition to the increased N2pc, errors were also associated with a decreased SPCN, and this effect was independent of the compatibility of the error trial. Previous studies have linked the SPCN to the maintenance and processing of target features (Jolicœur et al., 2008; Woodman and Vogel, 2008). The decreased SPCN on error trials might thus reflect that, even more than locating and selecting the target, the analysis of target features itself is a difficult and error-prone process. The decreased SPCN on error trials demonstrates that a considerable proportion of errors must have emerged during this target identification stage (Eimer, 2014a). Nonetheless, we found no evidence for adjustments to the SPCN on post-error trials. This could suggest either that post-error adjustments are not possible on this stage, or that these adjustments are not reflected in a change of the SPCN. It may be argued that subjects misinterpreted the instruction and in fact did not respond to the cut-off side of the target diamond but to the opposing side, perceiving it as an arrowhead. However, behavioral results argue against this possibility: Stimulus-response incompatible trials showed increased error rates (+3.4%) and incompatible errors

revealed that the reduction occurred only in incompatible trials, all ts(19) > 2.99, all ps < .019, suggesting that the post-error reduction of the robust N2pc was obtained only on incompatible pre/post-error trials.

Discussion The present study demonstrates adaptive post-error adjustments of selective attention in a visual search paradigm. Errors in this task could emerge on at least two stages, target selection and target identification, which are represented by two lateralized components, the N2pc and the SPCN. We reasoned that a post-error adjustment is adaptive in a source-specific way if, first, the adjustment occurs at the same processing stage as the source of the error, and second, the direction of the adjustment counteracts the deviation reflecting the error source. In our data, these two criteria apply to the N2pc. We found an increased N2pc on error trials relative to correct trials, but a decreased N2pc on post-error trials relative to pre-error trials.2 This might reflect the operation of a source-specific adaptive cognitive control mechanism which evaluates the error source and then initiates appropriate attentional adjustments counteracting this source. Whereas previous studies provided evidence for an evaluation of error sources (Maier et al., 2012; Maier and Steinhauser, 2013; Maier and Steinhauser, 2016; Steinhauser and Kiesel, 2011), and corresponding adaptive behavioral adjustments (Maier et al., 2011), the present study reveals neural signatures of source-specific adaptive post-error adjustments by showing that adjustments apply to the same neural marker as the error source. It appears to be surprising that errors on incompatible trials are associated with an increased N2pc given that a larger N2pc amplitude is normally interpreted as a marker of more efficient visual selection. The N2pc has previously been assumed to reflect successful target selection by suppressing distractor representations (Luck and Hillyard, 1990; Luck and Hillyard, 1994), enhancing target representations (Mazza et al., 2009), or both (Hickey et al., 2009). In the present task, however, target localization is task-beneficial only to a certain degree, as the actual objective is to identify a feature of the target stimulus, and the characteristic of this feature (the cut-off side of the diamond) can oppose the hemifield of the target location. Therefore, this characteristic of the task at hand can lead to a situation, in which more attention to the target location increases the probability that the target location activates the corresponding response side due to stimulus-response compatibility (Kornblum et al., 1990). In other words, errors are associated with an increased N2pc, because strong target location processing leads to strong activation of the response corresponding to the target hemifield, which, in case of an incompatible stimulus, opposes the task-relevant (cut-off) side of the target diamond and thus leads to an error. In accord with this idea, the N2pc is increased only for errors on incompatible trials because enhanced target location processing does not increase the risk of committing an error on compatible trials. From this perspective, post-error adjustments reflected by the N2pc can be viewed as an adaptive reduction of target location processing to prevent errors due to stimulus-response incompatibility. Similar adaptive post-error suppression of task-irrelevant stimulus features in conflict tasks has previously been found in fMRI studies on spatial 2 An overall analysis aiming to verify the reversal of increased N2pc in (incompatible) errors and decreased N2pc in post-error trials by integrating both the analysis on error causes and the analyses on post-error effects is unfeasible because these analyses are based on different comparisons and result in an incomplete design. Nonetheless, comparing the difference wave of “correct minus error in incompatible trials” (Fig. 4A) with the difference waves of “pre-error vs. post-error” as depicted in Fig. 5 substantiates our interpretation of a reversal of error-related increase and post-error related decrease of the N2pc (240 ms: p=.022 for the main effect with regard to compatibility of the error trial, p=.020 with regard to compatibility of the post-trial).

