Proportion congruency effects are the observation that the magnitude of the Stroop effect increases as the proportion of congruent trials in a block increases. Contemporary work shows that proportion effects can be specific to a particular context. For example, in a Simon task in which items appearing above fixation are mostly congruent and items appearing below fixation are mostly incongruent, the Simon effect is larger for the items appearing at the top. There is disagreement as to whether these context-specific effects result from simple associative learning or, instead, a type of conflict-mediated associative learning. Here, we address this question in an ERP study using a Simon task in which the proportion congruency effect was context-specific, manipulating the proportion of congruent trials based on location (upper vs. lower visual field). We found significant behavioral proportion congruency effects that varied with the specific contexts. In addition, we observed that the N2 response of the ERPs to the stimuli was larger in amplitude for the high congruent (high conflict) versus low congruent (low conflict) conditions/contexts. Because the N2 is known to be greater in amplitude also for trials where conflict is high and is believed to be an electrical signal related to conflict detection in the medial frontal cortex, this supports the idea that conflict-mediated associative learning is involved in the proportion congruency effect.
The conflict monitoring hypothesis (Botvinick, Cohen, & Carter, 2004; Botvinick, Braver, Barch, Carter, & Cohen, 2001; Carter et al., 1998) is the highly influential notion that cognitive conflict (e.g., response conflict) is detected in the brain, generating a signal that engages top–down control mechanisms that reduce conflict by modulating behavior. Although the original intent of the conflict monitoring hypothesis was to offer a solution to the homunculus problem of “what controls control,” the model has also provided an explanation for a wide range of findings, which have collectively challenged the traditional view that cognitive control is strictly rigid and top–down.
The most popular example of a dynamic, mainly stimulus-driven form of cognitive control is the trial-to-trial sequential compatibility effect, known simply as the Gratton effect (Gratton, Coles, & Donchin, 1992). The Gratton effect refers to the observation that the size of the interference effect in a range of tasks (e.g., Stroop, Flanker, and Simon tasks) is larger after a congruent trial than after an incongruent trial, where the term congruent refers to the nature of potentially conflicting signals in the stimulus (e.g., in the Stroop task, these would be the color name word and the color of the ink in which it appears; “red” in red ink is a congruent stimulus). Although there is evidence that feature priming contributes to this effect (Schmidt & de Houwer, 2011; Hommel, 2004; Mayr, Awh, & Laurey, 2003), most researchers also acknowledge the critical role of conflict adaptation.
In the conflict monitoring hypothesis, response conflict associated with an incongruent trial is detected by ACC. ACC then signals regions of the dorsolateral pFC that are involved in the maintenance of task rules, resulting in a control signal that increases attention to the task-relevant color dimension in the Stroop task. The result is that there is more attention to the task-relevant stimulus dimension after an incongruent trial resulting in slower congruent trials (less benefit from the irrelevant dimension) and faster incongruent trials (less interference from the irrelevant dimension). The neural mechanisms of the Gratton effect have been revealed by fMRI studies that identified dorsal ACC (dACC) as a region activated during conflict (Carter & van Veen, 2007; Kerns et al., 2004; MacDonald, Cohen, Stenger, & Carter, 2000; Carter et al., 1998). Many ERP studies have also found a larger N2 wave to conflict trials; this ERP peaks at 300 msec poststimulus, has a maximum amplitude over the midline central scalp, and has been shown to be consistent with a generator site in ACC based on neuroelectric modeling (van Veen & Carter, 2002a, 2002b; see Folstein & Van Petten, 2008, for a review). All of these neural results support the idea that dACC is part of a conflict detection mechanism (Botvinick et al., 2001, 2004).
This stimulus-driven account of control is highly appealing and has been applied to a wide range of findings (e.g., error-related slowing, semantic stem completion tasks, etc.), including those effects that have been traditionally dominated by strategic top–down explanations. Of these, the proportion congruency effect is perhaps the most well known. It refers to the observation that the size of an interference effect, the Stroop effect for example, is positively correlated with the frequency of congruent trials; that is, the size of the Stroop effect increases as the proportion of congruent trials increases (Blais, Harris, Guerrero, & Bunge, 2010; Lindsay & Jacoby, 1994; Cheesman & Merikle, 1986; Logan, Zbrodoff, & Williamson, 1984).
