Introduction

Much of our behavior is driven by routines. However, when an unexpected situational change occurs, we are able to overcome our automatic response in favor of a more appropriate one. This is referred to as cognitive control. For example, when our familiar road to home is blocked, we can suppress our urge to drive as usual and take the advised alternative route. The next day, taking the detour is already easier, perhaps due to the already formed association between the roadblock and the detour.

According to extant models, a behavioral impasse, like the roadblock, constitutes a trigger for cognitive control (e.g., conflict-monitoring model, Botvinick, Braver, Barch, Carter, & Cohen, 2001; error likelihood model, Brown & Braver, 2005; supervisory attentional system model, Norman & Shallice, 1986). For example, in the conflict-monitoring model, cognitive control is triggered by conflict between simultaneously active responses, addressing the issue of how the cognitive system knows when cognitive control is needed. The adaptation-by-binding model (Verguts & Notebaert, 2008, 2009) further specifies the mechanism of how such control may be implemented and addresses how the cognitive system learns to select the correct response. This model proposes that cognitive control emerges from fast binding between stimulus, action, and context events. Furthermore, the model proposes that such stimulus–action–context binding is modulated by salient (arousing) stimuli that arrive simultaneously. In the context of a flanker task (Eriksen & Eriksen, 1974; see Fig. 1), for instance, the response conflict caused by incongruent flankers increases binding between task demand representations, target stimulus, and response. The implementation of cognitive control upon the detection of conflict is typically considered as a reactive form of control, since one reacts upon the detection of difficulties. Other models describe how the system can be optimized in anticipation of difficulties, which is referred to as proactive control (Braver, 2012).

Fig. 1
figure 1

Schematic model (after Verguts & Notebaert, 2009) for the flanker task (incongruent stimulus, presented upper left). There are two visual input layers, representing input from the central position and the peripheral position. The dorsolateral prefrontal cortex (DLPFC) represents task demands. When instructed to respond to the central position (central arrow of the flanker stimulus), this dimension is prioritized due to the top-down influence of the DLPFC. Because of the activation of the peripheral position, the incorrect answer also becomes slightly activated. This simultaneous activation at the response layer is registered as conflict in the medial frontal cortex (MFC). The MFC consequently activates the neuromodulatory nucleus to release the neuromodulator. The neuromodulator increases binding between the most active representations, which results in increased task-relevant connections

Full size image

A frequently used index of cognitive control is the Gratton effect. It means that the congruency effect (difference in response time [RT] between incongruent and congruent stimuli) is smaller after incongruent than after congruent trials (Gratton, Coles, & Donchin, 1992). It has been argued by Hommel, Proctor and Vu (2004) and Mayr, Awh and Laurey (2003) that the Gratton effect is due to mere feature integration or repetition effects, respectively. However, recent research provides evidence that the effect is at least partly driven by cognitive control (Notebaert & Verguts, 2007; Ullsperger, Bylsma, & Botvinick, 2005). The model accounts for the Gratton effect, since it proposes that active representations (generally, task-relevant associations) are bound together after conflict. This will result in a stronger focus on task-relevant information because top-down control (dorsolateral prefrontal cortex) is increased, and target–response connections are strengthened. Although the model explains the Gratton effect by means of stronger connections, it does capture the generality of the effect, as long as the same relevant dimension is used. When the task-relevant dimension changes between two trials, the model explains no or a reversed Gratton effect. This has indeed been empirically observed (Notebaert & Verguts, 2008).