3 In our data, a comparison of the N1 in post-error and post-correct trials (as in Van der Borght et al. (2016)) failed to produce significant results. Rather, a trend towards an increased N1 in post-error trials could be observed, on the hemisphere ipsilateral to the target (p=.11). This opposite trend is in support of the idea that post-error changes to visual attention have to be treated differently in Van der Borght et al. (2016) and in this study.

276

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

potentially adaptive post-error adjustments, particularly on the neural level. If an adaptive control mechanism seeks to mend deviations from task-beneficial processing that gradually accumulate until peaking in the error trial (Steinhauser et al., 2012), its aim would certainly be to reduce such deviations back to the level that is found in correct trials. This adjustment is best reflected in a change relative to the pre-error baseline in which dysfunctional processing is already foreshadowed. In contrast, comparing post-error trials with post-correct trials reduces the power to find such an adjustment because post-correct trials are already associated with functional processing and there is no reason why control mechanisms should optimize processing beyond that seen in these post-correct trials. For the reasons described above, in contrast to most previous research, the present study reports an unbiased neural signature of post-error adjustment by applying the rationale of Dutilh et al. (2012), comparing post-error trials with pre-error trials.4 Taken together, these findings provide insight into the complex relationship of error sources and post-error adjustments. While we were able to identify two distinct sources of errors, only one error source – an increased N2pc on error trials presumably due to enhanced target location processing – was found to be adjusted in an adaptive way counteracting the error source on post-error trials. On the one hand, this shows that human performance monitoring indeed uses source-specific adaptive post-error adjustments to prevent the occurrence of future errors. On the other hand, adaptive post-error adjustments were applied only to some but not all error sources. Indeed, inconsistent adjustments across different error sources might explain why a post-error increase of accuracy as an index for adaptive adjustments is found in some studies (e.g. Danielmeier et al., 2011; Maier et al., 2011; Marco-Pallarés et al., 2008) but not in others (Hajcak and Simons, 2008; Notebaert et al., 2009; Danielmeier and Ullsperger, 2011). This suggests that research on performance monitoring and cognitive control would benefit if a broader range of tasks and possible error types were investigated to reveal the basic principles of adaptive cognitive control and learning from errors.