The dominant account of this finding is that participants learn that the color and word are correlated and then use this information to their advantage. Most commonly, this is framed as the participant adopting a strategy to allocate more attention to the word (e.g., read it) when there are many congruent trials or to disengage from the word (e.g., by focusing on a single letter within the word) when there are many incongruent trials. However, recent findings question the plausibility of this strategic account. First, it has been shown that participants are largely unaware of the proportion of congruent trials in a list. Blais et al. (2010) showed that, although participants were fairly accurate at judging the number of congruent trials in a list, for most blocks, they reported that they were “not sure” of the proportion in all but the most extreme proportion conditions (e.g., <15% or >85% congruent, Blais et al., 2010). Second, it is difficult to envision how this account could possibly explain the item-specific (e.g., Jacoby, Lindsay, & Hessels, 2003) and context-specific (e.g., Crump, Gong, & Milliken, 2006) class of proportion congruency effects.
These item- and context-specific proportion congruency effects refer to the observation that the size of the interference effect can be modulated by the proportion of congruent trials for specific items, or specific contexts, within a block of trials. For example, consider a Stroop task in which the set of items green, white, and red are presented congruently on 80% of trials and the set of items blue, yellow, and orange are presented congruently on 20% of trials. Although the items are randomly intermixed so that the participants cannot predict which item will come next, participants will show larger interference effects for the 80% congruent items than for the 20% congruent items (Jacoby et al., 2003). Similarly, if items appearing at one location are 80% likely to be congruent and items appearing at a different location are 20% congruent, then items appearing at the 80% location will show a larger interference effect (Crump et al., 2006).
There is a growing body of behavioral evidence for these item- and context-specific effects (e.g., Schouppe, De Ferrerre, Van Opstal, Braem, & Notebaert, 2014; Shedden, Milliken, Watter, & Monteiro, 2013; Crump & Milliken, 2009; Bugg, Jacoby, & Toth, 2008; Crump, Vaquero, & Milliken, 2008; Schmidt, Crump, Cheesman, & Besner, 2007; Crump et al., 2006). Although it is generally accepted that the effect results from conflict-mediated associative learning (Verguts & Notebaert, 2008, 2009; Blais, Robidoux, Risko, & Besner, 2007), an iteration of the conflict monitoring model (Botvinick et al., 2001) that explains these effects, there are some who argue that these effects have nothing to do with conflict adaptation but rather result from traditional associative learning (Schmidt & Besner, 2008) or compound-cue learning (e.g., Schmidt, 2013a, 2013b). All of these accounts acknowledge at some level that the participant must learn, probably implicitly (e.g., Crump et al., 2008; see Egner, 2014, for a review), that there is an association between the context and the degree to which the irrelevant and relevant dimensions are related to one another. They differ as to whether response conflict is the signal by which this implicit priming is necessary. The purpose of the current investigation is to use ERP to test whether we can adjudicate between these alternative views.
Throughout the paper we assume that the item-specific and context-specific proportion congruency effect arise via the same mechanism. Everyone in the literature does not share this view. For example, Schmidt and Besner (2008) have argued that the item-specific, but not the context-specific, effect can be explained by traditional associative learning mechanisms. However, more recently, Schmidt (2013a, 2013b) has argued that the context-specific effect can be explained by a compound-cue-based learning (see Risko, Blais, Stolz, & Besner, 2008, for similar accounts of proportion effects in spatial cueing). With respect to the current paper, the difference between conflict-mediated associative learning and these simple and compound-cue accounts is that the former claim that these proportion effects are driven by conflict.
As noted, an N2 component peaking around 300 msec over midline central scalp is consistent with cognitive conflict (Bartholow et al., 2005; van Veen & Carter, 2002a, 2002b; see Folstein & Van Petten, 2008, for a review). The current investigation therefore assesses whether the N2 component of the ERP waveform is sensitive to the context-specific proportion congruency effect. A larger N2 component in the high-conflict versus the low-conflict condition is consistent with an account in which these item- and context-specific proportion congruency effects result from conflict-mediated associative learning (e.g., Verguts & Notebaert, 2008, 2009; Blais et al., 2007). However, failure to observe an N2 would suggest that a traditional associative learning account (e.g., Schmidt & Besner, 2008) or compound-cue learning without conflict (e.g., Schmidt, 2013a, 2013b) is sufficient to explain these item- and context-specific proportion effects. To preview the results, we do observe an N2 over Cz approximately 300 msec poststimulus providing additional evidence that conflict-mediated associative learning is necessary to observe a context-specific proportion congruency effect.
