In everyday life, people frequently encounter conflicts between their long-term goals, moral values, or social norms, on the one side, and momentary impulses to satisfy immediate desires, on the other. Self-control denotes the ability to resist such temptations and override impulsive responses in order to render behavior consistent with superordinate goals (Baumeister, Vohs, & Tice, 2007; Hofmann, Baumeister, Förster, & Vohs, 2012; Inzlicht, Legault, & Teper, 2014). Whereas high self-control predicts higher educational achievement and social adjustment, better coping with stress, and less substance abuse (Mischel et al., 2010; Tangney, Baumeister, & Boone, 2004), deficient self-control entails harmful behaviors such as overeating and overspending and is a core characteristic of substance use disorders (Bühringer, Wittchen, Gottlebe, Kufeld, & Goschke, 2008; Goschke, 2014; Heatherton & Wagner, 2011). Self-control failures (SCFs) thus incur severe personal and societal costs, due to poor health, disability, and early death (Schroeder, 2007; Wittchen et al., 2011). Investigating the neurocognitive basis of individual differences in self-control is therefore of great scientific relevance.

Self-control requires the individual to suppress prepotent responses leading to unwanted behaviors (e.g., to eat cake, smoke a cigarette), and thus is assumed to depend on response inhibition, a cognitive control process associated with the right inferior frontal gyrus (rIFG; Aron, Robbins, & Poldrack, 2014; Goschke, 2014). Consequently, impaired self-control is commonly thought to reflect the reduced implementation of cognitive control (e.g., Heatherton & Wagner, 2011; Hofmann, Schmeichel, & Baddeley, 2012). In line with this view, two previous studies yielded evidence that real-life SCFs related to smoking (Berkman, Falk, & Lieberman, 2011) and eating (Lopez, Hofmann, Wagner, Kelley, & Heatherton, 2014) are associated with reduced activity in the rIFG in a response inhibition task. In contrast to these studies, which focused on the neural correlates of sustained inhibitory control across a task, here we aimed to investigate the role of error processing as an indicator of performance monitoring. Errors indicate performance failures—that is, actions that run counter to one’s intentions—and therefore signal a strong need for subsequent behavioral adaptations in order to continue goal-directed behavior (Danielmeier & Ullsperger, 2011). Thus, instead of sustained inhibitory control across tasks, here we focused on dynamic adjustments in situations in which increased control was needed. Moreover, we did not restrict data acquisition to a certain (clinical) type of SCF (such as smoking). Rather, aiming at a general understanding of self-control, we addressed SCFs in the context of a wide range of different desires.

On the basis of current neurobiological models of cognitive control, we assume that self-control rests on a performance-monitoring network (PMN) comprising the posterior medial frontal cortex (pMFC), which includes the anterior middle cingulate cortex (aMCC) and the presupplementary motor area (preSMA), as well as the anterior insulae (aINS) (Uddin, 2015; Ullsperger, Danielmeyer & Jocham, 2014). Although methodological challenges (e.g., spatial smoothing, group differences, image resolution) make a clear functional localization within the pMFC difficult, functional magnetic resonance imaging (fMRI) studies have consistently shown that the pMFC is sensitive to evaluative signals (King, Korb, von Cramon, & Ullsperger, 2010; Ridderinkhof, Ullsperger, Crone, & Nieuwenhuis, 2004; Ullsperger & von Cramon, 2004). Likewise, there is converging evidence that the often coactivated aINS assigns salience to behaviorally relevant stimuli and events (Craig, 2009; Klein et al., 2007; Uddin, 2015).

In line with conflict-monitoring theory (Botvinick, Braver, Barch, Carter, & Cohen, 2001; Kerns et al., 2004; Miller & Cohen, 2001), we assume that the PMN registers conflict-induced performance problems and signals the need for enhanced cognitive control and appropriate adjustments to the lateral prefrontal cortex. In the context of response interference tasks such as the Stroop, flanker, or Simon task, these adjustments include recruitment of the rIFG, which mediates the inhibition of prepotent but nonintended responses (Aron et al., 2014). On a behavioral level, these adjustments of cognitive control are reflected in posterror slowing (PES)—that is, the tendency to slow down after an error (Danielmeier & Ullsperger, 2011; King et al., 2010; Rabbitt, 1966; Ullsperger et al., 2014).