were linked to faster responses (−44 ms). If indeed subjects would have focused their attention on the assumed arrowhead, trials that are marked as incompatible based on the actual instruction (because target location and cut-off side differ) would have to be considered as compatible trials (because target location and arrowhead are on the same side) and vice versa. Consequently, error rates with regard to compatibility as well as error response time should be reversed as well, which is not the case. Taken together, the present results are in line with theories that postulate adaptive post-error adjustments. Models of adaptive cognitive control assumed that adjustments of cognitive control are sensitive to the source of errors or conflicts. For instance, in their seminal work on conflict monitoring theory, Botvinick et al., (2001) proposed that conflict due to incompatible flankers leads to an adjustment of spatial attention whereas conflict due to errors leads to an adjustment of response bias. However, their account is underspecified with respect to the question how the type of adjustment is chosen. Maier et al. (2011) proposed a two-stage model of error evaluation, consisting of an early evaluation stage, which operates during task processing and a late evaluation stage after the response. The early stage monitors attentional parameters during task processing, and provides the input for a cognitive control system that initiates source-specific adaptive adjustments following errors (see also Maier and Steinhauser, 2013). Whereas Maier et al. (2011) found behavioral post-error adjustments that matched the preceding error type, the present study found adaptive adjustments of the N2pc even following congruent error trials on which the N2pc did not deviate. It appears that in the present task, monitoring of attentional states during task processing was used to establish generic adjustments without conditionalizing the implementation of these adjustments on these attentional states. Future studies will have to reveal the boundary conditions under which adaptive adjustments are chosen based on an evaluation of the error type. In contrast to previous studies, we measured post-error adjustments while controlling for neural activity on the error trial. Previous studies often considered activity on post-error trials only (e.g., Purcell and Kiani, 2016), which can lead to false inferences in two ways: First, one could confuse a post-error adjustment with a persisting error source, i.e., post-error activity could actually reflect neural activity related to the error source that persists into the post-error trial. In our study, this could have been the case if the attentional state reflected by a decreased N2pc on error trials had persisted into the post-error trial again causing a decreased N2pc, which would then mistakenly have been interpreted as post-error adjustment. Second, one could confuse adaptive with non-adaptive post-error adjustments. Reduced activity in the post-error trial could erroneously be classified as a non-adaptive adjustment, if activity on the error trial is ignored. In the present study, we could exclude that the decreased N2pc reflects a non-adaptive adjustment or a persisting error source because the N2pc was increased on the error trial. The fact that error source and post-error adjustment are represented by deviations in opposite directions from the baseline is a direct marker for the source-specificity and thus adaptive nature of this adjustment. A further problem for the measurement of post-error adjustments are the confounding effects of persistent error sources as well as effects of global performance shifts (Dutilh et al., 2012). Previous studies frequently measured post-error adjustments by comparing post-error trials with post-correct trials. In a recent analysis of post-error slowing, Dutilh et al., (2012) showed that this contrast is biased when error probability is nonstationary, i.e., varies across the experiment. In this case, post-error trials are overproportionally sampled from periods in which the error rate is high. Altered activity on post-error trials could simply reflect activity changes associated with more error-prone periods of the experiment, and thus, a neural signature of the error source. To prevent this confound, Dutilh et al. (2012) proposed to use a post-error vs. pre-error contrast to quantify post-error adjustments. This approach has an additional advantage with regard to measuring

Acknowledgements This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Correspondence concerning this article should be addressed to Robert Steinhauser, Catholic University of Eichstätt-Ingolstadt, Ostenstraße 25, 85072 Eichstätt, Germany. E-mail: robert. [email protected]. References Bell, A.J., Sejnowski, T.J., 1995. An information-maximization approach to blind separation and blind deconvolution. Neural Comput. 7 (6), 1129–1159. http:// dx.doi.org/10.1162/neco.1995.7.6.1129. Benjamini, Y., Hochberg, Y., 1995. Controlling the false discovery rate: a practical and powerful approach to multiple testing. J. R. Stat. Soc. Ser. B (Methodol.) 57 (1), 289–300. Boldt, A., Yeung, N., 2015. Shared neural markers of decision confidence and error detection. J. Neurosci. 35 (8), 3478–3484. http://dx.doi.org/10.1523/ JNEUROSCI.0797-14.2015. Botvinick, M.M., Braver, T.S., Barch, D.M., Carter, C.S., Cohen, J.D., 2001. Conflict monitoring and cognitive control. Psychol. Rev. 108 (3), 624–652.

4 Nonetheless, we conducted analyses both with conventional (post-correct) baseline and a baseline based on two trials prior to the error. These analyses share the same general pattern as our main analysis on the descriptive level. However, some effects did not reach significance. [Analysis with regard to compatibility of the error trial: main effect of response type not significant in both baselines; Analysis with regard to compatibility of the post-trial: interaction marginally significant with post-correct baseline (ps < .063 at 200, 240, and 260ms), interaction significant with two-beforeerror baseline (ps < .033 at 200, 240, and 260ms)]. As discussed in the main text, this presumably reflects that a pre-error baseline is beneficial for detecting post-error adjustments because the error source is already foreshadowed on pre-error trials and the aim of these adjustments is to compensate this error source.