Sixteen healthy volunteer college students from the University of California, Davis, served as participants. They were neurologically normal and had normal or corrected-to-normal vision. They were tested in accordance with the procedures prescribed by the University of California, Davis, human subjects institutional review board, provided informed consent, and were offered course credit for participation in partial completion of course requirements. All students have the option of opting for alternative assignments, in line with the Psychology Department policies.
Equipment and Recording
Continuous EEG data were recorded from 58 scalp sites using tin electrodes mounted in an elastic electrode cap. These electrodes were referenced to the right mastoid for recording. EOG data were simultaneously recorded with two electrodes placed above and below the left eye to detect eye blinks and two electrodes placed at the outer ocular canthi to detect lateral eye movements. Additional electrodes were placed on the left and right mastoid processes. EEGs were amplified, bandpass filtered (0.15–100 Hz), and digitized at 1000 samples per second using Synamps2 amplifiers running under Neuroscan 4.5. The EEG data were processed using EEGLAB 12.0 (Delorme & Makeig, 2004) and ERPLAB 4.0 (Lopez-Calderon & Luck, 2014). All channels were resampled to 250 Hz and rereferenced offline to the algebraic average of the left and right mastoids.
Stimuli and Procedure
Participants participated in a modified Simon task, as follows. A central fixation marker remained on the screen for the duration of the block. The letter X or O was presented in one of four locations (top left, top right, bottom left, or bottom right). Participants were asked to respond to the letter on the screen by pressing a left- or right-hand button (counterbalanced across individuals) with the appropriate index finger. The target remained on the screen until the response was executed. This was immediately followed by a uniformly distributed variable intertrial interval lasting between 500 and 1000 msec inclusive, in 17-msec increments, during which the fixation marker remained on the screen. Thus, from the point of view of the participant, there was a “+” sign on the center of the screen, and then approximately every 1.5 sec (depending on their RT), an X or O would appear in one of four corners.
There were 80 trials in each block. Half of the trials were congruent, and the other half were incongruent. A congruent trial is one in which the location of the stimulus is consistent with the response mapping. For example, if the participant is to press the left button when they see an X, then a congruent trial occurs when the X is presented on the left half of the screen, and the O is presented on the right half of the screen. Whether the stimulus was on the top or bottom of the screen affected the probability of the stimulus being congruent. For half of the participants, items appearing on the top of the screen were 75% congruent (i.e., 30 congruent trials and 10 incongruent trials), and items appearing on the bottom of the screen were 25% congruent (i.e., 10 congruent trials and 30 incongruent trials). Each block of trials lasted approximately 120 sec. Participants performed 20 blocks1 of experimental trials that were preceded by one block of practice trials during which the experimenter sat in the room and provided feedback about maintaining fixation to minimize eye movements.
Stimuli were presented using E-Prime 2.0 (Psychology Tools, Inc., Pittsburgh, PA) on a 20-in. LCD flat panel monitor running at its native resolution of 1280 × 1024 pixels. Participants were seated in a comfortable reclining chair approximately 115 cm from the screen. The X and O stimuli were presented 1.3° above/below fixation and 5.0° to the left/right of fixation, in Arial Black font; each letter was approximately 1.0° in height. Participants were instructed to respond as quickly and accurately as possible (equal weighting) and were told that they should be going fast enough to make at least a few errors.
Errors (6.7%, range: 2.4–21.3%) and posterror (6.1%, range: 2.4–17.7%) trials were excluded from the analysis. Following Hommel (1994), the remaining correct RTs were considered outliers if they fell outside the range of 200 ≥ RT ≥ 1000 msec. This procedure resulted in the elimination of 0.38% of trials (range: 0–1.11% across participants). For the EEG data, an independent components analysis was used to identify the signal corresponding to eye blinks and eye movements. These components were then removed from the data. Finally, a moving window that was 60 msec wide moving in increments of 20 msec to detect peak-to-peak voltage differences exceeding 80 μV across any channel was used to remove trials with excess electrical noise. This resulted in the exclusion of a further 2.0% of trials (range: 0–3.9% across participants).2
The RT and error rate data were analyzed using a 2 (Context: mostly congruent vs. mostly incongruent) × 2 (Congruency: congruent vs. incongruent) repeated-measures ANOVA.