There is good support for the hypothesis that conflict is detected in the medial frontal cortex (MFC; Botvinick et al., 2001). In the adaptation-by-binding model, this triggers the norepinephrine (NE) system in the locus coeruleus (LC), leading to increased binding in the cortex and hippocampus. This theory is consistent with the fact that (1) the MFC projects to the LC (Jodo, Chiang, & Aston-Jones, 1998; see Aston-Jones & Cohen, 2005, for a review), (2) arousal enhances memory representations (e.g., McGaugh, 2006), (3) NE has been described as a “now print” signal (Harley, 2004; Livingston, 1967; Sara, 2009), (4) NE enhances binding by modulating Hebbian learning (Harley, 2004), and (5) Hebbian learning can develop rather quickly (e.g., within five trials, the smallest number tested in the paradigm reviewed in Weinberger & Bakin, 1998). However, dopamine (DA) is also described as a major learning signal in the brain (Reynolds, Hyland, & Wickens, 2001) and is also related to Hebbian learning. In vivo recordings in rodents show that DA (via D1 receptors) plays an essential role in both early and late long-term potentation in learning. In particular, DA may interact with the NMDA receptor, which is thought to implement Hebbian plasticity at the cellular level (Granado et al., 2008; see Lisman, Grace, & Duzel, 2011, for a review). FMRI research in humans shows that VTA activation, which suggests DA release, activates cortical areas related to cognitive control and predicts enhanced cognitive control in a task-switching paradigm (Savine & Braver, 2010). Similar to NE, DA is released after arousing events (see Bromberg-Martin, Matsumoto, & Hikosaka, 2010, for a review). Moreover, binding between a visual stimulus and a subsequent action is suggested to be mediated by DA (Colzato, van Wouwe, & Hommel, 2007; Colzato et al., 2012; Schnitzler & Gross, 2005) via the D1 receptor (Colzato & Hommel, 2008). Besides the D1 receptor type, there is the D2 receptor type, which is involved in cognitive flexibility (Van Holstein et al., 2011). In the present article, we address only D1 receptor type modulation.

Hence, either NE or DA (or both) could be the relevant neuromodulator for implementing cognitive control. Consistent with a role for DA (related to reward processing; e.g.,Schultz, 1998), Braem, Verguts, Roggeman, and Notebaert (2012) found that reward on a (correct) previous trial increases the Gratton effect, especially in highly reward-sensitive persons (but see Stürmer, Nigbur, Schacht, & Sommer, 2011; Van Steenbergen, Band, & Hommel, 2009). Here, we follow a complementary approach: Instead of presenting affective stimuli, we measure markers of the autonomic nervous system to investigate its modulation of the Gratton effect. To investigate the role of NE in cognitive control, we measure pupil dilation on a trial-to-trial basis. Pupil dilation provides a measure of both tonic and phasic LC activation in humans (Gilzenrat, Nieuwenhuis, Jepma, & Cohen, 2010; see Nieuwenhuis & Jepma, 2011, for a review). The LC is the main source of NE in the brain (Sara, 2009) and is connected via two intermediate synapses to both the sphincter and dilator muscle of the iris (Samuels & Szabadi, 2008). Moreover, NE agonists (e.g., clonidine) dilate the pupil, whereas NE antagonists (yohimbine) reduce pupil size (Koss, 1986). Hence, pupil dilation can act as an indirect measure of NE release. At the same time, to investigate the role of DA, we measure eye blinks on a trial-to-trial basis. Support for the assumption that eye blinks are related to DA processing comes from investigations of eye blink rate (EBR). DA increases the EBR, in both humans and rats (Blin, Masson, Azulay, Fondarai, & Serratrice, 1990; Taylor et al., 1999). Furthermore, Parkinson patients (with impaired DA functioning) exhibit decreased EBR (Deuschl & Goddemeier, 1998), and patients diagnosed with schizophrenia (with increased DA uptake in the striatum) show elevated EBR (Freed et al., 1980; Karson et al., 1983). Also, recreational cocaine users exhibit a reduced EBR, consistent with their reduced D2 receptor density (Colzato, van den Wildenberg, & Hommel, 2008). Similarly, there is a decreased EBR in chronic cannabis users (Kowal, Colzato, & Hommel, 2011). Moreover, a link between EBR, DA, and cognitive control was established by Dreisbach et al. (2005), demonstrating an influence of EBR on perseveration and distractibility, which was modulated by DRD4 polymorphism.

The adaptation-by-binding model (Verguts & Notebaert, 2008, 2009) proposes that the Gratton effect results from binding. Together with the reviewed literature on Hebbian learning and neuromodulation, the model motivates our hypotheses that NE and DA, as they relate to binding, might play a role in conflict adaptation. Accordingly, we investigated the modulatory influence of pupil dilation (putative marker of NE) and eye blink (putative marker of DA) on the Gratton effect. More specifically, we tested whether congruency predicts increased pupil dilation and blinking and whether pupil dilation and/or blinking predicts increased adaptation (larger Gratton effect) from trial to trial.