277

NeuroImage 150 (2017) 270–278

R. Steinhauser et al.

33 (40), 15705–15709. http://dx.doi.org/10.1523/JNEUROSCI.2785-13.2013. Maier, M.E., Steinhauser, M., 2016. Error significance but not error expectancy predicts error-related negativities for different error types. Behav. Brain Res. 297, 259–267. http://dx.doi.org/10.1016/j.bbr.2015.10.031. Maier, M.E., Yeung, N., Steinhauser, M., 2011. Error-related brain activity and adjustments of selective attention following errors. NeuroImage 56 (4), 2339–2347. Marco-Pallarés, J., Camara, E., Münte, T.F., Rodríguez-Fornells, A., 2008. Neural mechanisms underlying adaptive actions after slips. J. Cogn. Neurosci. 20 (9), 1595–1610. http://dx.doi.org/10.1162/jocn.2008.20117. Mazza, V., Turatto, M., Caramazza, A., 2009a. An electrophysiological assessment of distractor suppression in visual search tasks. Psychophysiology 46 (4), 771–775. http://dx.doi.org/10.1111/j.1469-8986.2009.00814.x. Mazza, V., Turatto, M., Caramazza, A., 2009b. Attention selection, distractor suppression and N2pc. Cortex; J. Devoted Study Nerv. Syst. Behav. 45 (7), 879–890. http:// dx.doi.org/10.1016/j.cortex.2008.10.009. Mazza, V., Turatto, M., Umiltà, C., Eimer, M., 2007. Attentional selection and identification of visual objects are reflected by distinct electrophysiological responses. Exp. Brain Res. 181 (3), 531–536. http://dx.doi.org/10.1007/s00221007-1002-4. Notebaert, W., Houtman, F., Opstal, F., Van, Gevers, W., Fias, W., Verguts, T., 2009. Post-error slowing: an orienting account. Cognition 111 (2), 275–279. Notebaert, W., Houtman, F., Opstal, F., Van, Gevers, W., Fias, W., Verguts, T., 2009. Post-error slowing: an orienting account. Cognition 111 (2), 275–279. http:// dx.doi.org/10.1016/j.cognition.2009.02.002. Parra, L.C., Alvino, C., Tang, A., Pearlmutter, B., Yeung, N., Osman, A., Sajda, P., 2002. Linear spatial integration for single-trial detection in encephalography. NeuroImage 17 (1), 223–230. http://dx.doi.org/10.1006/nimg.2002.1212. Parra, L.C., Spence, C.D., Gerson, A.D., Sajda, P., 2005. Recipes for the linear analysis of EEG. NeuroImage 28 (2), 326–341. http://dx.doi.org/10.1016/ j.neuroimage.2005.05.032. Philiastides, M.G., Ratcliff, R., Sajda, P., 2006. Neural representation of task difficulty and decision making during perceptual categorization: a timing diagram. J. Neurosci.: Off. J. Soc. Neurosci. 