Trial‐By‐Trial Changes in Neural Indices of Performance Monitoring Uniquely Correspond to Behavioral Adjustments During a Flanker Task Authors: Miranda C. Lutz, Bohyun Park, Philippe Rast, Scott A. Baldwin, Michael J. Larson, Peter E. Clayson Published: 2026 ABSTRACT ABSTRACTBehavioral and neural indices of performance monitoring are key to understanding behavioral adaptation during task performance. However, associations between performance monitoring event‐related potentials (ERPs) and task behavior have been inconsistent. This inconsistency may partly reflect reliance on single‐subject averages that obscure trial‐by‐trial changes in ERPs and behavior, and a tendency to examine only one or two ERP indices at a time. Our objective was to uncover how neural variability during performance monitoring contributes to behavioral adaptation, revealing variability as a functional signature of cognitive control. We investigated whether current‐trial response times (RTs) and accuracy can be predicted from previous‐ and current‐trial congruency and accuracy and ERP indices of performance monitoring (N2, P3, error‐related negativity [ERN], error positivity, [Pe]). Flanker data from 291 healthy participants (54% female) were analyzed using multilevel location‐scale modeling. This modeling framework facilitates simultaneous examination of mean and variance relationships of single‐trial data. Previous‐ and current‐trial ERP amplitudes uniquely predict current‐trial RTs and accuracy, beyond previous‐ and current‐trial congruency and accuracy effects. Previous‐ and current‐trial N2, P3, ERN, and Pe were concurrently related to the mean and variance of RTs and to accuracy. The observed within‐person changes in the relationship between performance‐monitoring ERPs and task behavior indicate that trial‐by‐trial neural fluctuations reflect dynamic adjustments in cognitive control across successive actions. These findings demonstrate the value of modeling intraindividual variability in neurophysiological measures to understand adaptive behavior. Introduction Performance monitoring, a core component of cognitive control (Luna et al. 2015), is the ability to oversee and adapt one's actions to achieve goal‐directed behavior (Ullsperger et al. 2014). This ability allows for learning and behavioral change, making performance monitoring essential for the remarkable flexibility that healthy people demonstrate in the real world. Speeded forced‐choice response tasks are commonly used to examine performance monitoring by measuring accuracy, response times (RTs), and event‐related brain potentials (ERPs). Traditionally, psychophysiological research on performance monitoring focuses on between‐person differences in average scores of ERP components. However, this approach overlooks potentially informative trial‐by‐trial relationships between behavior and ERP components. Another limitation is that most studies have focused on a narrow set of performance monitoring indices, such as only examining stimulus‐ or response‐locked ERPs rather than multiple ERP components that index distinct aspects of performance monitoring. As a result, it is unclear which candidate ERP components are optimal for studying trial‐by‐trial adjustments in cognitive control. In the present study, we adopted a comprehensive approach to identify interindividual and intraindividual differences in ERP‐behavior relationships. The aims of this study were (1) to determine the correspondence between behavioral and neurophysiological indices of performance monitoring on a trial‐by‐trial basis during a flanker task and (2) to identify key ERP components that uniquely predict current‐trial RTs and accuracy. Trial‐By‐Trial Behavioral Adjustments Adaptive changes in task behavior, such as variations in reaction times, often occur when individuals monitor interference, detect conflicting information, or process decisions (Botvinick et al. 2001). These dynamics are commonly studied using the flanker task (Eriksen and Eriksen 1974), during which flankers either match (congruent) or oppose (incongruent) the target response. Typically, longer RTs and higher error rates are observed during difficult (incongruent) trials than during easy (congruent) trials. According to the conflict monitoring theory (Botvinick et al. 2001, 2004), once conflict is detected (e.g., incongruent trial), compensatory mechanisms adjust behavior by engaging adaptive control, resulting in improved performance. In general, RTs are faster, and error rates are lower when an incongruent trial follows another incongruent trial (high conflict) than when it follows a congruent trial (low conflict) (Larson et al. 2014). These trial‐by‐trial adjustments have been referred to as Gratton effects (Gratton et al. 1992), congruency sequence effects (Braem et al. 2019), and conflict adaptation effects (Clayson and Larson 2011a; Ullsperger et al. 2005). In short, it is important to incorporate congruency or sequence effects when examining trial‐to‐trial task performance. In addition to trial congruency, performance‐monitoring research highlights the role of accuracy in prompting changes to subsequent behavior. Several hypotheses address the significance of errors during performance monitoring (Larson et al. 2014). For instance, errors are threatening events (Weinberg et al. 2016) and are motivationally salient (Gehring et al. 1993). Incorrect choices are thought to signal the need to adjust behavior, typically to prevent future errors. However, post‐error accuracy (PEA; Danielmeier and Ullsperger 2011; Schroder and Moser 2014) and post‐correct accuracy (PCA; Schroder et al. 2020) behavioral adjustments are rarely examined. The magnitude of PEA appears to depend on factors such as error awareness (e.g., Klein et al. 2007) or response–stimulus intervals (Buzzell et al. 2017), and both PEA and PCA are sensitive to inter‐individual differences (Buzzell et al. 2019; Danielmeier and Ullsperger 2011). Therefore, the present study tested whether previous‐trial accuracy modulates trial‐by‐trial changes in RTs and accuracy, and whether these effects vary across individuals. A related behavioral metric is post‐error slowing (PES), which refers to the slowing of RTs after an error (Danielmeier and Ullsperger 2011; Rabbitt 1966). Larger PES is associated with improved task performance on subsequent trials, purportedly indicative of enhanced cognitive control (Dutilh et al. 2013). To date, studies typically have examined PES in isolation or compared it with a single behavioral index (e.g., Buzzell et al. 2017; Kirschner et al. 2021; Li et al. 2022). For example, Li et al. (2022) observed errors following congruent or incongruent trials consistently elicited PES, but accuracy adjustments depended on trial type, sequence effects, and previous‐trial accuracy. Understanding how multiple trial‐level variables jointly influence task performance elucidates how performance monitoring occurs in real time. The present study, therefore, examined the unique contributions of previous‐ and current‐trial congruency and accuracy on current‐trial RTs as well as accuracy within participants. Performance‐Monitoring ERPs Stimulus‐locked (N2, P3) and response‐locked (error‐related negativity [ERN], error positivity [Pe]) ERP components are widely used as indices of performance monitoring during flanker tasks. When confronted with conflicting stimuli, a negative deflection at 200–350 ms is often observed, here referred to as N2 (Botvinick et al. 2001; Ullsperger et al. 2014; Van Veen and Carter 2002; Yeung et al. 2004; Yeung and Cohen 2006). N2 is an index of conflict monitoring (Larson et al. 2014), representing a manifestation of anterior cingulate cortex (ACC) activity (Van Veen and Carter 2002; Yeung et al. 2004) and is sensitive to congruency sequence effects. N2 is smaller on incongruent trials preceded by another incongruent trial than on incongruent trials preceded by a congruent trial, putatively reflecting the engagement of cognitive control following an incongruent trial (e.g., Clayson and Larson 2011b, 2013b; Larson, Clayson, Keith, et al. 2016; Larson, Clayson, Kirwan, et al. 2016; Liu et al. 2018). The stimulus‐locked P3 follows N2 and appears as a centro‐parietal positive deflection roughly 350–500 ms after stimulus onset. The functional significance of P3 and its neural generator may be paradigm‐specific, but a flanker P3 has been linked to inhibitory control (Frühholz et al. 2011; Hervault et al. 2024; Wessel and Aron 2015) and adjustments in attention, with inferior frontal cortex often implicated in these processes (Nee et al. 2007; Polich 2007). The flanker P3 is larger on incongruent trials than on congruent trials. Similar to N2, flanker P3 is smaller for incongruent trials preceded by other incongruent trials than for incongruent trials preceded by congruent trials, suggesting that flanker P3 is sensitive to dynamic changes in cognitive control (Clayson and Larson 2011a, 2011b). ERN and Pe are two