There was a main effect of Context on RT, F(1, 15) = 12.83, p < .005, η2 = .461. RTs were faster in the mostly incongruent (420 msec) than the mostly congruent (426 msec) condition. There was also a main effect of Congruency (i.e., the Simon effect), F(1, 15) = 76.26, p < .001, η2 = .836. RTs made with the hand on the same side as the target were faster (405 vs. 436 msec). Importantly, the interaction between Context and Congruency was significant, F(1, 15) = 27.86, p < .001, η2 = .650. The size of the Simon effect in the mostly incongruent condition (410 vs. 430 msec = 20 ± 4 msec; t(15) = 5.26, p < .001, η2 = .648) was smaller than the size of the Simon effect in the mostly congruent condition (405 vs. 436 msec = 43 ± 5 msec; t(15) = 9.40, p < .001, η2 = .855).
For error rates, the main effect of Context was significant, F(1, 15) = 7.87, p < .05, η2 = .344. Fewer errors were made in the mostly incongruent condition (6.8% errors) than in the mostly congruent condition (7.5% errors). Again, there was a main effect of Congruency, F(1, 15) = 32.76, p < .001, η2 = .686, indicating a Simon effect for error rates (4.3% vs. 10.0% errors). Importantly, the interaction between Context and Congruency was significant, F(1, 15) = 31.91, p < .001, η2 = .680. The size of the Simon effect in the 25% congruent condition (4.8% vs. 8.7% = 3.9 ± 0.9% error rate; t(15) = 4.25, p < .001, η2 = .546) was smaller than the size of the Simon effect in the 75% congruent condition (3.8% vs. 11.3% = 7.4 ± 1.1% error rate; t(15) = 6.50, p < .001, η2 = .738).
The distribution of instantaneous scalp voltages at 0, 80, 160, 240, 320, 400, 480, and 560 msec poststimulus is shown in Figure 1. Figure 1A shows the Simon effect ERP difference waves potential maps (incongruent minus congruent) for the mostly incongruent and mostly congruent conditions, and Figure 1B shows the size of the context-specific proportion congruency interaction (mostly congruent Simon effect difference wave map minus mostly incongruent Simon effect difference wave map). Visual inspection of Figure 1B confirms our prediction of an N2 component that has maximal amplitude over midline central electrode sites, peaking at electrode site Cz (i.e., the vertex; e.g., Jurcak, Tsuzuki, & Dan, 2007). Therefore, as is common in the N2 literature (e.g., Nieuwenhuis, Yeung, van den Wildenberg, & Ridderinkhof, 2003), we focused our analyses on electrode Cz. ERP waveforms for each of the four conditions (Context × Congruency) are shown in Figure 2A, and Simon difference waves for each of the mostly congruent and mostly incongruent conditions are shown in Figure 2B.
For statistical analyses, we first determined the N2 amplitude peak latency at Cz by sliding a 36-msec (10 sample points) wide window over the difference wave for the overall Simon effect (i.e., collapsed over the proportion congruency effect). The peak of this window was 310 msec (between 292 and 328 msec) and is in line with the value of 350 msec obtained in a flanker task with mean RTs approximately 60 msec longer than in the current Simon task (van Veen & Carter, 2002b). We then quantified the mean voltage amplitude around the N2 peak latency, using the window from 292- to 328-msec poststimulus latency for each participant, in each condition, and analyzed these data using a 2 × 2 ANOVA with the same factors as in the behavioral analyses.