26 (35), 8965–8975. http://dx.doi.org/10.1523/ JNEUROSCI.1655-06.2006. Purcell, B.A., Kiani, R., 2016. Neural mechanisms of post-error adjustments of decision policy in parietal cortex. Neuron 89 (3), 658–671. http://dx.doi.org/10.1016/ j.neuron.2015.12.027. Ridderinkhof, K.R., Ullsperger, M., Crone, E.A., Nieuwenhuis, S., 2004. The role of the medial frontal cortex in cognitive control. Science 306 (5695), 443–447. Scherbaum, S., Dshemuchadse, M., Fischer, R., Goschke, T., 2010. How decisions evolve: the temporal dynamics of action selection. Cognition 115 (3), 407–416. http:// dx.doi.org/10.1016/j.cognition.2010.02.004. Simon, J.R., Rudell, A.P., 1967. Auditory S-R compatibility: the effect of an irrelevant cue on information processing. J. Appl. Psychol. 51 (3), 300–304. Steinhauser, M., Eichele, H., Juvodden, H.T., Huster, R.J., Ullsperger, M., Eichele, T., 2012. Error-preceding brain activity reflects (mal-)adaptive adjustments of cognitive control: a modeling study. Front. Human. Neurosci. 6, 97. http://dx.doi.org/ 10.3389/fnhum.2012.00097. Steinhauser, M., Ernst, B., Ibald, K., 2016. Isolating component processes of posterror slowing with the psychological refractory period paradigm. J. Exp. Psychol.: Learn. Mem. Cogn.. http://dx.doi.org/10.1037/xlm0000329, Advance online publication. Steinhauser, M., Kiesel, A., 2011. Performance monitoring and the causal attribution of errors. Cogn. Affect. Behav. Neurosci. 11 (3), 309–320. http://dx.doi.org/10.3758/ s13415-011-0033-2. Steinhauser, M., Yeung, N., 2010. Decision processes in human performance monitoring. J. Neurosci.. http://dx.doi.org/10.1523/JNEUROSCI.1899-10.2010. Van der Borght, L., Braem, S., Notebaert, W., 2014. Disentangling posterror and postconflict reduction of interference. Psychon. Bull. Rev. 21 (6), 1530–1536. http:// dx.doi.org/10.3758/s13423-014-0628-z. Van der Borght, L., Schevernels, H., Burle, B., Notebaert, W., 2016. Errors disrupt subsequent early attentional processes. PLoS One 11 (4), e0151843. http:// dx.doi.org/10.1371/journal.pone.0151843. Weissman, D.H., Roberts, K.C., Visscher, K.M., Woldorff, M.G., 2006. The neural bases of momentary lapses in attention. Nat. Neurosci. 9 (7), 971–978. http://dx.doi.org/ 10.1038/nn1727. Winer, B.J., Brown, D.R., Michels, K.M., 1991. Statistical Principles in Experimental Design (3rd ed.). McGraw-Hill, New York. Wolfe, J.M., 2007. Guided Search 4.0: current Progress with a model of visual search. In: Gray, W. (Ed.), Integrated Models of Cognitive Systems. Oxford Univ. Press, 99–119, (Retrieved from Wolfe (2007) - Guided Search4.pdf). Woodman, G.F., Vogel, E.K., 2008. Selective storage and maintenance of an object's features in visual working memory. Psychon. Bull. Rev. 15 (1), 223–229. http:// dx.doi.org/10.3758/PBR.15.1.223.