Method

Participants

Forty-eight students participated for a monetary reward of 8 Euros (mean age = 22.1 years [range, 18–33], 38 female, 46 right-handed). All participants gave their written informed consent.

Apparatus

Eye data were recorded using an Eye Link 1000 Tower Mount (SR Research, Ontario, Canada) eye tracker. Sample rate was 500 Hz. Responses were collected with a Microsoft SideWinder ® Plug & Play Game Pad.

Stimuli

The stimuli on each trial of the flanker task consisted of five arrows (e.g., >><>>). The middle arrow was the target stimulus; the two arrows left and right of the target arrow were flanker arrows. Arrows were oriented left (<) or right (>). This resulted in four different stimuli, since both target directions had congruent and incongruent flankers. The stimuli were presented in a random order in equal proportions. The arrows were presented in black on a white background. All stimuli had the same luminance to prevent luminance confounds on the pupil dilation recordings. An error was followed by a low tone.

Procedure

Participants were instructed to respond to the direction of the target arrow by pressing the left or right button with their left or right index finger, respectively. In order to acquire enough trials with pupil dilation data, we asked participants to blink less than usual, but not to refrain from blinking. If they experienced dry eyes, they were encouraged to blink during drift correction or breaks. Calibration and validation of gaze position were carried out with a 9-point grid. Viewing was binocular throughout the experiment, but pupil dilation was recorded for the right eye only. A chinrest and a brace at forehead height were used to restrict head movements. Participants executed a training block consisting of 20 trials; then they performed eight test blocks consisting of 100 trials each. In both the training block and the test blocks, each trial started with stimulus presentation for 200 ms. In the case of a correct response, the stimulus was followed by a fixation cross during 2,000 ms. An incorrect response was followed by a low tone for 250 ms, after which again a fixation cross was presented during 4,000 ms. To ensure accurate eye data collection, we applied drift correction for gaze position every tenth trial. In order to investigate a possible effect of luminance, half of the participants (n = 22) executed the experiment in a brightly lit room, and the other half of the participants (n = 26) in a dimly lit room. Within each group, we held luminance constant. The experiment lasted for approximately an hour.

Data analysis

Only correct trials were analyzed, with RT as the dependent variable. Since only incorrect trials were followed by feedback, trials of interest contained no feedback. The data were analyzed with both repeated measures ANOVA and hierarchical generalized linear modeling (HGLM) applying the “summary statistic” approach (Lorch & Myers, 1990; Notebaert & Verguts, 2007). When the dependent variable was continuous, the first-level model was linear (so HGLM was actually just a hierarchical linear model); when the dependent variable was binary (i.e., in the case when blink was the dependent variable), the first-level model was logistic. The HGLM procedure allows taking into consideration several correlated factors simultaneously by including them as regressors. For instance, we expect that pupil dilation will be affected by congruency. Therefore, the factor previous pupil dilation will be correlated with the factor previous congruency. Hence, if we want to show that previous pupil dilation modulates the Gratton effect, it is important to show that this effect is not caused by previous congruency. HGLM will reveal unique variance that is explained by each factor independently. In this way, HGLM is complementary to the more standard ANOVA approach. Results that turn out to be consistent across analyses can be considered robust. In the HGLM analyses, for each individual participant separately, we applied linear (or logistic in the case in which blink was the dependent variable) regression, including each trial as one data point. Incorrect trials were not included, nor were trials following errors in the n−1 regressors included. The regression coefficients (betas) for the different regressors from the individual-subject analyses were subsequently investigated with one-sample t-tests, testing whether they deviated from zero at the group level.