According to the function of performance monitoring outlined above, a hypoactive PMN should lead to reduced recruitment of cognitive control, not only under laboratory conditions but also in real-life situations requiring self-control: For instance, a dieter who views a fast food commercial may fail to inhibit intrusive food-related thoughts, thus allowing a problematic desire to emerge. In such situations, aberrant performance monitoring would entail an inability to efficiently mobilize cognitive control processes, and thus likely leads to full-blown SCFs.

Following a brain-as-predictor approach (Berkman & Falk, 2013), here we tested whether everyday SCFs assessed via experience sampling are predicted by the neural correlates of performance monitoring, operationalized as error-related brain activity in a Stroop task under fMRI. On the basis of the assumption that SCFs result from deficient monitoring of performance problems and/or insufficient subsequent control adaptations, we hypothesized that low self-control would be associated with (i) low error-related activity in the PMN (aMCC, preSMA, aINS), (ii) low posterror recruitment of the rIFG, and (iii) reduced PES. To determine the discriminant validity of these neural and behavioral measures of performance monitoring with respect to self-ascribed self-control abilities, we furthermore investigated whether error-related brain activity and PES explain variance in SCFs beyond what is predicted by participants’ self-reports of trait self-control.

Materials and method

Participants

A total of 142 young adults (79 female, 63 male; age 20 to 26 years, M = 22.04 years, SD = 1.73) were recruited from a representative community sample for an ongoing longitudinal study on the role of cognitive control in the onset and early course of addictive disorders. Thus, although addiction was not the focus of the present investigation, participants were screened for symptoms of addictive disorders and addiction-like behaviors (see Appendix 1). Participants were excluded if they had a limited ability to provide informed consent and to understand the questionnaires and tasks, disorders that might influence cognition or motor performance (e.g., craniocerebral injury, multiple sclerosis), magnetic-resonance contraindications, lifetime schizophrenia or psychotic symptoms, bipolar disorder, somatoform, anxiety, obsessive-compulsive or eating disorders, or major depression in the last four weeks. Participants were paid €40 for completing the scanning session and experience sampling. All participants provided written informed consent. One participant was excluded due to technical problems during the fMRI acquisition; two participants were excluded due to error rates of 100% in the Stroop task, and 20 participants were excluded due to a complete lack of errors. We excluded one further participant whose parameter estimates deviated from the normal distribution. Thus, 118 participants (64 females, 54 male; age 20–26 years, M = 22.18, SD = 1.82) were included in the analyses reported here. This sample size is exceptionally large in comparison to previous studies that have used a brain-as-predictor approach.

Experience sampling procedure

Real-life SCFs across a wide range of behavioral domains were assessed via experience sampling (Hofmann, Baumeister, et al., 2012). Participants were given identical smartphones on which a customizable experience-sampling application was running while all other functions were blocked. Participants were shown how to use the device during a brief meeting with a research assistant. The experience-sampling period, during which the participants carried the devices with them at all times, started the next morning and lasted for seven consecutive days. Each day, eight alarms were emitted within a 14-h time window, which was adjusted to the participant’s habitual waking hours (starting at either 8, 9, or 10 a.m.). The exact time for each alarm was randomly selected, with the constraint that two alarms be at least 1 h apart. Whenever participants accepted an alarm, they completed a short questionnaire consisting of up to seven questions on the device. First, they were asked whether or not they had experienced a desire to enact a realizable behavior at some point during the last hour. If they reported a desire, they were asked to indicate the desire strength, on a scale from 1 (very weak) to 6 (very strong), and to select the respective desire type, from a list of 19 categories (eating, drinking, drinking alcohol, smoking, using some other substance, using the internet, playing a computer game, watching TV, buying something, gambling, exercising, sleeping, resting, retreating, misbehaving, socializing, having sex or intimacy, using a bathroom, and other). They were then asked whether they had a reason not to enact the desire (i.e., whether there was a conflict). If they reported a conflict, they were asked to indicate the conflict strength, on a scale from 1 (very weak) to 6 (very strong), and whether or not they had attempted to resist the desire. Eventually they were asked whether or not they had enacted the desired behavior. In summary, up to four dichotomous variables (desire, conflict, resistance, and enactment), one categorical variable (desire type), and two continuous variables (desire and conflict strength) were acquired per questionnaire. Depending on response rates, each participant completed up to 56 questionnaires. SCFs were operationalized as occasions on which the participant failed to resist temptation—that is, enacted a conflict-laden desire.