For the N2, there was no main effect of Context (F < 1), but there was a main effect of Congruency; scalp potentials in the 292- to 328-msec window at Cz were significantly more negative in voltage for incongruent trials than for congruent trials (5.06 vs. 6.02 μV), F(1, 15) = 6.42, p = .023, η2 = .300. Critically, the interaction of these factors was significant, F(1, 15) = 7.37, p < .05, η2 = .329, indicating that the size of the Simon effect in the mostly incongruent condition (5.78 vs. 5.45 μV = −0.33 ± 0.47 μV; t(15) < 1, η2 = .032) was smaller than the size of the Simon effect in the mostly congruent condition (6.27 vs. 4.69 μV = −1.60 ± 0.42 μV; t(15) = −3.78, p = .002, η2 = .488). Rather than use a constant 292- to 328-msec window for each participant, we can temporally localize the N2 for each participant using their overall Simon effect and compute mean amplitude over a 36-msec wide window centered on their peak N2 (e.g., see Varnum, Blais, Hampton, & Brewer, 2015). This approach to measure the relation between N2 amplitude and RTs yields r14 = −.46, p < .05, one-tailed. This relationship is shown in Figure 3.3
Readers for an earlier draft of this paper noted that we may in fact be looking at the MMN (see Folstein & Van Petten, 2008, for a review). Most commonly used in the study of auditory perception, the MMN is an ERP component related to the presentation of an odd stimulus in a sequence of stimuli. For example, in a sequence of 500-Hz tones occurring 80% of the time, a 2000-Hz tone would generate an MMN. The MMN also occurs in the visual domain (e.g., Tales, Newton, Troscianko, & Butler, 1999). It is a potential confound here because the very nature of proportion congruency effects as implemented here necessitates that there are frequently occurring trials (e.g., the congruent stimuli in mostly congruent condition and the incongruent stimuli in the mostly incongruent condition) and infrequently occurring trials (the incongruent stimuli in mostly congruent condition and the congruent stimuli in the mostly incongruent condition). Furthermore, the time window of this effect is roughly 300 msec with a scalp distribution similar to the conflict-mediated N2.
In the current experiment, we can look at the MMN by comparing the frequent stimuli (e.g., X in the top left and in the bottom right; O in the top right and in the bottom left) with the rare stimuli (e.g., O in the top left and in the bottom right; X in the top right and in the bottom left). Figure 4 shows the distribution of instantaneous scalp voltages at 0, 80, 160, 240, 360, 480, and 560 msec poststimulus for this contrast. The top and middle rows show voltages for the rare and frequent conditions, respectively, whereas the bottom row shows the MMN effect. It should be noted that, in the context of this experiment, the MMN is mathematically half the size of the context-specific proportion congruency effect. Thus, we cannot rule out that the MMN contributes to the effect—but clearly, the context-specific proportion congruency effect is more than just the MMN.
Congruency Sequential Effects
In the mostly congruent context, it is more likely that a congruent trial was preceded by a congruent trial than an incongruent trial. Similarly, in the mostly incongruent context, it is more likely that an incongruent trial was preceded by an incongruent trial than a congruent trial. This is problematic because we know that the size of the conflict effect is dependent on the congruency of the previous trial (e.g., Gratton et al., 1992). To access whether this could account for the context-specific proportion congruency effect we observe, we performed a Previous congruency × Context × Congruency analysis on the RTs (errors showed a similar pattern of results). There was a main effect of Congruency, F(1, 15) = 76.8, p < .001, η2 = .837. There was a Previous congruency × Congruency interaction, F(1, 15) = 102.2, p < .001, η2 = .872 (i.e., Gratton effect), indicating that the size of the Simon effect when the previous trial was congruent was larger (381 vs. 445 msec = 64 ± 6 msec; t(15) = 10.9, p < .001, η2 = .888) than the size of the Simon effect when the previous trial was incongruent (414 vs. 413 msec = −2 ± 4 msec; t(15) < 1, η2 = .025). Critically, there was also a Context × Congruency interaction, F(1, 15) = 22.9, p < .001, η2 = .603 (i.e., context-specific proportion congruency effect), indicating that the size of the Simon effect was larger in the mostly congruent context (393 vs. 434 msec = 41 ± 4 msec; t(15) = 9.7, p < .001, η2 = .862) than in the mostly incongruent context (402 vs. 423 msec = 21 ± 4 msec; t(15) = 5.3, p < .001, η2 = .652). All other effects from this analysis were not reliable (Fs < 1).
On the basis of this pattern of results, we conclude that the context-specific proportion congruency effect is not a by-product of the fact that the relative frequency of previous trial congruent differs across the mostly congruent and mostly incongruent contexts. Given that the proportion of congruent trials has little to no impact on the magnitude of the Gratton effect (e.g., Blais, Stefanidi, & Brewer, 2014), this is not particularly surprising.