Trials with missing pupil data were considered blink trials (containing one or more blinks) and, accordingly, were included in the blink analysis. Since we asked participants to blink less than usual, but not to refrain from blinking, there was, on average, an equal number of blink and pupil trials. In the HGLM analysis, the regressors were congruency (Cn), previous-trial congruency (Cn−1), interaction between congruency and previous-trial congruency (i.e., the Gratton effect; Cn*Cn−1), blink (Bn), previous-trial blink (Bn−1), Cn*Bn−1, and finally Cn*Cn−1*Bn−1. Both congruency and blink were modeled with dummy variables (congruent = 0, incongruent = 1; no blink = 0, blink = 1). Since there was a small break after every tenth trial to perform drift correction, every tenth trial was excluded from all previous-trial regressors.

For the pupil analyses, raw pupil data was preprocessed using MATLAB 7.9. Only trials without blinks were included in the pupil dilation analysis. From each trial, a window of 2 s starting at stimulus presentation was analyzed. As baseline, we took pupil dilation at the start of each trial (first three samples). Analyses based on data with just the first sample of each trial as baseline yielded very similar results. Baseline pupil dilation at the start of each trial was subtracted from maximum pupil dilation within a trial to prevent an influence of drift. This defined pupil size on each trial. For the ANOVA, a pupil was defined as large if it was in size percentile 50 or more for that subject, and as small otherwise. In the HGLM analysis, the regressors were similar to those for the blink analyses, except that pupil size on the current (Pn) and previous trial (Pn−1) replaced the blink variables. Congruency was again modeled with a dummy variable, but pupil dilation was continuous (the larger the number [scale in arbitrary units], the larger the pupil). As in the blink HGLM analysis, every tenth trial was excluded from previous-trial regressors.

Results

Pupil data were collected on 52 % of the trials. Blinking occurred on 42 % of the trials. The average error rate was 6 %.

Before the main analysis, we first checked whether congruency on the current trial predicts blink on the current trial with an HGLM analysis with congruency on the current trial as regressor. This was indeed the case, t(47) = 1.878, SE = .041, p = .034, one-sided, indicating that blinking increases on incongruent trials, as compared with congruent trials.

We applied a blink ANOVA on RTs, with congruency on the current trial, congruency on the previous trial, and blink (yes / no) on the previous trial as independent variables. We found a significant main effect for congruency on the current trial, F(1, 47) = 693, MSE = 938, p < .001, indicating that congruent trials are faster (see Fig. 2a, b). We also obtained a significant main effect for congruency on the previous trial, F(1, 47) = 5.13, MSE = 238, p = .03, replicating earlier reports of postconflict slowing (Verguts, Notebaert, Kunde, & Wühr, 2011). There was a significant two-way interaction between congruency on the current trial and congruency on the previous trial, F(1, 47) = 37.5, MSE = 383, p < .001, indicating an overall Gratton effect. The three-way interaction between congruency on the current trial, congruency on the previous trial, and blink on the previous trial was also significant, F(1, 47) = 16.4, MSE = 206, p < .001, indicating a larger Gratton effect after a blink trial, as compared with after a no-blink trial (see Fig. 2a, b). Post hoc tests revealed that participants had a Gratton effect after a blink, F(1, 47) = 43.4, MSE = 365, p < .001, but also after a no-blink trial, F(1, 47) = 8.53, MSE = 223, p = .005. We also found a significant main effect for blink on the previous trial, F(1, 47) = 5.80, MSE = 1,592, p = .02, and a significant two-way interaction between congruency on the previous trial and blink on the previous trial, F(1, 47) = 6.43, MSE = 243, p = .02.

Fig. 2
figure 2

The Gratton effect after a no-blink trial (a), after a blink trial (b), after a small pupil trial (c), and after a large pupil trial (d). The y-axis shows average RTs; for each average, the confidence interval is depicted

Full size image

In the HGLM blink analysis (see Table 1), we found a significant main effect for congruency (Cn), t(47) = 18.4, p < .001, β = 90.3, indicating that participants were slower on incongruent trials. We found a significant two-way interaction between congruency on the current and the previous trials (Cn*Cn−1), t(47) = −3.32, p = .002, β = −14.6, indicating an overall Gratton effect. Importantly, and consistent with the blink ANOVA, we found a significant three-way interaction between congruency on the current trial, congruency on the previous trial, and blink on the previous trial (Cn*Cn−1*Bn-1), t(46) = −2.22, p = .03, β = −11.4, indicating a larger Gratton effect on trials following a blink trial, as compared with trials following a no-blink trial (cf. Fig. 2). This shows that blinking on the previous trial has a significant effect on cognitive control even after controlling for the effect of congruency on the current and previous trials. We also found a significant main effect for blink on the current trial (Bn), t(46) = 2.21, p = .03, β = 5.98. There were no significant differences between the groups (tested under bright vs. dimmed light conditions; all two-sample ts < 1.48, all ps > .15).