Self-reports of trait self-control

Participants’ self-evaluations of trait self-control were assessed with a German version of the Brief Self-Control Scale (BSCS; Bertrams & Dickhäuser, 2009; Tangney et al., 2004). The BSCS comprises 13 items (e.g., “I am good at resisting temptation”; “People would say that I have iron self-discipline”; “Sometimes I can’t stop myself from doing something, even if I know it is wrong”) and is one of the most widely used questionnaires in self-control research (Duckworth & Kern, 2011). High BSCS scores indicate high levels of trait self-control.

Stroop task and computation of PES

Error-related brain activity was measured using blood-oxygen level-dependent (BOLD) fMRI in a counting Stroop task (Bush et al., 1998), with a rapid event-related design. Figure 1 illustrates an exemplary sequence of trials. Each trial started with the presentation of a fixation cross with a jittered duration of 2–4 s (2.0, 2.5, 3.0, 3.5, or 4.0 s). The fixation cross was then replaced by a row of either one, two, three, or four identical digits from 1 to 4 for 1 s, yielding an average trial duration of 4 s. Participants held the index and middle fingers of both hands over four response buttons and were instructed to respond as quickly as possible to the number of digits while ignoring the denoted numbers. The stimulus–response mapping was natural: a one-digit stimulus required a response on the leftmost button, a two-digit stimulus required a response on the second-leftmost button, and so forth. The task-relevant number of digits and the task-irrelevant denotation of the digits were either congruent (“1,” “22,” “333,” “4444”) or incongruent (“111,” “2222,” “3,” “44”). A total of 80 trials (40 congruent, 40 incongruent) were presented in a randomized sequence, yielding a total task duration of 5 min 20 s. To clarify the task, participants completed a practice session of 15 trials inside the scanner before the actual experiment started.

Fig. 1
figure 1

Exemplary sequence of two trials in the counting Stroop task. After a jittered interstimulus interval (2–4 s), a row of identical digits was shown. Participants had to indicate the number of digits while ignoring the digits’ denotations by pressing one of four response keys. The number and denotation of the digits were either congruent (“1,” “22,” “333,” “4444”) or incongruent (“111,” “2222,” “3,” “44”). In all, 40 congruent and 40 incongruent trials were presented in randomized order

Full size image

As Dutilh et al. (2012) have demonstrated, PES calculated in the traditional way (posterror vs. postcorrect) is prone to global fluctuations in response times (RTs) over the course of the experiment. To reduce noise and minimize the impact of global RT fluctuations, we applied the robust method of PES calculation (Dutilh et al., 2012): The mean RT of correct pre-error trials was subtracted from the mean RT of correct posterror trials [RT (E+1) – RT (E–1)] (Table 1). Only the RTs from correct pre- and posterror trials were used for the PES calculation. For each participant, the z-standardized PES score was incorporated into subsequent hierarchical linear modeling (HLM) analyses predicting real-life SCFs as a Level-2 predictor (see below).

Table 1 Behavioral results of the Stroop task across participants (N = 118)
Full size table