response‐related performance monitoring indices, reflecting distinct stages of error processing. ERN, a negative peak occurring 50‐100 ms after an incorrect response (Falkenstein et al. 1991; Gehring et al. 1993, 2018), is interpreted as a neural correlate of the conflict between the mental representation of a correct response and the observed error response (behavioral) (Larson et al. 2014; Yeung et al. 2004), although there is considerable debate around the functional significance of ERN (Clayson, Baldwin, and Larson 2021; Clayson, Baldwin, Rocha, et al. 2021; Clayson, Brush, and Hajcak 2021; Clayson, Carbine, et al. 2021; Clayson, Kappenman, et al. 2021). Pe is a positive deflection peaking between 200 and 400 ms (Falkenstein et al. 1991; Nieuwenhuis et al. 2001) and is regarded as the affective and/or conscious appraisal of the error, the awareness of the error, or awareness of a behavioral change after the error (Overbeek et al. 2005; Wessel 2018). Both ERN and Pe represent a manifestation of ACC activity (Ridderinkhof et al. 2004; Ullsperger et al. 2014; Van Veen and Carter 2002). ERN and Pe are sensitive to congruency sequence effects (Bartholow et al. 2005; Larson, Clayson, and Baldwin 2012). That is, ERN is larger following incongruent trials than following congruent trials (Larson, Clayson, and Baldwin 2012), which is consistent with the notion that larger ERN relates to greater attention (Larson and Clayson 2011) and task engagement (Tops and Boksem 2010; Tops et al. 2006). Similarly, congruency sequence effects have been observed for Pe, with smaller amplitudes following incongruent trials, but, unlike ERN, changes in amplitude were not specific to error trials (Larson, Clayson, and Baldwin 2012). Taken together, performance‐monitoring ERP components appear sensitive to trial‐by‐trial adjustments in cognitive control. However, it remains unclear whether and how performance monitoring ERPs relate to trial‐by‐trial adjustments in behavior. ERP‐Behavior Relationships The relationship between performance‐monitoring ERP components and task behavior is unclear, with considerable variability across findings and no consistent pattern emerging in the research to date (Cavanagh and Shackman 2015; Fischer et al. 2016; Kalfaoğlu et al. 2018; Larson, Clawson, et al. 2012; Larson, Clayson, and Baldwin 2012; Lin et al. 2022; Liu et al. 2018; Sandre et al. 2020; Valadez and Simons 2018). Some studies report associations between single‐trial ERP component amplitude and subsequent behavior (mostly with ERN: Beatty et al. 2020; Buzzell et al. 2019; Debener et al. 2005; Park et al. 2024; Steinhauser and Andersen 2019), whereas others fail to find any association (e.g., Buzzell et al. 2017; Kirschner et al. 2021; Li et al. 2022). The mixed results on the relationship between ERN and task‐related behavior (PES and PEA) may be due to ACC activity relating to many downstream instantiations of cognitive control, the indirect effect of ERN on behavior, as well as the method used to measure the ERPs (LoTemplio et al. 2023; Sandre et al. 2020). Consequently, the relationship between task behavior and performance‐monitoring ERPs requires further investigation. Studies of ERP‐behavior relationships typically focus on only one or two ERP components during previous or current trials and on only stimulus‐ or response‐related ERP components. Because different ERP components index distinct stages of cognitive control, examining each in isolation obscures how they jointly shape performance (Clayson 2024). Therefore, the present study used single‐trial models to examine whether N2, P3, ERN, and Pe amplitudes from previous and current trials predicted current‐trial RT and accuracy within participants. These models included previous‐ and current‐trial congruency and previous‐trial accuracy as covariates; models of RT also included current‐trial accuracy. By modeling behavioral and ERP indices simultaneously, we tested whether performance‐monitoring ERP component amplitudes explained unique variance in current‐trial behavior beyond congruency‐sequence effects and accuracy effects. Intraindividual Variability of Performance Monitoring Indices Views of variability and consistency in neurophysiological measures of performance have shifted. Intraindividual variability (the dynamic changes of the measure observed within a person) has generally been treated as ‘noise’ or ‘error’, resulting in ignoring the intraindividual variability or statistically adjusting for it. However, such variability can be meaningful: intraindividual variability has been related to various experiment‐related or cognitive processes (e.g., arousal; McGinley et al. 2015; Tops and Boksem 2011; response inhibition; Bellgrove et al. 2004) and to the heterogeneity of the sample, especially in psychiatric patients. Indeed, variability and the pattern of relationships between behavioral and neural variability have proven to be meaningful to understand adaptive changes in behavior (Garrett et al. 2013; Renart and Machens 2014; Waschke et al. 2021). Intraindividual variability in performance‐monitoring indices is often observed (e.g., Park et al. 2024; Clayson, Joshi, et al. 2022; Clayson, Rocha, et al. 2022; Egner 2011; Wang et al. 2014). That is, in these studies, robust intraindividual variability was observed for indices, like ERN and Pe, which point to the possibility of these dynamic changes within people representing key signals of interest. Interestingly, within‐subject performance could show a reversed direction of the ERP‐RTs relationship than between‐subject performance (e.g., for P3 and RT in LoTemplio et al. 2021). Furthermore, several studies have underscored the importance of disentangling mean scores and variability of performance monitoring in psychiatric samples, including studies with different forms of psychopathology (Clayson, Joshi, et al. 2022; Clayson, Rocha, et al. 2022; Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025; Kaiser et al. 2008), schizophrenia (Wootton et al. 2023), psychotic disorders (Holbrook et al. 2025), autism (Karalunas et al. 2014), and attention deficit hyperactivity disorder (Tamm et al. 2012). In many of these studies, the (increased) variability in RT was a predictor of cognitive outcomes, such as inhibitory control, attention, and performance monitoring, or psychiatric symptoms that are of importance in these psychopathologies. In healthy participants, within‐person task performance was slower and showed less variability in response to incongruent trials, but a larger intraindividual variability is more typically observed at a between‐person level (Williams et al. 2021). More variability in the ERN has also been related to larger ERN, suggesting heightened conflict processing (Clayson, Joshi, et al. 2022; Clayson, Rocha, et al. 2022). Although there are several sources of variation that can account for intraindividual variability (see Williams et al. 2021), the present study will identify how within‐person changes in mean ERP amplitudes and amplitude variability relate to subsequent changes in RTs and accuracy. Traditional multilevel models are well‐suited for examining trial‐by‐trial changes in behavioral and ERP data, because these models accommodate the nested nature of the proposed dataset, wherein individual trials are nested within participants (Volpert‐Esmond 2022; Volpert‐Esmond et al. 2021). Traditional multilevel models examine changes in outcomes based on fixed and random effects while assuming fixed residual variances across participants. Multilevel location‐scale models relax the assumption of fixed residual variance across participants used in traditional multilevel models (i.e., location only), simultaneously modeling the mean and variance structures of the dataset and the impact of covariates on each portion of the model (Walters et al. 2018). Furthermore, multilevel location‐scale models provide more robust parameter estimates against potential model misspecifications than traditional multilevel models (Leckie 2014; Walters et al. 2018) and have been successfully applied to RT (Park et al. 2024) and ERP datasets (Clayson, Baldwin, and Larson 2021; Clayson, Baldwin, Rocha, et al. 2021; Clayson, Brush, and Hajcak 2021; Clayson, Carbine, et al. 2021; Clayson, Kappenman, et al. 2021; Clayson, Joshi, et al. 2022; Clayson, Rocha, et al. 2022; Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025; Holbrook et al. 2025; Mon et al. 2024). Therefore, the present study used location‐scale models to examine the impact of previous‐ and current‐trial covariates on mean scores (location portion; e.g., do previous‐trial ERN scores predict mean changes in current‐trial RTs?) and variances (scale portion; e.g., do previous‐trial ERN scores predict variability in current‐trial RTs?). Present Study In sum, the primary aim of the current study was to investigate whether stimulus‐ and response‐locked ERP components recorded from the previous and current trial predict current‐trial