Brewer, N., Smith, G.A., 1989. Developmental changes in processing speed: influence of speed-accuracy regulation. J. Exp. Psychol.: General 118 (3), 298–310. http:// dx.doi.org/10.1037/0096-3445.118.3.298. Danielmeier, C., Eichele, T., Forstmann, B.U., Tittgemeyer, M., Ullsperger, M., 2011. Posterior medial frontal cortex activity predicts post-error adaptations in taskrelated visual and motor areas. J. Neurosci. 31 (5), 1780–1789. http://dx.doi.org/ 10.1523/JNEUROSCI.4299-10.2011. Danielmeier, C., Ullsperger, M., 2011. Post-error adjustments. Front. Psychol. 2. http:// dx.doi.org/10.3389/fpsyg.2011.00233. Delorme, A., Makeig, S., 2004. EEGLAB: an open source toolbox for analysis of singletrial EEG dynamics including independent component analysis. J. Neurosci. Methods 134, 9–21. Dudschig, C., Jentzsch, I., 2009. Speeding before and slowing after errors: is it all just strategy? Brain Res. 1296, 56–62. Dutilh, G., Van Ravenzwaaij, D., Nieuwenhuis, S., Van der Maas, H.L.J., Forstmann, B.U., Wagenmakers, E.J., 2012. How to measure post-error slowing: a confound and a simple solution. J. Math. Psychol. 56 (3), 208–216. http://dx.doi.org/10.1016/ j.jmp.2012.04.001. Dutilh, G., Vandekerckhove, J., Forstmann, B.U., Keuleers, E., Brysbaert, M., Wagenmakers, E.-J., 2012. Testing theories of post-error slowing. Atten. Percept. Psychophys. 74 (2), 454–465. http://dx.doi.org/10.3758/s13414-011-0243-2. Dutilh, G., Wagenmakers, E., Visser, I., Maas, H.L.J. Van Der, 2011. A phase transition model for the speed-accuracy trade-off in response time experiments. Cogn. Sci. 35, 211–250. http://dx.doi.org/10.1111/j.1551-6709.2010.01147.x. Eimer, M., 1998. The lateralized readiness potential as an on-line measure of central response activation processes. Behav. Res. Methods Instrum. Comput. 30 (1), 146–156. http://dx.doi.org/10.3758/BF03209424. Eimer, M., 2014a. The neural basis of attentional control in visual search. Trends Cogn. Sci. 18 (10), 1–10. http://dx.doi.org/10.1016/j.tics.2014.05.005. Eimer, M., 2014b. The time course of spatial attention: insights from event-related brain potentials. Oxf. Handb. Atten.. http://dx.doi.org/10.1093/oxfordhb/ 9780199675111.013.006. Eimer, M., Hommel, B., Prinz, W., 1995. S-R compatibility and response selection. Acta Psychol. 90 (1–3), 301–313. http://dx.doi.org/10.1016/0001-6918(95)00022-M. Eriksen, B.A., Eriksen, C.W., 1974. Effects of noise letters upon the identification of a target letter in a nonsearch task. Percept. Psychophys. 16 (1), 143–149. Ghorashi, S., Enns, J.T., Klein, R.M., Lollo, V. Di, 2010. Spatial selection and target identification are separable processes in visual search. J. Vision. 10 (3), 1–12. http:// dx.doi.org/10.1167/10.3.7.Introduction. Gómez-Herrero, G., 2008. Automatic Artifact Removal (AAR) toolbox for MATLAB. Transform. Methods Electroencephalogr. (EEG), Retrieved from 〈http://germangh. github.io/pubs/aardoc07.pdf〉. Hajcak, G., Simons, R.F., 2008. Oops!.. I did it again: an ERP and behavioral study of double-errors. Brain Cogn. 68 (1), 15–21. http://dx.doi.org/10.1016/ j.bandc.2008.02.118. Hickey, C., Di Lollo, V., McDonald, J.J., 2009. Electrophysiological indices of target and distractor processing in visual search. J. Cogn. Neurosci. 21 (4), 760–775. http:// dx.doi.org/10.1162/jocn.2009.21039. Houtman, F., Notebaert, W., 2013. Blinded by an error. Cognition 128 (2), 228–236. Jentzsch, I., Dudschig, C., 2009. Why do we slow down after an error? Mechanisms underlying the effects of posterror slowing. Q. J. Exp. Psychol. 62 (2), 209–218. Jentzsch, I., Leuthold, H., 2006. Control over speeded actions: a common processing locus for micro- and macro-trade-offs? Q. J. Exp. Psychol. 59 (8), 1329–1337. http://dx.doi.org/10.1080/17470210600674394. Jolicœur, P., Brisson, B., Robitaille, N., 2008. Dissociation of the N2pc and sustained posterior contralateral negativity in a choice response task. Brain Res. 1215, 160–172. http://dx.doi.org/10.1016/j.brainres.2008.03.059. King, J.A., Korb, F.M., von Cramon, D.Y., Ullsperger, M., 2010. Post-error behavioral adjustments are facilitated by activation and suppression of task-relevant and taskirrelevant information processing. J. Neurosci. 30 (38), 12759–12769. http:// dx.doi.org/10.1523/JNEUROSCI.3274-10.2010. Kornblum, S., Hasbroucq, T., Osman, a, 1990. Dimensional overlap: cognitive basis for stimulus-response compatibility–a model and taxonomy. Psychol. Rev. 97 (2), 253–270. http://dx.doi.org/10.1037/0033-295X.97.2.253. Luck, S.J., Hillyard, S.A., 1990. Electrophysiological evidence for parallel and serial processing during visual search. Percept. Psychophys. 48 (6), 603–617. http:// dx.doi.org/10.3758/BF03211606. Luck, S.J., Hillyard, S.A., 1994. Electrophysiological correlates of feature analysis during visual search. Psychophysiology 31, 291–308. Maier, M.E., Di Pellegrino, G., Steinhauser, M., 2012. Enhanced error-related negativity on flanker errors: error expectancy or error significance? Psychophysiology 49 (7), 899–908. http://dx.doi.org/10.1111/j.1469-8986.2012.01373.x. Maier, M.E., Steinhauser, M., 2013. Updating expected action outcome in the medial frontal cortex involves an evaluation of error type. J. Neurosci.: Off. J. Soc. Neurosci.

278