fMRI data acquisition and analysis

Functional images were acquired using a T2*-weighted gradient-echo echo-planar imaging (EPI) sequence (TE = 25 ms, TR = 2 s, flip angle 78°, slice thickness 3.2 mm, matrix 64 × 64, FOV 19.2 cm, in-plane resolution 3 × 3 mm) on a Siemens MAGNETOM Trio A Tim 3-T scanner with a 32-channel head coil. Thirty-four axial slices, oriented parallel to the AC–PC line covering the whole brain, were acquired. In addition, high-resolution anatomical images were acquired (TE = 2.26 ms, TR = 1,900 ms, flip angle 9°, matrix 256 × 256, FOV 25.6 cm, 176 sagittal slices, slice thickness 1 mm) and co-registered with the functional images. SPM8 (www.fil.ion.ucl.ac.uk/spm/) was used for preprocessing and statistical analyses of the fMRI data. After realignment and slice-time correction (to the middle slice), the data were normalized to standard Montreal Neurological Institute (MNI) space, using a unified segmentation approach based on the separation of gray matter, white matter, and cerebrospinal fluid (voxel size 3 mm). The data were spatially smoothed using an 8-mm full-width at half-maximum (FWHM) Gaussian filter. For baseline correction, the data were high-pass-filtered with a cutoff period of 128 s.

We derived whole-brain measures of error-related brain activity by estimating a general linear model (GLM) of BOLD activity, which included six regressors of interest: (i) congruent correct trials, (ii) incongruent correct trials, (iii) congruent error trials, (iv) incongruent error trials, (v) postincongruent error trials, and (vi) postincongruent correct trials.Footnote 1 Motion parameters were included as regressors of no interest, and the regressors were convolved with a canonical form of the hemodynamic response. Two first-level single-subject contrasts were computed to assess the neural correlates of error processing during and after error commission: an error contrast (incongruent error vs. incongruent correct trials), and a posterror contrast (postincongruent error vs. postincongruent correct trials).Footnote 2 The corresponding contrasts for congruent errors were not calculated because of the low number of errors on congruent trials (see Table 1). At the second level, group contrasts were computed using one-sample t tests on both of the previously described single-subject contrasts.

Extraction of parameter estimates from regions of interest (ROIs)

At first we confirmed that the PMN was activated during error trials and that the rIFG was activated during posterror trials (see the Results and Appendix 2 for whole-brain activations). Then we defined ROIs as the overlap between the anatomical masks and task activity, thresholded at p < .001 (FWE-corrected, whole-brain). We constructed one mask for the PMN (bilateral MCC, SMA, and INS; Uddin, 2015; Ullsperger et al., 2014) and a second mask for the rIFG (pars triangularis, pars orbitalis, and pars opercularis; Aron et al., 2014) using the Automated Anatomical Labeling (AAL) toolbox (Tzourio-Mazoyer et al., 2002). Since we were exclusively interested in error-related brain activity within the PMN and rIFG ROIs, parameter estimates were extracted from those voxels that were significantly activated in the respective contrast (error or posterror contrast) and that overlapped with the respective mask (PMN or rIFG).

The Marsbar Toolbox (http://marsbar.sourceforge.net/) was used to extract raw parameter estimates. For each participant, z-standardized parameter estimates were incorporated into the HLM analyses as Level-2 predictors.

Integration of fMRI and experience sampling data with hierarchical linear modeling

To test whether error-related brain activity and PES predicted SCFs while taking into account the nested structure of the experience sampling data, fMRI, behavioral, and experience sampling data were subjected to multilevel regression analyses, using the Hierarchical Linear Modeling software package (HLM 7; Raudenbush, Cheong, Congdon, & DuToit, 2011). Since desire enactment indicates a lack of self-control only when the desires are at conflict with superordinate goals, only those situations in which such a conflict was reported were analyzed. Because the dependent variable, enactment, was binary, logistic multilevel regression analyses were applied using the Bernoulli model in the HLM. For each analysis, a hierarchical two-level model was built with situations (Level 1) nested within participants (Level 2). In each model, the Level-1 predictors were person-mean centered, whereas the Level-2 predictors were grand-mean centered. To avoid overparameterization and keep computational complexity low, only significant random variance components were kept in the models (Hox, 2010).

In a first step, a basic model was built to examine how behavior enactment was affected by desire strength and conflict strength, while controlling for possible effects of desire content by including desire type as a set of effects-coded variables at Level 1. In a second step, this basic model was extended by incorporating individually averaged parameter estimates from the PMN (aMCC, preSMA, aINS) and the rIFG, as well as PES slowing scores as z-standardized Level-2 predictors. To avoid problems due to collinearity among the predictor variables, three separate analyses were applied in which only one Level-2 predictor at a time was included. In a third step, each of these three models was extended by including z-standardized BSCS scores as a second Level-2 predictor.