behavior (RTs, accuracy) beyond previous‐ and current‐trial accuracy and congruency. We expected that changes in current‐trial behavior would be predicted by previous‐ and current‐trial accuracy and congruency. We hypothesized that ERP component amplitudes would improve the prediction of current‐trial behavior over and above previous‐ and current‐trial accuracy and congruency. Performance‐monitoring ERP component amplitudes (N2, P3, ERN, and Pe), congruency (congruent, incongruent), and accuracy (correct, error) were evaluated simultaneously within the same participants to clarify the nuanced inter‐ and intra‐individual relationships between behavioral and neurophysiological indices of performance monitoring. Method The present procedures and analyses were preregistered on Open Science Framework (OSF; https://osf.io/mzexu/). To test the relationships between trial‐level ERPs and behavior, 334 healthy participants who completed the same version of the flanker task from studies published by our lab were initially included (Baldwin et al. 2015; Clayson, Baldwin, and Larson 2021; Clayson, Baldwin, Rocha, et al. 2021; Clayson, Brush, and Hajcak 2021; Clayson, Carbine, et al. 2021; Clayson, Kappenman, et al. 2021; Clayson and Larson 2011a, 2011b; Larson and Clayson 2011; Larson, Clawson, et al. 2012; Larson, Clayson, and Baldwin 2012; Larson, Steffen, and Primosch 2013). The present analyses focused on behavioral adjustments and their relationship with previous and current‐trial ERP amplitudes, which were not reported in any of the cited studies. Participants Participants were recruited from undergraduate psychology courses and received course credit for participation. All participants provided written informed consent as approved by the local Institutional Review Board. Exclusion criteria for the original studies were assessed via participant self‐report and included current or previous diagnosis of a psychiatric disorder, psychoactive medication use, substance use or dependence, neurological disorders, head injury with loss of consciousness, left‐handedness, or uncorrected visual impairment. The preregistered sample size justification was to include all participants with at least five observations for each relevant RT and ERP condition (congruent trial, incongruent trial, correct trial, error trial, error trials for previous‐ and current‐trial combinations). This threshold was used because wider trial ranges improve the model stability and reliability of estimates. Moreover, this threshold was based on simulation studies that indicate multilevel location‐scale models yield sufficient power (> 0.80), good coverage, and minimal bias with as few as 10 observations per person for small effect sizes in a sample (n = 25) much smaller than that of the present study (see documentation for Leckie 2014). Following artifact rejection (see Electrophysiological Data Recording and Reduction section), 291 participants (160 women) were included in the present analyses. Participants were an average of 21 years old (SD = 3.2) with an average of 14 years of education (SD = 1.5). Age and education information were missing for one participant. Experimental Task Participants completed a modified version of the Eriksen flanker task (Eriksen and Eriksen 1974). Each trial comprised a congruent (<<<<<, >>>>>) or incongruent (<<><<, >><>>) stimulus presented in white on a black background of a 17 in. computer monitor approximately 20 in. from the participant's head. Participants were instructed to respond as quickly and accurately as possible with a right‐hand key press. An index‐finger button press was used if the middle arrow (i.e., the target stimulus) pointed to the left and a middle‐finger button press was used if the middle arrow pointed to the right. Flanker stimuli were presented for 100 ms prior to the onset of the target stimulus, which remained on the screen for 600 ms. The response window was 1600 ms. If the participant responded after 1600 ms, the trial was counted as an error of omission. To decrease expectancy effects, the inter‐trial interval (ITI) varied pseudo‐randomly between 800 ms, 1000 ms, and 1200 ms. Three blocks of 300 trials (900 total trials) were presented; 45% of trials within a block were congruent and 55% of trials were incongruent. Participants completed 24 practice trials prior to beginning the experimental task to ensure adequate understanding of instructions. Electrophysiological Data Recording and Reduction EEG was recorded from 128 passive scalp sites (Ag/AgCl) using a 129‐channel hydrocel geodesic sensor net and Electrical Geodesics Inc. (EGI; Eugene, OR) amplifier system (20 K nominal gain, delta‐sigma sinc lowpass filter with a half‐power cutoff of 2000 Hz). The sensor layout of the hydrocel net is shown in Clayson et al. (2011). EEG was online referenced to the vertex electrode, and a ground sensor was located at PCz. EEG was digitized continuously at 250 Hz with a 24‐bit analog‐to‐digital converter. Impedances were kept below 50 kΩ. Data processing steps are based on the optimized pipeline for ERN recorded during a flanker task (Clayson, Baldwin, and Larson 2021; Clayson, Baldwin, Rocha, et al. 2021; Clayson, Brush, and Hajcak 2021; Clayson, Carbine, et al. 2021; Clayson, Kappenman, et al. 2021). Continuous EEG was filtered offline using a sixth‐order (36 dB/oct) infinite impulse response (IIR) Butterworth filter with half‐amplitude cutoffs at 0.01 and 15 Hz implemented in ERPLab v8.02 (Lopez‐Calderon and Luck 2014). Response‐locked epochs were extracted separately for each participant using a temporal window from 400 ms prior to the participant's button press to 800 ms following a button press. Stimulus‐locked epochs were extracted separately for each participant using a temporal window from 250 ms prior to the stimulus to 800 ms following the presentation of the stimulus. Eye blinks and horizontal and vertical saccadic eye movements were removed from the segmented waveforms using independent components analysis (ICA). For the ICA procedure, epoched EEG data from all channels were processed through a binary version of EEGLab's runica function called binica (Delorme and Makeig 2004). Any ICA components that correlated at 0.9 or above with the scalp topography of a blink template or at 0.8 or above with the scalp topography of vertical and horizontal saccade templates were removed from the data. Artifact templates for blinks and saccades were created by the present authors using the collected data and automatically generated by the ERP PCA Toolkit (Dien 2010). Following artifact correction, channels with more than a 100 μV step within 100 ms intervals, a voltage difference of 300 μV through the duration of the epoch, or an absolute correlation with the nearest six neighboring channels that fell below 0.4 were marked as bad for the epoch. Channels marked as bad for more than 20% of epochs were considered globally bad, and bad channels were interpolated using spherical splines (Perrin et al. 1989). If more than 10% of channels were marked bad for an epoch, then the entire epoch was rejected. EEG was then rereferenced to an average reference scheme, and the period from 400 to 200 ms prior to participant response was used for baseline adjustment. Response‐locked and stimulus‐locked ERPs were analyzed. To reduce the biasing effects of background EEG noise on ERP measurements, ERPs were scored using a time‐window mean amplitude approach (Clayson et al. 2013; Luck 2014). Regions of interest (ROIs), which are created by averaging data across multiple electrodes, were used to provide improved signal‐to‐noise ratio over data from a single electrode (Baldwin et al. 2015; Huffmeijer et al. 2010). Time windows and sensors used for scoring were preregistered and based on other studies of cognitive control from our lab (Clawson et al. 2013; Clayson and Larson 2011a, 2013b; Larson, Clawson, et al. 2013; Larson, Clayson, Keith, et al. 2016; Larson, Clayson, Kirwan, et al. 2016). ERN (and correct‐trial negativity [CRN], the correct‐trial counterpart of ERN) was extracted from response‐locked epochs as the average activity over fronto‐medial sites (6 [FCz], 129 [Cz], 7, 106; for electrode configuration, see Clayson and Larson 2013a) and quantified using a time‐window mean amplitude (average activity from 0 to 100 ms). Pe was extracted from response‐locked epochs as the average activity of centro‐parietal sites (54, 55 [CPz], 61, 62 [Pz], 78, 79) and scored using a time‐window mean amplitude from 200 to 400 ms. Stimulus‐locked N2 was scored over fronto‐medial sites (FCz, Cz, 7, 106) using a time‐window mean amplitude from 300 to 350 ms, and stimulus‐locked P3 was scored over central sites (Cz, 31, CPz, 80) using a time‐window mean amplitude from 350 to 500 ms. It was proposed in the preregistration to use a collapsed localizer approach if these time windows did not capture ERP component activity of interest, but the time windows appeared appropriate based on visual inspection