Contextual Sequence Effects
King, Korb, and Egner (2012) reported that context-specific proportion congruency effects depend on the previous context. We thus performed a Previous context (mostly congruent vs. mostly incongruent) × Context (mostly congruent vs. mostly incongruent) × Congruency (congruent vs. incongruent) analysis. There is a main effect of Congruency, F(1, 15) = 82.32, p < .001, η2 = .846. There is also a Previous context × Congruency interaction, F(1, 15) = 50.0, p < .001, such that the size of the Simon effect is smaller after a mostly incongruent context (403 vs. 419 msec, 16 ± 3 msec; t(15) = 5.0, p < .001, η2 = .625) compared with a mostly congruent context (388 vs. 435 msec, 47 ± 4 msec; t(15) = 9.7, p < .001, η2 = .862). Critically, there is also a Context × Congruency interaction, F(1, 15) = 18.9, p < .001, η2 = .558, such that the size of the Simon effect is larger in the mostly congruent context (390 vs. 432 msec; t(15) = 9.5, p < .001, η2 = .857) than in the mostly incongruent context (401 vs. 422 msec; t(15) = 5.1, p < .001, η2 = .631). All other effects, including the three-way interaction, are not reliable (Fs < 1).4
The proportion of congruent trials in blocks of stimulus trials influences the magnitude of interference effects like those observed in Stroop or Simon tasks. As the proportion of congruent trials in a block increases, so does the degree of interference effects. Furthermore, these so-called proportion congruency effects can be specific to particular contexts. Here, we used ERPs to test whether context-specific proportion congruency effects result from simple associative learning as opposed to some type of conflict-mediated associative learning.
The conflict-mediated associative learning accounts (Verguts & Notebaert, 2008, 2009; Blais et al., 2007) propose that ACC mediates the rapid learning involved in the context-specific proportion congruency effect. Because the N2 component of the ERP waveform, which is maximal in amplitude over central scalp sites, is thought to arise via conflict adaptation from activity in ACC, it provides a way to test the role of ACC in context-specific proportion congruency effect (van Veen & Carter, 2002b). Hence, the current study sought to adjudicate between the two accounts of the context-specific proportion congruency effect by measuring the N2 component in a modified Simon tasks that included context-specific variations in the proportion of congruent trials within blocks of trials. If the conflict-mediated associative learning hypothesis is correct, a larger N2 effect will be observed for the high conflict (high congruent) versus low conflict (low congruent) conditions/contexts. More traditional associative learning accounts (Schmidt, 2013a, 2013b; Schmidt & Besner, 2008) explicitly state that the proportion congruency effect is unrelated to conflict adaptation; if correct, there should be no difference in the magnitude of the N2 component in the high versus low conflict condition. The present experiment revealed a larger N2 effect for the high conflict condition, supporting a conflict-mediated account of the context-specific proportion congruency effect.
Using behavioral results alone, it is admittedly difficult to disambiguate between these two accounts of the context-specific proportion congruency effect. However, studies looking at neural correlates have proven to be fruitful. In addition to the current study showing that the magnitude of the N2 varied as a function of contextual control, a recent fMRI study provides support for the conflict-mediated learning hypothesis. Blais and Bunge (2010) designed an fMRI experiment to reveal the local (item-specific) and global (list-specific) contributions to the proportion congruency effect using a Stroop task. It was found that ACC and the dorsolateral pFC, to a lesser degree, are involved in item-specific cognitive control. Specifically, activity in these regions tracked the amount of item-specific conflict but was uncorrelated with any independent form of global conflict. Given that the Schmidt and Besner (2008) account states that conflict adaption does not play a role in proportion congruency effects, it is difficult to see how either the current ERP study or the Blais and Bunge (2010) fMRI study could be reconciled with their account. Conversely, both studies fit well with the wide range of studies that support the idea that activity in ACC is associated with conflict resolution in processes such as response selection (e.g., Shackman et al., 2011).