Results

Experience-sampling response rates

On average, participants responded to 42.83 (SD = 10.17) of the 56 issued alarms. They reported 30.64 (SD = 10.53) desires, 10.48 (SD = 6.33) of which were conflict-laden. SCFs were operationalized as all occasions on which the participant failed to resist temptation and enacted a desire that conflicted with a long-term goal. Of the conflict-laden desires, 5.64 (SD = 4.67) were reported to result in SCFs, on average (for a categorization of occasions by desire types, see Appendix 3).

Behavioral results of the Stroop task

Table 1 shows RTs and numbers of errors for the different conditions of the Stroop task. As is commonly found in Stroop experiments, increased RTs, t(117) = 24.753, p < .001, and errors, t(117) = 12.38, p < .001, were found on incongruent as compared to congruent trials: In the incongruent condition, the mean RT was increased by 51 ms (SD = 23), and the mean number of errors was increased by 1.97 (SD = 1.73). Moreover, significant PES (using the robust method; Dutilh et al., 2012) was observed following incongruent errors (M = 34 ms, SD = 106), t(117) = 3.49, p < .001.Footnote 3 No corresponding PES effect was observed following congruent errors (M = – 10 ms, SD = 123), t(57) = – 0.43, p = .669.

Significant PES was also observed when using the traditional method for calculation (posterror minus postcorrect; M = 34 ms, SD = 106 ms), t(117) = 3.49, p < .001. PES calculated in this way was not correlated with error-related activity in either the PMN (r = .115, p = .215) or rIFG (r = .102, p = .271) and did not significantly predict SCFs (blog = – 0.102, p = .116, one-sided).

Note that PES was not accompanied by posterror changes in response accuracy (PIA): Response accuracy was 94.6% on trials following incongruent hits and 96.1% on trials following incongruent errors; thus, there was a posterror accuracy increase of 1.5%. This increase was not statistically significant, t(117) = 1.33, p = .187. Note, however, that the interpretability of PIA in this dataset is limited due to a marked ceiling effect in postincongruent error response accuracies (104 of 118 participants showed a response accuracy of 100% on trials following incongruent errors; Danielmeier & Ullsperger, 2011).

Error-related brain activity

Figure 2 shows whole-brain activations (p < .001, FWE-corrected) in the error and posterror contrasts. As expected, the PMN was activated on error trials, and the rIFG was activated on posterror trials. Global peak activations in the error- and posterror contrasts were located in the right aINS (30/20/– 8) and the rIFG (57/26/– 2), respectively (for all peak activations, see Appendix 2). Correlations with PES were nonsignificant for error-related brain activity in both the PMN (r = .115, p = .216) and rIFG (r = .091, p = .329).

Fig. 2
figure 2

Whole-brain activations for the error and posterror contrasts, thresholded at p < .001 (FWE-corrected; for all peak activations, see Appendix 2). The performance-monitoring network (PMN; including aMCC, preSMA, and aINS) was activated in the error contrast, and the right inferior frontal gyrus (rIFG) was activated in the posterror contrast. Overlaps between these activations and predefined anatomical masks were used as the regions of interest from which parameter estimates were extracted in order to predict self-control failures (see Fig. 3)

Full size image

Prediction of SCFs

Situational variables

First, the effects of the situational variables, desire strength and conflict strength, on the occurrence of SCFs were examined without considering any predictors on the second level (Table 2). As one would expect (Hofmann, Baumeister, et al., 2012), the frequency of SCFs increased with desire strength (blog = 0.40, p < .001) and decreased with conflict strength (blog = – 0.78, p < .001): SCFs were most likely in situations in which desire strength was high and/or conflict strength was low.

Table 2 Multilevel logistic regression results: Predicting SCFs from desire strength, conflict strength, neural correlates of error processing, and trait self-control
Full size table

Error-related brain activity and PES

Consistent with our first prediction, we observed that participants showing weaker activation in the PMN (blog = – 0.20, p = .018, one-tailed) on error relative to correct trials were more likely to commit SCFs than were participants showing stronger error-related activation of the PMN (Table 2 and Fig. 3). Consistent with our second prediction, SCFs were also significantly predicted by activity in the rIFG on posterror trials: More SCFs were reported by participants showing weaker posterror activation in the rIFG (blog = – 0.25, p = .004, one-tailed) (Fig. 3, Table 2). Consistent with our third prediction, we found that reduced PES predicted more SCFs (blog = – 0.19, p = .013, one-tailed).