Table 1 Blink analysis: coefficients of the HGLM analysis with response time as dependent variable
Full size table

To examine the role of pupil dilation, before the main analysis, we first checked whether congruency on the current trial predicts pupil dilation with an HGLM analysis with congruency on the current trial as regressor, and it did, t(47) = 5.14, SE = 6.60, p < .001. This indicates that pupil dilation increases more on incongruent trials, as compared with congruent trials.

We then applied a pupil dilation ANOVA on RTs, with congruency on the current trial, congruency on the previous trial, and pupil dilation (large/small) on the previous trial as independent variables. We found a significant main effect for congruency on the current trial, F(1, 47) = 857, MSE = 958, p < .001, indicating that congruent trials are faster (Fig. 2c, d). We found a significant two-way interaction between congruency on the current trial and congruency on the previous trial, F(1, 47) = 9.30, MSE = 365, p = .004, indicating an overall Gratton effect. The three-way interaction between congruency on the current trial, congruency on the previous trial, and pupil dilation on the previous trial was also significant, F(1, 47) = 11.2, MSE = 360, p = .002. However, this indicated that participants had a larger adaptation effect after a small pupil size than after a large pupil on the previous trial. Post hoc tests revealed that participants had a significant adaptation effect after a small pupil size, F(1, 47) = 25.6, MSE = 289, p < .001, but not after a large one, F(1, 47) = .032, MSE = 436, p = .86. We also found a significant main effect for pupil dilation on the previous trial, F(1, 47) = 13.5, MSE = 339, p = .001.

In the HGLM pupil dilation analysis (see Table 2), we found a significant main effect for congruency (Cn), t(47) = 7.17, p < .001, β = 74.4, indicating that incongruent trials were slower. We also found a significant main effect for pupil dilation (Pn), t(47) = 3.53, p = .001, β = .086, indicating that pupil dilation leads to a longer RT. Importantly, the near zero value of the nonsignificant three-way interaction regression parameter indicates that the modulation by pupil size suggested by the pupil dilation ANOVA and visual inspection of Fig. 1c, d could be due to confounding variables. This indicates that we cannot draw any conclusions about the effect of pupil dilation on cognitive control. There was no difference between the two groups (tested under bright vs. dimmed light conditions; all two-sample ts < 1.17, all ps > .25).

Table 2 Pupil size analysis: coefficients of the HGLM Analysis with response time as dependent variable
Full size table

Discussion

First, we found that congruency on the current trial predicts both pupil dilation and blink on the current trial, consistent with findings of Siegle, Ichikawa, and Steinhauer (2008). More broadly, our findings on pupil dilation and blinking suggest that incongruent stimuli can trigger the autonomic system (Critchley, Tang, Glaser, Butterworth, & Dolan, 2005; Kobayashi, Yoshino, Takahashi, & Nomura, 2007).

Next, in our main analysis, we investigated the modulatory effect of pupil dilation (NE related) and eye blink (DA related) on the Gratton effect. We found a significant modulation of the Gratton effect by eye blink on the previous trial, even when we controlled for the effect of congruency on the current and previous trials. We failed to find a similar modulatory effect for pupil dilation.