These and similar findings have led to the proposal that attentional control can be both proactive and reactive (de Pisapia & Braver, 2006). Proactive control is consistent with the traditional concept of cognitive control. It refers to the set of mechanisms responsible for functions such as maintaining arbitrary task rules, sustaining attention to the task, and overt strategies for optimizing performance. The notion of automatic, or reactive, control is relatively recent. It is related to stimulus-driven, bottom–up forms of control, which are thought to modulate the proactive control mechanisms set in place by the task demands. In the context of the Stroop task, proactive control mechanisms related to activity in the dorsolateral pFC (MacDonald et al., 2000) implement the instruction “attend, and respond, to the color.” Reactive control mechanisms related to performance monitoring area of cortex such as dACC continually assess the need for control using cues such as (i) response conflict (Botvinick, Nystrom, Fissell, Carter, & Cohen, 1999), (ii) the likelihood of committing an error (Brown & Braver, 2005), or (iii) actually committing an error (Rabbitt, 1966). Detecting a strong signal for any of these cues has a similar effect, increasing reliance on proactive control.
What are the characteristics of control in these context-specific proportion congruency tasks? It would seem that many of the “rapid” control effects, such as the Gratton sequential compatibility (Gratton et al., 1992) and item-specific (Jacoby et al., 2003) and context-specific proportion congruency effects (Crump et al., 2006), are likely the result of this newer notion of reactive control. Furthermore, it has argued elsewhere (e.g., Blais, 2010) that many aspects of reactive control are likely implicit, in contrast to the (often) deliberate aspects of proactive control. Therefore, it becomes important to specify several characteristics for the specific type of cognitive control to which one is referring.
With respect to the item- and context-specific congruency effects, it appears that they are implicit and reactive. Crump et al. (2008) report two important pieces of evidence that favor this view. First, participants are largely unaware of the manipulations of proportion (e.g., see also Blais et al., 2010; Schmidt et al., 2007). Second, informing participants of the proportion manipulations has no impact on performance. That is, the magnitude of the item-specific/context-specific proportion congruency effects is the same for both naive and informed participants.
King et al. (2012) have provided evidence that these effects were not proactive. They had participants perform a face-gaze variant of a location-specific proportion congruency Erikson flanker task, in which high-conflict stimuli were presented to one visual hemifield and low conflict stimuli were presented to the other hemifield. The use of faces allowed them to look at sustained versus transient activity in the fusiform face area. If proactive control was used in this task, there should be persistent activity in the contralateral hemifield (i.e., “Targets on the left are harder; therefore, I will continuously attend to that side of space.”). In contrast, however, they observed activity in the contralateral fusiform face area only when a target was presented, consistent with the idea that the control system is reacting to the stimulus and not “preparing” for it.
Schmidt (2013a, 2013b) has proposed that the proportion congruency effect arises not from changes in conflict adaption but rather from a shift in learning when to respond. In his computational model, this is implemented as decreasing the threshold before a response is made. Thus, system adjusts this response threshold as a function of the amount of time taken to make a response. This results in fast trials (i.e., congruent trials) having a lower response threshold than slow trials (i.e., incongruent trials). On the surface, it seems very similar to how the conflict monitoring models (e.g., Blais & Verguts, 2012; Verguts & Notebaert, 2008; Blais et al., 2007; Botvinick et al., 2001) operate.
In the proof of concept experiment, Schmidt (2013a, 2013b) presented participants with bright or dim letters in contexts that were either mostly bright or mostly dim. These data show an interaction such that there is a larger contrast effect in the mostly bright than in the mostly dim condition. However, the pattern of this interaction is quite different than what is typically seen in proportion congruency experiments.
However, it seems to us that any of the existing models (e.g., Blais & Verguts, 2012; Verguts & Notebaert, 2008; Blais et al., 2007; Botvinick et al., 2001) can readily explain this “temporal learning” result. In these models, the input to the conflict monitoring node is Hopfield energy, which is defined as the product of all the activations in the output layer. Lateral inhibition within the output layer ensures a winner-takes-all system where the sum of all responses is constrained. Such a system has an important property: When uncertainty is high (as will be the case when all of the responses have relatively equal output), the amount of Hopfield energy is high, and when uncertainty is low (as will be the case when one of the responses is very different from the others), the amount of Hopfield energy is low. In the mostly dim condition, activation for the correct response is going to accrue, on average, at a slower rate than in the mostly bright condition precisely because there are more dim trials. Critically, this means that the output layer will remain in a state of higher uncertainty under the mostly dim condition leading to relatively slower RTs on bright trials and relatively faster RTs on dim trials in comparison with the mostly bright condition.