Fig. 3
figure 3

Effects of error-related performance-monitoring network (PMN) and right inferior frontal gyrus (rIFG) activity, desire strength, and conflict strength on the probability of real-life self-control failures (SCFs). Anatomically defined ROI masks are shown in red in the online figure. Brain activity was extracted from voxels that were significantly activated in response to errors (p < .001, FWE-corrected, whole-brain) and that overlapped with the ROI masks (shown in yellow in online version). Predicted log-odds for SCFs have been transformed into probabilities. The x-axes range from the 5th to the 95th percentiles of values for PMN activation and rIFG activation, respectively. Separate plots are shown for high (1 SD above the mean) and low (1 SD below the mean) desire and conflict strength. SCFs were less likely to occur with low desire strength or high conflict strength, and in participants who showed stronger PMN and rIFG activations

Full size image

Note that (despite reduced sample sizes) the results remained consistent when participants were excluded who committed only one error on incongruent trials (n = 81; PMN: blog = – 0.35, p = .001; rIFG: blog = – 0.31, p = .011), two errors on incongruent trials (n = 52; PMN: blog = – 0.42, p = .014; rIFG: blog = – 0.39, p = .047), or three errors on incongruent trials (n = 37; PMN: blog = – 0.44, p = .038; rIFG: blog = – 0.44, p = .074.

In none of the analyzed brain regions did error-related activity significantly interact with desire or conflict strength. Because analyses of correctly performed conflict trials might also be of interest, these results can be found in Appendix 4. The results for between-group analyses are reported in Appendix 1.

Of note, SCFs were also significantly predicted by higher numbers of incongruent errors (blog = 0.317, p = .013). Nonetheless, error rates did not explain the effects of error-related brain activity and PES on SCFs: Even after including the number of errors as an additional predictor, the effects of PMN (blog = – 0.189, p = .028) and rIFG (blog = – 0.216, p = .015) activity, as well as of PES (blog = – 0.191, p = .014), on SCF remained significant. Error rates were not correlated with PES (r = – .061, p = .514).

Trait self-control

The negative main effects of error-related brain activity and PES remained significant when BSCS scores were included as an additional Level-2 predictor (Table 2), showing that these neural and behavioral measures of performance monitoring predicted variability in SCFs that was not explained by trait self-control. SCFs were nonetheless also significantly predicted by BSCS scores in all models: As expected, participants with low as compared to high trait self-control were significantly more likely to commit SCFs. Significant cross-level interactions of BSCS scores with desire and conflict strength indicated that the effect of trait self-control on SCFs was stronger for weak desires and strong conflicts.

Discussion

The goal of the present study was to further elucidate the neurocognitive mechanisms underlying individual differences in real-life SCFs. Therefore, we combined neuroimaging (fMRI) with ecological assessment of self-control via experience sampling in a large sample of participants. Whereas it is commonly assumed that impaired self-control reflects the reduced implementation of cognitive control, here we investigated the more specific assumption that the reduced implementation of cognitive control during real-life SCFs may result from insufficient mobilization of control processes due to deficient performance monitoring. Consistent with this idea, our results showed that individual differences in real-life self-control were reliably predicted by error-related brain activation in the performance-monitoring network, comprising the pMFC and aINS, posterror activation of the rIFG, and, on a behavioral level, PES. These results suggest that individual differences in real-life self-control reflect variability in both initial (error detection) and subsequent (implementation of cognitive control) components of performance monitoring.