of the data, so no changes were made. Psychometric Internal Consistency Estimates of group‐level psychometric internal consistency were examined, which is recommended as routine practice (Clayson 2024, 2025; Clayson and Miller 2017b; Rocha et al. 2026). Participants were not excluded based on estimates of internal consistency because the focus of the present study was on within‐person relationships, not between‐person relationships. Generalizability theory was used to calculate dependability (ϕ), which is a measure of internal consistency analogous to coefficient alpha from classical test theory (see Baldwin et al. 2015; Brennan 2001; Clayson, Baldwin, et al. 2021; Clayson, Kappenman, Gehring, et al. 2021; Shavelson and Webb 1991). Dependability estimates for average ERP component amplitude, response time, and mean accuracy as a function of event (current‐trial: correct, error; congruent, incongruent) were examined using the ERP Reliability Analysis (ERA) Toolbox (Clayson, Baldwin, and Larson 2021; Clayson and Miller 2017a). Data Analysis Model Estimation of RTs Multilevel location‐scale models were used to examine trial‐level data for RT analyses (Walters et al. 2018; Williams et al. 2022, 2021), and these models have successfully been applied to other ERN datasets (Clayson, Kappenman, Gehring, et al. 2021; Clayson, Joshi, et al. 2022; Clayson, Varkala, et al. 2025; Clayson, Rocha, et al. 2022; Clayson, Mcdonald, et al. 2025; Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025). Multilevel models are well suited for ERP data (Volpert‐Esmond 2022; Volpert‐Esmond et al. 2021) and accommodate the nested nature of the proposed dataset (i.e., trials nested within participants). Unlike in traditional multilevel models, which assume fixed residual variance across participants (i.e., location only), location‐scale models relax this assumption and simultaneously model the mean and variance data structures (Leckie 2014; Walters et al. 2018). Multilevel location‐scale models allow for the examination of ERP‐RT relationships. For instance: for the location portion of the model: do ERN scores predict mean post‐error RTs? On the scale portion of the model: can ERN scores predict variability in post‐error RT? All current‐trial RT models included a population intercept, fixed effects for the Previous‐Trial Congruency × Current‐Trial Congruency interaction, fixed effects for Previous‐Trial Accuracy × Current‐Trial Accuracy interaction, fixed effects for constituent main effects, random intercepts for participants, and a random slope for current‐trial congruency. Congruency included two levels (congruent [reference level], incongruent), and accuracy included two levels (error [reference level], correct) for previous‐ and current‐trials. Previous‐trial referred to the congruency or accuracy of the previous trial, and current‐trial referred to the congruency or accuracy of the current trial. The dependent variable for all models was the current‐trial RT. This base model is described below using Wilkinson notation for transparency (Wilkinson and Rogers 1973), and the location and scale portions of the model included the same predictors. RTBase Model:Current−trialRT~1+Previous−Trial Congruency+Current−Trial Congruency+Previous−TrialCongruency*Current−Trial Congruency+Previous−Trial Accuracy+Current−Trial Accuracy+Previous−Trial Accuracy*Current−Trial Accuracy+1+Current−Trial CongruencyParticipant. All unique combinations of the following independent variables were evaluated as fixed effects added onto the Base Model described above with two constraints: (1) if an interaction was included in the model, then the constituent main effects were also included, and (2) the predictors on the location and scale portion were the same. The tested models included Previous‐Trial Accuracy × Previous‐Trial N2, Previous‐Trial Accuracy × Previous‐Trial P3, Previous‐Trial Accuracy × Previous‐Trial ERN, Previous‐Trial Accuracy × Previous‐Trial Pe, Current‐Trial Accuracy × Current‐Trial N2, Current‐Trial Accuracy × Current‐Trial P3, Current‐Trial Accuracy × Current‐Trial ERN, and Current‐Trial Accuracy × Current‐Trial Pe. All ERP component amplitudes were entered as continuous and were standardized (z‐score transformed) to a grand mean of 0 and a standard deviation of 1. A total of 6561 unique combinations of predictors were evaluated. To address the resulting high‐dimensional model space and avoid overfitting, we employed a Bayesian hierarchical model utilizing the regularized horseshoe (HS) prior, following the approach outlined by Peltola et al. (2014), as well as the parametrization described in Piironen and Vehtari (2017). The regularized horseshoe model used in our analyses incorporates both location and scale parameters, with separate regularized horseshoe priors applied to each. The model structure is designed to manage the inherent complexities in ERP and RT datasets, accommodating the sparse nature of significant predictors while controlling for potential overfitting. The HS prior consists of a mix of parameters that govern the shrinkage by combining a global parameter τ that pulls all the weights of the parameters globally toward zero, and local slab parameters λi that allow some coefficients to escape this shrinkage due to their heavy‐tailed distribution. The local parameters λi are derived from components that include a standard normal and an inverse gamma influenced parameter, while the global parameter τ is based on a Cauchy (0, 2.5) prior. This balance enables the model to aggressively suppress noise while retaining signals of potential importance. Bayesian models were fit within R (R Core Team 2021) using cmdstanr (Gabry et al. 2024), which is a front‐end wrapper for Stan (Stan Development Team 2021). Model Estimation of Accuracy Accuracy was analyzed using multilevel logistic regression models with a Bernoulli‐distributed outcome and a logit link. In such analyses, there is no scale portion of the model to estimate, so these models examined only mean accuracy scores (but see Hedeker et al. 2009). All current‐trial accuracy models included a population intercept, fixed effects for previous‐trial accuracy, fixed effects for the Previous‐Trial Congruency × Current‐Trial Congruency interaction, fixed effects for constituent main effects, random intercepts for participants, and a random slope for current‐trial congruency. The dependent variable for all models was the current‐trial accuracy (0, incorrect; 1, correct). This base model is described below. Accuracy Base Model:Current−trial Accuracy~1+Previous−Trial Congruency+Current−Trial Congruency+Previous−TrialCongruency*Current−Trial Congruency+Previous−Trial Accuracy+1+Current−Trial CongruencyParticipant. A different set of predictors from the RT analyses was used for estimating current‐trial accuracy because RT analyses included current‐trial accuracy as a predictor. Therefore, unique combinations of the following predictors were evaluated as fixed effects added onto the Base Model, and constituent main effects were included for all interaction analyses. The tested models included Previous‐Trial Accuracy × Previous‐Trial N2, Previous‐Trial Accuracy × Previous‐Trial P3, Previous‐Trial Accuracy × Previous‐Trial ERN, Previous‐Trial Accuracy × Previous‐Trial Pe, Current‐Trial N2, Current‐Trial P3, Current‐Trial ERN, and Current‐Trial Pe. All ERP component amplitudes were entered as continuous and were standardized (z‐score transformed) to a grand mean of 0 and a standard deviation of 1. A total of 1296 unique combinations of predictors were evaluated. Bayesian hierarchical models using the regularized HS prior and were fit within Stan. Model Comparisons Models of current‐trial RT and of current‐trial accuracy were separately compared based on predictive accuracy. Pareto smoothed importance sampling (PSIS‐LOO) was used to compare model fits (Vehtari et al. 2016) using the R package loo (Vehtari et al. 2020). Conceptually speaking, PSIS‐LOO compares the predictive accuracy of models, which provides an estimate of how the model would perform on a new, unseen dataset. A given model is expected to show better fit over another when (1) the difference in expected log predictive density (elpd`loo) is greater than 4 and (2) the difference in elpd`loo is greater than 2 times the standard error of elpd`loo (SE (elpd`loo)). Two models are considered to have comparable fits when (1) the difference in elpd`loo is less than 4 or (2) the difference in elpd`loo is less than 2 times the SE (elpd`loo) (Vehtari et al. 2016). If the two models show similar fit, then the model with fewer predictors will be considered optimal to avoid overfitting. Then, post‐estimation contrasts on the posterior parameter estimates were used to characterize the effects of covariates on the dependent variables. Deviations From Preregistration To improve the efficiency of fitting models, a population intercept was included, but this has no functional impact on the models themselves—only