Modulation of the Gratton effect by neuromodulatory markers is predicted by the adaptation-by-binding model (Verguts & Notebaert, 2008). This model proposes that the traditional roles of these neuromodulators in learning (Berridge & Waterhouse, 2003; Lisman et al., 2011) and cognitive control are not two separate functions but, instead, are strongly related. In particular, the DA burst underlying the blink may increase binding between stimulus, action, and more general task-relevant context elements. The modulatory role of DA in cognitive control is well-established in both computational models (e.g., Braver & Cohen, 2000; O’Reilly & Frank, 2006) and empirical studies (see Cools & D’Esposito, 2011, for a review). In particular, DA is involved in working memory (Brozoski, Brown, Rosvold, & Goldman, 1979) and has been proposed to be involved in regulating different computational trade-offs, such as flexibility versus stability (Cools & D’Esposito, 2011; Hazy, Frank, & O’Reilly, 2007). Dreisbach et al. (2005) observed that elevated DA levels increased distractibility; Dreisbach and Goschke (2004) found that positive affect (presumably triggering DA) had the same effect. In contrast, we find that a phasic burst of DA increases control. This is not necessarily in contradiction. First, in the task used by Dreisbach and colleagues, a task switch situation (switching between different goals) was implemented, in contrast to ours, where the goal was always the same. Also, the data pattern of Dreisbach and colleagues was modulated by DRD4 genotype (D4 being a D2-like receptor), whereas our effect is presumably D1-receptor dependent (see above). However, the exact relationships between these tasks and data remain to be determined. Similarly, theoretical and modeling accounts of NE in cognitive control focused on its role in regulating exploration versus exploitation (Aston-Jones & Cohen, 2005; Jepma & Nieuwenhuis, 2011), although empirical validation remains currently mixed (e.g., Jepma, Te Beek, Wagenmakers, Van Gerven, & Nieuwenhuis, 2010).

Since we find that blink, but not pupil dilation, modulates the Gratton effect, it is possible that specifically reward, not generally arousal, drives DA bursts in the current task. Reward is intimately connected to the dopaminergic system (Lisman et al., 2011; Schultz, 1998), and it may be more rewarding to execute an incongruent trial than a congruent trial (Molapour & Morsella, 2011; Schouppe et al., submitted; Silvetti, Seurinck, & Verguts, 2011). Incongruent trials themselves are aversive (Dreisbach & Fischer, 2012; Schouppe, De Houwer, Ridderinkhof, & Notebaert, 2012; Van Steenbergen et al., 2009), but successfully executing such difficult trials may be more rewarding than executing the easier congruent trials (Schouppe et al., submitted). This reward by self-evaluation may influence performance on the next trial(s) and improve binding between task-relevant representations (Waszak & Pholulamdeth, 2009). Also, in a Stroop task in which some stimuli are rewarded and others not, reward decreases the congruency effect (Krebs, Boehler, & Woldorff, 2010). Finally, reward increases the Gratton effect when reward is performance related, especially in reward-sensitive subjects (Braem et al., 2012). Of course, we cannot unambiguously equate the current DA modulation as a reward modulation phenomenon, given that different DA neurons have different functional characteristics (e.g., sensitivity to reward vs. novelty; see Bromberg-Martin et al., 2010, for a review). This needs to be further investigated with the current behavioral paradigm.

We expected to find a similar effect for pupil dilation, since NE was proposed to be the relevant neuromodulator by Verguts and Notebaert (2009). However, because of the inconsistency between the ANOVA and HGLM analysis, no strong conclusions can be drawn from the present data with respect to NE. The inconsistent outcomes of the ANOVA and the HGLM analysis may be due to the fact that the factors in the analyses are not orthogonal, since congruency causes pupil dilation. It may be possible that we did not find an effect for pupil dilation because trait anxiety determines the direction of the effect of NE release (De Rover et al., 2012). Another possibility is that the null effect for pupil dilation in the HGLM analysis is due to the fact that NE is released only in the early trials of the experiment when there is a high uncertainty in how to respond (Aston-Jones, Rajkowski, & Cohen, 1999; Dayan & Yu, 2006; Yu & Dayan, 2005). Since we used a congruent arrow flanker task (press left button for left arrow), the task may have been (over)learned rather quickly. Future research is needed to investigate the modulatory roles of eye blink and pupil size in more novel tasks.

We presently investigated cognitive control and found a phasic influence of DA. It would be of interest to see whether tonic and phasic DA would have a different effects on proactive and reactive control, respectively (Braver, 2012). In addition, since eye blink and pupil dilation are only indirect measures of DA and NE, respectively, direct neuroimaging and pharmacological manipulations are needed to investigate the roles of DA and NE nuclei in early and late stages of cognitive control.