This pattern goes beyond previous reports of associations between SCFs and activation of the rIFG (Berkman et al., 2011; Lopez et al., 2014), and emphasizes the role of performance monitoring in real-life self-control: Consistent with conflict-monitoring theory (Botvinick et al., 2001; Kerns et al., 2004; Miller & Cohen, 2001), according to which the implementation of cognitive control depends on a signal generated by the PMN, we showed that real-life SCFs were more likely in those individuals who exhibited low error-related activation of the PMN in a Stroop task. This indicates that among individuals with low self-control, unwanted action outcomes or desires conflicting with superordinate goals may be less likely to be detected and/or do not elicit sufficiently strong salience signals, leading to reduced recruitment of cognitive control. Converging, albeit more indirect, evidence supporting this view comes from the fact that PMN dysfunctions are often found in substance-related and addictive disorders, of which impaired self-control is a core characteristic (Carey, Nestor, Jones, Garavan, & Hester, 2015; Goschke, 2014; Luijten et al., 2014).

Note that although correlates of impaired cognitive control implementation (reduced posterror rIFG activation, reduced PES) also predicted real-life SCFs, this does not necessarily imply a lack of cognitive control competencies in those individuals who act in a less self-controlled way. In line with current models of self-control that have incorporated conflict-monitoring theory (Inzlicht, Bartholow, & Hirsh, 2015; Kotabe & Hofmann, 2015), less self-controlled behavior could alternatively be explained by reduced mobilization of cognitive control as a consequence of insufficient performance monitoring. This is consistent with the more general view that strong control competencies alone are not sufficient for successful self-control. Instead, even individuals with high capacities for exerting cognitive control may exhibit low self-control if they fail to efficiently mobilize control in response to desire-goal conflicts.

In contrast to Lopez et al. (2014), we did not find an interaction between desire strength and rIFG activation. Furthermore, we found no interactions between neural and behavioral measures of error processing and conflict strength. This suggests that, in contrast to the ability to exert sustained inhibitory control, which may be especially important when facing strong desires, an efficient mobilization of control is crucial for self-control regardless of desire and conflict strength. An alternative explanation for the lack of interaction with desire and conflict strength is that third variables that are unrelated to desire and conflict strength may explain the observed association between error-related brain activity and SCFs.

We found no correlation between PES and PMN activity. This might be explained by the fact that the PMN as defined in the present study (preSMA, MCC, insula) only partially overlaps with neural substrates described for PES, which, in addition to the preSMA, also include right-hemispheric lateral frontal areas and the subthalamic nucleus (Danielmeier & Ullsperger, 2011).

Response conflicts and resulting performance problems have been shown to be inherently aversive (Aarts, De Houwer, & Pourtois, 2012), and this negative affect appears to play a functional role in initiating required adjustments in cognitive control (Dreisbach & Fischer, 2012; Inzlicht et al., 2015; Saunders, Lin, Milyavskaya, & Inzlicht, 2017; Wiswede, Münte, Goschke, & Rüsseler, 2009). This suggests that impaired mobilization of control in self-control situations may reflect a relative inability to process conflict-related aversive states adaptively. Instead of effectively reducing negative affect by mobilizing control, thus addressing the conflict itself and solving it in accordance with superordinate goals, individuals with low self-control may use less conflict-centered emotion regulation strategies or even ruminate about the emotion itself. In line with this conjecture, in a recent study (Wolff et al., 2016) we found that high executive control competencies predicted lower proneness to daily SCFs in individuals with a disposition toward “action orientation” (who have been shown to efficiently recruit control in response to conflict; Goschke & Bolte, 2018; Kuhl & Beckmann, 1994), but not in “state-oriented” individuals, who tend to respond to conflicts with rumination rather than enhanced control recruitment (Fischer, Plessow, Dreisbach, & Goschke, 2015; Jostmann & Koole, 2007).

Apart from ineffective responses to negative affect, reduced mobilization of control and low error-related PMN activations may also result from low control motivation—that is, a low expected payoff in relation to costs from engaging control (Shenhav, Botvinick, & Cohen, 2013). Thus, in line with previous evidence suggesting an inverse relationship between cognitive demand avoidance and the efficacy of self-control (Kool, McGuire, Wang, & Botvinick, 2013), our finding of low error-related brain activity in individuals with low self-control may in part reflect individual differences in effort discounting.