on the interpretation of parameter estimates. There were no additional deviations from preregistration. Results Summary data for RTs and ERP component amplitudes are shown in Table 1. Overall accuracy of current congruent trials was 97% (SD: 7%) and for incongruent trials: 89% (SD 12%). Grand average waveforms are shown in Figure 1. All continuous variables, including RTs, were z‐scored transformed (M = 0, SD = 1) prior to model fitting. TABLE 1Summary statistics for response times (ms), and event‐related potential Amplitude (μV) data.Current congruentCurrent incongruent M (SD) ϕ (95% CrI) M (SD) ϕ (95% CrI)Current correct trialRT375 (30)0.99 (0.99, 0.99)443 (31)0.99 (0.99, 0.99)ERN2.2 (1.6)0.98 (0.97, 0.98)2.6 (2.0)0.98 (0.98, 0.99)Pe−0.2 (1.2)0.96 (0.95, 0.97)−0.3 (1.4)0.97 (0.97, 0.98)N21.9 (1.7)0.98 (0.97, 0.98)0.8 (1.9)0.98 (0.98, 0.98)P33.0 (1.6)0.97 (0.97, 0.98)3.6 (2.0)0.98 (0.98, 0.99)Current error trialRT390 (53)0.75 (0.70, 0.80)324 (39)0.94 (0.93, 0.95)ERN−0.4 (3.1)0.64 (0.56, 0.71)−0.8 (2.4)0.91 (0.90, 0.93)Pe3.6 (3.3)0.75 (0.69, 0.80)3.9 (2.8)0.94 (0.93, 0.95)N20.9 (3.2)0.63 (0.54, 0.70)−0.3 (2.9)0.92 (0.90, 0.93)P31.6 (3.1)0.58 (0.48, 0.66)2.5 (2.6)0.92 (0.91, 0.93)Previous correctPrevious error M (SD) ϕ (95% CrI) M (SD) ϕ (95% CrI)Current correct trialRT409 (30)0.99 (0.99, 0.99)418 (34)0.90 (0.88, 0.91)ERN2.5 (1.7)0.99 (0.99, 0.99)2.3 (1.8)0.85 (0.82, 0.87)Pe−0.3 (1.2)0.98 (0.98, 0.98)−0.3 (1.5)0.78 (0.74, 0.81)N21.3 (1.7)0.99 (0.98, 0.99)1.2 (1.7)0.81 (0.77, 0.84)P33.3 (1.8)0.99 (0.99, 0.99)2.6 (1.8)0.83 (0.80, 0.86)Current error trialRT334 (38)0.92 (0.91, 0.93)351 (68)0.88 (0.85, 0.91)ERN−0.9 (2.3)0.90 (0.88, 0.92)−0.1 (3.3)0.81 (0.75, 0.86)Pe4.0 (2.7)0.93 (0.92, 0.94)1.7 (3.2)0.71 (0.61, 0.79)N2−0.1 (2.7)0.90 (0.89, 0.92)0.4 (3.5)0.71 (0.61, 0.79)P32.3 (2.5)0.91 (0.89, 0.93)2.0 (3.2)0.68 (0.56, 0.77)Abbreviations: 95% CrI, 95% credible interval; ERN, error‐related negativity; Pe, error positivity; RT, response time; ϕ, dependability (estimate of internal consistency). Summary statistics for response times (ms), and event‐related potential Amplitude (μV) data. Abbreviations: 95% CrI, 95% credible interval; ERN, error‐related negativity; Pe, error positivity; RT, response time; ϕ, dependability (estimate of internal consistency). Grand average stimulus‐locked (top) and response‐locked (bottom) waveforms. N2 waveforms are shown in the top left and represent activity averaged across four fronto‐central sites. P3 waveforms are shown in the top right and represent activity averaged across four central sites. Error‐related negativity (ERN) waveforms are shown in the bottom left and represent activity averaged across four fronto‐central sites. Error positivity (Pe) waveforms are shown in the bottom right and represent activity averaged across six centro‐parietal sites. Response Times A total of 6561 unique combinations of predictors were compared using PSIS‐LOO, and a best‐fitting model with the 20 predictors on the location and scale portions of the model was identified. The 20 included predictors were the Previous‐Trial Congruency × Current‐Trial Congruency interaction, Previous‐Trial Accuracy × Current‐Trial Accuracy interaction, Previous‐Trial Accuracy × Previous‐Trial N2 interaction, Previous‐Trial Accuracy × Previous‐Trial P3, the main effects of previous‐trial ERN and previous‐trial Pe, Current‐Trial Accuracy × Current‐Trial N2, Current‐Trial Accuracy × Current‐Trial P3, Current‐Trial Accuracy × Current‐Trial Pe, and the main effect of current‐trial ERN (constituent main effects of each interaction were included). The best‐fitting model is described in detail below, and all parameter estimates are shown in the Supporting Information. The regularized HS prior was used to accommodate the sparse nature of significant predictors while mitigating the risk of potential overfitting. The estimated posterior density curves of fixed effects and their associated uncertainty intervals for the location portion of the model are shown in Figure 2, and the scale portion effects of the model are shown in Figure 3. These plots help to convey the relative impact of different fixed effects on the predicted RTs. For example, congruency effects and accuracy effects have a numerically larger impact on RTs than predictors for ERP component scores on the location portion of the model. Figure 4 shows the point estimates (average) of all estimated relationships in the RT model to allow a visual comparison of the impact of the effects of ERP components, congruency, and accuracy on location and scale portions of the model. The estimated posterior density curves of fixed effects and their associated uncertainty intervals for the location portion of the response time (RT) model, with the 95% credible intervals shaded around the posterior mean. Continuous fixed effects and RTs were z‐score transformed. The estimated posterior density curves of fixed effects and their associated uncertainty intervals for the scale portion of the response time (RT) model, with the 95% credible intervals shaded around the posterior mean. Continuous fixed effects and RTs were z‐score transformed. The relationships between all fixed effect estimates from both the location (A, B) and scale (C, D) portions of the model predicting current‐trial response times (RTs). The columns differentiate whether the effects are associated with the previous or current trial. In panels A and C, when “correct” or “error” appears next to a fixed effect, it indicates that the model included a Component × Accuracy interaction. If “correct” or “error” is not shown, the model included only that specific component amplitude as a predictor, without the interaction effect. In panels A and C there were no interactions between effects from previous trials and effects from current trials. However, in panels B and D, all effects represent Previous‐Trial × Current‐Trial interactions. For example, the model included a Previous‐Trial Accuracy × Current‐Trial Accuracy interaction, and the effects of current‐trial accuracy are shown as a function of the accuracy of the previous trial, as indicated in the preceding column. Lines are weighted according to the size of the point estimate (average) printed on top of each line. Red lines indicate negative parameter estimates, and blue lines indicate positive parameter estimates. For N2 and CRN/ERN, larger amplitudes indicate more negative values; therefore, positive and negative parameter estimates should be interpreted in relation to this scoring direction. Gray lines are used when the 95% credible interval included zero. ERN, error‐related negativity; Pe, error positivity. Trial Congruency, Accuracy and Response Times Congruency sequence effects were observed in z‐scored RTs on the location portion of the model (see Clayson and Larson 2011a, 2011b; Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025), such that congruent trials preceded by congruent trials (cC trials) were fastest (95% credible interval [CrI]: −0.53, −0.44) and became slower in the following order: congruent trials preceded by incongruent trials (iC trials; 95% CrI: −0.28, −0.20), incongruent trials preceded by incongruent trials (iI trials; 95% CrI: 0.44, 0.52), then incongruent trials preceded by congruent trials (cI trials; 95% CrI: 0.49, 0.57). A congruency sequence effect index was calculated using the following equation: (cI—cC)—(iI—iC) (see Clayson and Larson 2012; Nieuwenhuis et al. 2006; van Steenbergen et al. 2010), and a robust congruency sequence effect was observed for z‐scored RTs (95% CrI: 0.28, 0.31). For the scale portion of the model, cI (95% CrI: −0.43, −0.39) trials and iI trials (95% CrI: −0.43, −0.39) were related to the least variability, and the variability was similar across cI and iI trials (95% CrI: −0.01, 0.01). Variability increased in the order of cC trials (95% CrI: −0.25, −0.21), then iC trials (95% CrI: −0.22, −0.19). Congruency sequence effects were also observed on the scale portion of the model (95% CrI: 0.01, 0.03). The fastest RTs were for error trials preceded by correct trials (95% CrI: −1.89, −1.80), then slowed in speed in the order of error trials preceded by error trials (95% CrI: −1.21, −1.08), correct trials preceded by other correct trials (95% CrI: −0.53, −0.44), and correct trials preceded by error trials (95% CrI: −0.47, −0.38). Post‐error slowing was observed for current correct trials (95% CrI: 0.04, 0.06) and for current error trials (95% CrI: 0.65, 0.74). When the current trial was correct, RT variability was smaller when preceded by a correct trial (95% CrI: −0.71, −0.60) than when preceded by an incorrect trial (95% CrI: −0.17, −0.13; contrast: 95% CrI: 0.45, 0.56). When the current trial was incorrect, RT variability was similar regardless of whether the previous trial was correct (95% CrI: 0.13, 0.18) or incorrect (95% CrI: 0.11, 0.17; contrast: 95% CrI: −0.01, 0.05). Stimulus‐Related ERPs and Response Times Larger previous, error‐trial N2 (i.e., more