It is difficult to rule out whether participants are adapting to conflict or adapting to time-to-response. However, given that the temporal learning hypothesis (1) provides a relatively poor account of the list-level proportion effect and (2) is unable to account for many of the other “conflict monitoring” effects such as error-related slowing and the Gratton effect and that (3) Schmidt (2013a, 2013b) concedes “[t]hese inconsistencies might indicate that something different, such as conflict adaptation, occurs on top of the temporal learning effect…” (p. 10), we feel more evidence is necessary to warrant whether the temporal learning hypothesis remains viable.
The fact that a larger N2 effect is observed in the mostly congruent context where a larger behavioral interference is observed and there is a correlation between N2 and RT measures of the context-specific proportion congruency effect provides compelling evidence for a correlation between the RT measure of interference and the ERP measure of conflict. It does not, however, tell us definitively that the mechanism by which this occurs is through conflict-mediated Hebbian learning. For example, Burle, Allain, Vidal, and Hasbroucq (2005) investigated whether it is conflict or compatibility that determines sequential compatibility effects in the Simon task. They found that sequential effects were not determined on the conflict measured on the previous trial but on whether the previous trial was compatible or incompatible, independently of whether more or less conflict was experienced on the previous trial. The same could be true here; more conflict (indexed by N2) is measured in the mostly incongruent context, but we do not know whether this is because people learned the association between how much conflict they experienced and this context or the likelihood that the trial would be compatible or incompatible on this context. Thus, participants might learn that, in a given context, the presented target would be incompatible. The context might activate the appropriate processing for dealing with this specific conflict, and therefore, less conflict would be observed. This is consistent with the observed pattern of results in the literature that there can be several sources of the conflict measured on a given trial, but only the source of conflict (e.g., Simon compatibility) occurring on the previous trial that is of the same type of conflict as the one measured in the current trial does in fact produce sequential effects (i.e., Torres-Quesada, Funes, & Lupianez, 2013; Funes, Lupianez, & Humphreys, 2010).5
There are now at least two dozen published studies demonstrating the presence of item, or, more generally, context-specific proportion congruency effects. It is difficult to conceive of an account of these effects that fails to rely on some type of associative learning. The few studies that have looked at the neural correlates of this activity (e.g., the current study, King et al., 2012; Zurawska Vel Grajewska, Sim, Hoenig, Herrnberger, & Kiefer, 2011; Blais & Bunge, 2010) have consistently found regions implicated in the conflict monitoring components of the cognitive control networks (e.g., ACC). It is difficult to reconcile these neural findings with traditional simple associative learning algorithms.
The data were collected, and an early version of the paper was written, when C. B. was a postdoctoral fellow at the Center for Mind and Brain. C. B. is now at Arizona State University. We thank Sharon Corina, Javier Lopez-Calderon, and Steven J. Luck for their support and assistance and Juan Lupiáñez and an anonymous reviewer for their critiques on an earlier version of this paper. This research was supported by NIMH grant MH055714 to G. R. M.
Reprint requests should be sent to Chris Blais, Department of Psychology, Arizona State University, Tempe, AZ, or via e-mail: firstname.lastname@example.org.
Computer failures resulted in the loss of 13 blocks of EEG data spread across seven participants. These blocks were repeated.
The pattern of results reported is stable both when all artifacts are left in the data (indicating that the occurrence of blinks/saccades is random with respect to stimulus onset) and when bipolar vertical and horizontal electrooculogram channels are computed offline and used to detect trials on which a blink or saccade occurred between stimulus onset and 600 msec by using a moving window that was 60 msec wide moving in increments of 20 msec to detect peak-to-peak voltage differences exceeding 80 μV resulting in the exclusion of a further 19.8% of trials (range: 1.9–39.7% across participants).
Using the constant 292- to 328-msec window for each participant yields a correlation between N2 amplitude and the context-specific proportion congruency effect in RTs of r14 = −.42, p = .10.
Another way to do this analysis is to test whether the previous context repeats/switches. If the variance is parsed in this manner, the Previous context × Congruency two-way interaction is now labeled the Previous context repeat/switch × Context × Congruency three-way interaction, and we come to the same conclusion that size of the context-specific proportion congruency effect is sensitive to the previous context.
We thank Juan Lupiáñez for pointing this out.