Since SCFs are a core characteristic of various mental disorders such as addiction, depression, and certain anxiety disorders (Goschke, 2014), our study is also of theoretical and practical relevance for clinical conditions: Besides attempts to improve control competencies with, for instance, working memory training (Karbach & Verhaeghen, 2014; Klingberg, 2010) or training of inhibitory control (Berkman, Kahn, & Merchant, 2014), patients might also benefit from therapeutic interventions that include modules to strengthen performance monitoring—for example, via mindfulness-based interventions (e.g., Jha, Krompinger, & Baime, 2007; Mrazek, Franklin, Phillips, Baird, & Schooler, 2013; Witkiewitz, Bowen, Douglas, & Hsu, 2013; Saunders, Rodrigo, & Inzlicht, 2016).

Note that although we showed that predictions based on conflict-monitoring theory yield convincing results with regard to real-life self-control, this does not mean that alternative accounts of cognitive control are less valid. We are aware of the fact that a simple conflict monitoring view has been challenged (e.g., Grinband et al., 2011), and we do not suggest that the role of the medial prefrontal cortex is restricted to a single cognitive function (Ebitz & Hayden, 2016). Rather, apart from multiple forms of performance monitoring (Neta, Schlaggar, & Petersen, 2014), the medial prefrontal areas have been implicated in reward processing (Holroyd & Coles, 2002; Holroyd & Yeung, 2012), learning (Alexander & Brown, 2011), execution of control, action selection and computing the value and cost of control (Shenhav et al., 2013). At the same time, however, we stress that our conclusion that real-life SCFs can partly be explained by insufficient performance monitoring leading to reduced behavioral adaptations is neither invalidated by alternative accounts of cognitive control nor inconsistent with accounts according to which brain structures involved in performance monitoring may serve more general functions such as cost-benefit analyses underlying the adaptive regulation of control (Shenhav et al., 2013).

Having shown that everyday SCFs are predicted by neural processes related to performance monitoring and cognitive control, the present study demonstrates that brain data serve as an informative window into the psychological processes underlying important real-life behaviors. When following such a brain-as-predictor approach, it is critical to consider whether the predictive information of neural data goes beyond what could be obtained otherwise (Berkman & Falk, 2013). Here we found error-related brain activity (and PES) to predict variability in SCFs that was not explained by trait self-control, indicating that these measures have unique predictive power and show discriminant validity with respect to self-reports.

A potential methodological limitation of the present study results from the necessity of using a task that produces maximally salient—and thus relatively rare—error events, which might have led to a decreased signal-to-noise ratio in error- and posterror trials. Nevertheless, consistent with current models of cognitive control (Aron et al., 2014; Ridderinkhof et al., 2004; Ullsperger et al., 2014), we found extremely robust task activations related to performance monitoring (aINS, aMCC, SMA) and inhibition (rIFG), even when using a very conservative threshold (p < .001 FWE-corrected; see Fig. 2 and Appendix 2)—and these were indeed predictive of real-life self-control. However, to obtain more error trials while preserving error saliency, future studies might employ longer tasks or administer the tasks repeatedly.

To conclude, our findings suggest that real-life SCFs may result from deficient performance monitoring, leading to insufficient recruitment of cognitive control in response to action outcomes that conflict with superordinate goals. The present study adds to the still sparse attempts to bridge the gap between experimental laboratory research and the assessment of real-life self-control using a brain-as-predictor approach and provides further support for the validity of that approach. Although currently correlational, this promising approach allows for identifying possible neurocognitive mechanisms underlying real-world behavioral problems. This may serve as the basis for future prospective and interventional studies testing causal relations and may help improve the prediction, diagnosis, and modification of self-control impairments in mental disorders as well as of maladaptive social and economic behaviors.

Author note

T.G. conceptualized the research hypothesis and study design. Data analysis was carried out by K.M.K. and M.W. The manuscript was written primarily by K.M.K. and M.W., with contributions from T.G., H.M., A.K., M.S., and G.B. We thank Joseph A. King for helpful comments on the manuscript. Funding for this research was provided by a grant from the German Research Foundation (DFG) for the Collaborative Research Centre “Volition and Cognitive Control” (SFB 940/1-2015 and SFB 940/2-2016). The authors declare no conflicts of interest.