negative) was related to shorter current‐trial RTs (95% CrI: 0.02, 0.04). This effect was smaller when the previous trial was correct (95% CrI: 0.004, 0.01; contrast: 95% CrI: −0.03, −0.01). Similarly, when the previous trial was incorrect, larger previous‐trial N2 related to smaller RT variability (95% CrI: 0.02, 0.04), and this effect was smaller when the previous trial was correct (95% CrI: 0.004, 0.01; contrast: 95% CrI: −0.04, −0.01). When the current‐trial was an error, larger current‐trial N2 related to shorter RTs (95% CrI: 0.12, 0.16), but an opposite pattern was observed for correct trials. Larger current‐trial N2 related to longer RTs (95% CrI: −0.04, −0.03; contrast 95% CrI: −0.19, −0.16). For error trials, larger current‐trial N2 related to smaller RT variability (95% CrI: 0.06, 0.08). For correct trials, the effect was smaller (95% CrI: 0.02, 0.03; contrast: 95% CrI: −0.06, −0.03). Larger previous, error‐trial P3 related to shorter current‐trial RTs (95% CrI: −0.05, −0.03), and this effect was smaller when the previous‐trial was correct (95% CrI: −0.01, −0.005; contrast 95% CrI: 0.02, 0.04). When the previous trial was correct, larger previous‐trial P3 related to less RT variability (95% CrI: −0.013, −0.005). However, this effect was not observed when the previous trial was an error (95% CrI: −0.02, 0.002), but the contrast indicated similar effects following previous‐correct and previous‐error trials (95% CrI: −0.01, 0.01). Larger current, error‐trial P3 related to shorter RTs (95% CrI: −0.15, −0.12). The effect was smaller for correct trials (95% CrI: −0.023, −0.016; contrast 95% CrI: 0.10, 0.13). Larger current, error‐trial P3 related to less variability in RTs (95% CrI: −0.11, −0.09), and the effect was smaller for current correct trials (95% CrI: −0.06, −0.05; contrast 95% CrI: 0.03, 0.06). Response‐Related Potentials and Response Times A larger previous‐trial ERN (i.e., more negative) related to longer current‐trial RTs (95% CrI: −0.012, −0.005) and smaller RT variability (95% CrI: 0.003, 0.01). Current‐trial ERN was not related to RTs for error trials (95% CrI: −0.03, 0.01). However, larger current‐trial CRN was related to longer RTs (95% CrI: −0.07, −0.06; contrast 95% CrI: −0.08, −0.04). When the current trial was an error, the larger current‐trial ERN was related to smaller RT variability (95% CrI: 0.02, 0.04), but the opposite pattern was observed for correct trials. Larger current CRN related to greater RT variability (95% CrI: −0.075, −0.068; contrast 95% CrI: −0.12, −0.09). Larger previous‐trial Pe related to longer RTs (95% CrI: 0.002, 0.01), but previous‐trial Pe was unrelated to RT variability (95% CrI: −0.001, 0.005). Larger current, error‐trial Pe was related to faster RTs (95% CrI: −0.07, −0.04), but on correct trials, larger post‐correct positivity (Pc; correct‐trial counterpart to Pe) was related to longer RTs (95% CrI: 0.003, 0.01; contrast: 95% CrI: 0.05, 0.07). When the current‐trial was an error, larger current‐trial Pe related to less RT variability (95% CrI: −0.06, −0.04), but when the current‐trial was correct, larger current‐trial Pc related to greater RT variability (95% CrI: 0.03, 0.04; contrast 95% CrI: 0.07, 0.09). Interim Summary Response Times Regarding previous‐trial ERP components, larger previous‐trial N2 and P3 related to shorter current‐trial RTs, and larger previous‐trial ERN and Pe related to longer current‐trial RTs. For current‐trial ERP components, larger error‐trial N2 related to faster RTs, whereas larger correct‐trial N2 related to longer RTs. During error trials, larger P3 and Pe were related to shorter RTs. While for correct trials, this effect is smaller for P3 and reversed for Pc, which was related to longer RTs. Additionally, during correct trials, larger CRN related to longer RTs, and the same effect was found for Pc. No RT‐ERN relation was found during current error trials, but for Pe, faster RTs were observed. Regarding the impact of previous‐trial ERP components on variability of current‐trial RTs, larger previous‐trial N2 was related to smaller RT variability following error trials, and the effect was smaller following correct trials. Larger previous‐trial P3 following correct responses related to less RT variability, and larger previous‐trial ERN related to smaller RT variability. For current‐trial ERP components, both larger N2 and larger P3 related to reduced RT variability. Larger error‐trial ERN related to smaller RT variability, but larger correct‐trial CRN related to greater variability. Larger error‐trial Pe related to less RT variability, but larger correct‐trial Pc related to greater RT variability. Accuracy A total of 1296 unique combinations of predictors were compared, and the best‐fitting model included 16 fixed effects. The included fixed effects were a Previous‐Trial Congruency × Current‐Trial Congruency interaction; main effects of previous‐trial accuracy, P3, and ERN; and main effects of current‐trial N2, P3, ERN, and Pe. The best‐fitting models are described in detail below, and all parameter estimates are shown in the Supporting Information. Similar to the RT models, the regularized HS prior was used to accommodate the sparse nature of significant predictors and to mitigate the risk of potential overfitting. The estimated posterior density curves of fixed effects and their associated uncertainty intervals are shown in Figure 5. Figure 6 shows the point estimates (average) of all estimated relationships in the model to allow a visual comparison of the impact of the effects of ERP components and congruency on current‐trial accuracy. The estimated posterior density curves of fixed effects and their associated uncertainty intervals for the accuracy model, with the 95% credible intervals shaded around the posterior mean. Continuous fixed effects were z‐score transformed. The relationships between all fixed effect estimates from the model predicting current‐trial accuracy. The columns differentiate whether the effects are associated with the previous or current trial. Panel A shows ERP component effects, and panel B shows the Previous‐Trial Congruency x Current‐Trial Congruency interaction. In panel A there were no interactions between effects from previous trials and effects from current trials. However, in panel B, the model included a Previous‐Trial Congruency × Current‐Trial Congruency interaction, and the effects of current‐trial congruency are shown as a function of the congruency of the previous trial, as indicated in the preceding column. Lines are weighted based on the size of the point estimate (average) printed on top of each line. Red lines indicate negative parameter estimates, blue lines indicate positive estimates, and gray lines are used when the 95% credible interval included zero. ERN, error‐related negativity; Pe, error positivity. Trial Congruency and Accuracy Accuracy was lower for current incongruent trials than for current congruent trials (95% CrI: −2.26, −1.99). Evidence of congruency sequence effects was observed. For current incongruent trials, accuracy was lower when preceded by a congruent trial than when preceded by an incongruent trial (95% CrI: −0.94, −0.85), but accuracy was similar across congruent trials preceded by congruent or incongruent trials (95% CrI: −0.05, 0.10). A robust congruency sequence effect, (cI—cC)—(iI—iC) was observed (95% CrI: −0.95, −0.78). A current trial was more likely to be correct when the previous trial was correct (95% CrI: 0.96, 1.08). Stimulus‐Related ERPs and Accuracy Larger current‐trial N2 (i.e., more negative) was related to better current‐trial accuracy (95% CrI: −0.09, −0.05). Larger current‐trial P3 related to higher current‐trial accuracy (95% CrI: 0.10, 0.15). Larger previous‐trial P3 was not related to current‐trial accuracy (95% CrI: −0.04, 0.01). Response‐Related ERPs and Accuracy Larger previous‐trial ERN (i.e., more negative) related to improved accuracy on the subsequent trial (95% CrI: −0.16, −0.12). Larger current‐trial ERN (i.e., more negative) related to lower accuracy (95% CrI: 0.75, 0.80). Larger current‐trial Pe was related to lower accuracy (95% CrI: −0.93, −0.89). Interim Summary Accuracy As predicted, ERP component amplitudes uniquely predict accuracy over and above congruency and accuracy effects. Larger current‐trial N2 and P3 were related to higher accuracy, larger previous‐trial ERN related to improved accuracy, and larger current‐trial ERN and Pe were related to lower accuracy. Discussion The aim of this study was to determine whether previous‐ and current‐trial stimulus‐ and response‐locked ERP components uniquely predict current‐trial RT and accuracy over and above previous‐ and current‐trial accuracy and congruency. Consistent with expectations, current‐trial RT means and variances were uniquely predicted by previous‐ and current‐trial accruacy, congruency, and ERP component amplitudes, and current‐trial accuracy was uniquely predicted by previous‐trial accuracy, congruency, and selected ERP component amplitudes. Specifically, the final model identified 20 predictors for the means (location) and variances (scale) of current‐trial RTs and 16 predictors contributing to current‐trial accuracy. The present findings highlight the value of modeling both location and scale when evaluating ERP‐behavior relationships during performance monitoring by demonstrating that amplitude variability, not just shifts in average amplitudes, is systematically linked to adjustments in behavior. Predicting Current‐Trial Response Times In the current study, we observed the traditional congruency sequence effects, where both previous‐trial and current‐trial congruency predicted current‐trial RT averages and variances. 1 This has been observed in similar flanker paradigms (Clawson et al. 2013; Larson, Clayson, Kirwan, et al. 2016; Larson, Clayson, Keith, et al. 2016; Clayson et al. 2013; Clayson and Larson 2013a, 2013b; Von Gunten et al. 2018; Ullsperger et al. 2005). Participants varied most in their RTs when the current trial was congruent, regardless of previous‐trial congruency, which contradicts a behavioral study by Williams et al. (2021). The discrepancy is likely due to the inclusion of many covariates, as Williams et al. only examined RTs. A typical PES was observed, and previous‐ and current‐trial accuracy predicted current‐trial RT averages and variances. Differences in RT variability following current‐trial congruency and previous‐trial accuracy reflect the real‐time adaptive behavioral component of conflict processing; participants differed in the time needed to process conflicting stimuli and ensure accurate responses. Larger previous‐trial N2 was related to shorter and less variable current‐trial RTs, and larger current‐trial N2 was related to less RT variability. However, the direction of the association with mean RT differed by current‐trial accuracy, such that larger current‐trial N2 was related to shorter RTs on error trials but longer RTs on correct trials. This N2‐RT pattern reflects heightened conflict detection, which may contribute to same‐trial adjustments in cognitive control (Silton et al. 2010) and to a more efficient evidence accumulation on the same and subsequent trials (Nunez et al. 2017). Furthermore, the commission of an error influences the relationship between N2 amplitude and RT averages and variances. The relationships between current‐trial N2 amplitude and RT variability were similar across error and correct trials, but the relationships with mean RT differed by current‐trial accuracy. For current error trials, larger N2 amplitudes were related to faster RTs and less variance in RTs, which indicates that N2 is related to conflict monitoring of the stimulus as well as behavior (Clayson and Larson 2011a; Hoyniak and Petersen 2019). For flanker P3, effects were more consistently in one direction; both previous‐ and current‐trial amplitudes predicted shorter and less variable RTs. As P3 may be indicative of inhibitory control (Hervault et al. 2024; Wessel and Aron 2015), our results suggest that flanker P3 is involved in dynamic changes in cognitive control, similar to N2 and consistent with Clayson and Larson (2011a, 2011b). For error‐related ERPs, we found two distinct patterns. First, a relationship between larger amplitudes and longer RTs was observed for previous‐trial ERN and Pe and current‐trial CRN and Pc. This result could be interpreted as a participant still experiencing the affective effect of the error from the previous trial and consequently allocating attention to respond quickly and correctly on the current trial (Dali and Hester 2022). This implies that error detection or awareness from previous trials may exert more cognitive control to “correct” behaviors on the current trial, as ERN reflects error detection (Ullsperger et al. 2014) and affective processing (Dignath et al. 2020), whereas Pe is associated with later error awareness (Kirschner et al. 2021). Second, for current correct trials, a larger CRN predicted longer and more variable RTs, and a larger Pc predicted greater RT variability. This suggests that even in the absence of errors, recruitment of error‐monitoring processes may signal internal conflict or uncertainty, leading to more cautious and inconsistent responding. For current error trials, larger ERN was not related to mean RTs but was related to less RT variability, whereas a larger Pe predicted faster and less variable RTs. This suggests that when errors are consciously registered (as reflected in Pe), participants may engage in more efficient post‐error adjustments, resulting in quicker and more stable responses. The unique relations between the current‐trial RTs and two error‐monitoring ERP components illustrate the post‐decision evidence‐accumulation process of performance monitoring (Desender et al. 2021). Results for the later error‐processing component, captured by Pe, showed a pattern distinct from ERN, underscoring Pe as a distinct error processing component. On error trials, larger Pe amplitudes predicted faster and less variable RTs, suggesting consistent shifts in error processing across participants. In contrast, when the current trial was correct, larger Pc amplitudes were linked to longer and more variable RTs, a pattern consistent with the evidence‐accumulation process during error processing (van Steenbergen et al. 2010; Ullsperger et al. 2014) or the process of evaluating performance. In addition, the distinct relationships between flanker P3 and Pe with subsequent RTs suggest that P3 and Pe reflect at least in part, distinct processes, as suggested by Overbeek et al. (2005). Future work could examine how Pe and P3 separately but concurrently contribute to performance adjustments are encouraged. Predicting Current‐Trial Accuracy We replicated the typical congruency sequence effect for accuracy: accuracy was higher for incongruent trials preceded by other incongruent trials than for incongruent trials preceded by congruent trials. The results indicate the importance of considering the interaction effect of previous‐ and current‐trial congruency on accuracy during performance monitoring tasks. Neglecting these congruency sequence effects, along with previous‐trial accuracy, may help explain why some studies observe inconsistent accuracy adjustments following errors and correct responses. Consistent with transient and reactive trial‐bound control (e.g., Silton et al. 2010), previous stimulus‐locked ERPs were unrelated to current‐trial accuracy, suggesting that stimulus‐related conflict processing exerts its primary effect within the same trial. For response‐locked ERPs, previous‐trial ERN and current‐trial ERN and Pe predicted current‐trial accuracy. However, this study found that previous‐trial Pe was not related to current‐trial accuracy, which is consistent with other studies (Larson et al. 2011; Li et al. 2022). Given that Pe is viewed as an index of conscious error awareness (Falkenstein et al. 1991; Overbeek et al. 2005), this result suggests that error awareness alone may not be sufficient to drive effective behavioral adjustments in subsequent trials. In other words, it could be that participants were aware of the error being committed in the previous trial, which could have hijacked attention that should have been allocated to the processing of the stimulus on the subsequent trial (Lavro et al. 2018). Strengths and Limitations Performance‐monitoring ERPs (N2, P3, ERN, and Pe) index at least partially distinct control processes, and the present findings show that each component uniquely tracks trial‐by‐trial adjustments in behavior. The within‐person changes in these ERP components, likely manifesting from partially overlapping, functionally related brain regions (including ACC activity, for example Wilk and Morton 2012), are translated into measurable and meaningful changes in RT and accuracy. Typically, classic behavioral markers of performance monitoring, such as PES and conflict adaptation effects, are studied using mean response times. In the present study, analyzing trial‐by‐trial fluctuations within individuals revealed relationships that would have remained hidden if only between‐person averages were considered. More specifically, these findings argue against interpreting PES and conflict adaptation as unitary behavioral signatures of a single control process. Instead, trial‐to‐trial adjustment appears to reflect partially dissociable contributions from conflict detection, evaluative processing, early error detection, and error awareness, as indexed by N2, P3, ERN, and Pe. Behavioral slowing after errors or conflict should therefore not be treated as a direct proxy for any one neural mechanism. Importantly, ERP and behavioral indices were not constant over trials, and their variability is not simply “noise”. Current findings show how a multilevel approach incorporating trial‐to‐trial data can isolate unique effects of specific events (e.g., whether N2 on a congruent trial has the same relationship as N2 on an incongruent trial). Traditional analyses of subject averages cannot uncover these nuanced, within‐person relationships (Volpert‐Esmond et al. 2021). Moreover, present findings show that RT variability differed systematically as a function of congruency sequence, accuracy, and ERP component amplitudes, suggesting heightened engagement of cognitive control processes following high conflict events—whether response‐ or stimulus‐related conflict. Therefore, within‐participant fluctuations appear to provide complementary information to subject average scores, and participants may exhibit performance‐monitoring profiles that are best evaluated by examining their own trial‐level data. This perspective aligns with similar observations in Kirschner et al. (2021). There are limitations to the current study. First, this study does not investigate why participants vary in their behavior and ERPs, only how they vary in behavior, nor is it clear why some individuals vary more than others (and whether this is an adaptive or maladaptive function). Sources of variability in cognitive tasks could be person‐specific, such as age, speed‐accuracy trade‐off, fatigue, attention, motivation, and emotion (Judd et al. 2024). Another source of variability is related to persons' cognitive or motor control: the degree to which a participant was aware or confident in their performance (Kirschner et al. 2021; Wessel 2018) or the functional activity of the ACC and resting state connectivity of the dlPFC to other parts of the PFC (Wang et al. 2014). Although this was a healthy sample, other underlying characteristics known to affect performance monitoring may also contribute to individual differences. A recent test of this hypothesis failed to establish how traits like anxiety and worry affect trial‐to‐trial ERN‐PES relationships (LoTemplio et al., preprint). The present data show that there are individual differences in ERP‐behavior relationships, and these data provide a foundation for future research investigating the meaning of these individual differences. Second, this study used one version of a typical flanker task, making it unclear whether other task‐ or experiment‐related variables influence ERP‐behavior relationships and intraindividual variability (task indices differ across tasks). For instance, in the current study, we quantified ERP components using mean amplitudes with predefined time windows. Although this approach captures amplitude differences, it does not allow for the investigation of potential latency differences of individuals. Future studies could apply latency‐sensitive measures to examine whether temporal shifts in ERP components also contribute to performance monitoring. Another possible limitation is that congruent‐trial errors were too infrequent, accounting for an average of only 1.4% of each participant's errors, to permit a direct test of whether post‐error adaptation differed between congruent and incongruent trials. Future research could examine the interaction between previous‐trial congruency and previous‐trial accuracy using different paradigms to determine whether errors following high‐conflict and low‐conflict trials reflect different mechanisms. Also, future investigations adopting this study's approach could examine the effect of different manipulations within and between paradigms, such as number of trials, congruency ratios, response–stimulus intervals, intertrial intervals, or the effects of feedback and valenced stimuli, on behavioral adaptations and intraindividual variability of ERP components (e.g., as demonstrated in Zhang and Luck 2023) or other manifestations (time‐frequency domains). These variables are known confounders of performance monitoring indices and behavioral adaptation. Still, this study contributes to the growing body of literature that cognitive variability is an omnipresent characteristic in human task performance, with task‐specific properties contributing to the variance patterns (Judd et al. 2024) that require single‐trial (Debener et al. 2007) and multilevel and multiverse analyses (for a review, see Clayson 2024). Finally, although the predictors survived strong statistical testing, the results indicate that task effects (congruency, accuracy) appear to have numerically larger effects on behavior than ERP component amplitudes. Future studies could examine how these within‐ and between‐person differences translate to meaningful functional outcomes (e.g., Hajcak et al. 2024), such as whether greater variability in ERP component amplitudes and task performance relates to worse outcomes or symptoms. Other work suggests that within‐person ERN variability was not related to some clinical measures (Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025; Holbrook et al. 2025), implying that any relationship between within‐person variability and clinical symptoms might be indirect or moderated by other factors (e.g., symptom severity). Implications and Conclusion The present study has important methodological implications. Traditional statistical models that ignore within‐person variability have neglected a potentially key dimension that may contribute to individual differences in cognitive control. Most of the performance‐monitoring ERP indices measured in the present study represent manifestations of ACC activity, but there are several other brain areas (e.g., dorsal posterior cingulate cortex: Cieslik et al. 2024) or networks (Dali et al. 2023) that are conjointly involved in (aware) error or performance processing. For instance, the pre‐supplementary motor area has been linked to the PES‐ERN relationship (Fu et al. 2019) and the pregenual ACC has been related to trial‐level RT variability (Johnson et al. 2015). Promising work has demonstrated that enhancing functional connectivity can improve adaptive behavior at the trial level, underscoring future research directions, for example, in work on PES (Reinhart et al. 2015). Additionally, traditional statistical approaches that average single‐trial data before analysis appear inappropriate for studies of RTs during a flanker task. For example, the present study observed intraindividual variability in RTs. Failing to account for such within‐person variance leads to highly biased parameter estimates and inflated Type I errors in traditional (location‐only) multilevel models (Leckie 2014; Walters et al. 2018). The use of location‐scale multilevel models provides a powerful framework for examining the intraindividual variability in RTs at the between‐ and within‐person level. Another implication for researchers is to consider what role variability should have in their studies. This includes making predictions about within‐person changes in RT means and variances. When planning extensive cognitive testing, placing more demanding tasks earlier in the session may help minimize fatigue‐induced variability. Greater variability (caused by, for example, fatigue) may contribute to reduced psychometric reliability (Clayson, Rocha, et al. 2024; Clayson, Shuford, et al. 2024; Clayson, Baldwin, et al. 2025) or data quality (Zhang and Luck 2023) over the course of longer tasks. Similarly, the stability of performance‐monitoring indices can be influenced by participant age and other demographic factors (e.g., N2: Liu et al. 2018; ERN: Boen et al. 2022). The current study adopted an innovative analytic approach to link behavioral and neural indices of performance monitoring on a trial‐by‐trial basis in a healthy sample. By jointly modeling past‐ and current‐trial congruency and accuracy alongside stimulus‐ and response‐locked ERP components, the analysis demonstrates how these factors collectively shape performance in cognitive control tasks. Importantly, single‐trial modeling of multiple indices captures performance monitoring as it unfolds in real time, illuminating how a host of interconnected processes support adaptive changes in behavior to achieve task goals. Finally, the pronounced within‐person variability in performance‐monitoring measures highlights their importance for explaining adaptive behavior following conflict and error detection. Author Contributions Miranda C. Lutz: conceptualization, writing – original draft. Bohyun Park: visualization, writing – review and editing. Philippe Rast: software, formal analysis, data curation, writing – review and editing. Scott A. Baldwin: writing – review and editing. Michael J. Larson: investigation, resources, writing – review and editing. Peter E. Clayson: methodology, software, formal analysis, data curation, writing – review and editing. Funding M.C.L., B.P., P.R., S.A.B., M.J.L.: No funding statements. Writing of the present manuscript was supported by a NIMH grant awarded to Peter Clayson (MH128208). This funding source was not involved in data analysis, interpretation, or the preparation of this submission for publication. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Data S1: hbm70538‐sup‐0001‐Supinfo.html.