A genuine marker of the automaticity of reading in the stroop task
For the past four decades or so, an idea contrasting early definitions of automaticity
and claiming that automatic processes can be controlled, has dominated the literature
(Logan, 1980, 1985; Norman and Shallice, 1986; Tzelgov et al., 1990). The interference
effect found in the Stroop task is usually considered to be a marker of automaticity
of reading, while the modulation of its magnitude is referred to as a marker of control.
In the present article, we emphasize the frequently overlooked notion that what we
refer to as a marker of automaticity has in fact multiple origins. MacLeod and MacDonald
(2000) and Goldfarb and Henik (2007) suggested that two types of conflict—the task
and informational conflicts—contribute to Stroop interference. The informational conflict
(henceforth IC) represents competition between two color concepts: one that is activated
through color naming and the second that is activated by the reading process (e.g.,
the concepts red and blue respectively, when the stimulus is the word BLUE written
in red ink). However, according to the same authors, some amount of interference is
obtained even with color-unrelated words, since all words automatically activate the
irrelevant reading task, setting in motion the competition between two possible tasks
(henceforth task conflict; TC) (see Kalanthroff et al., 2013a; Entel et al., submitted,
for behavioral evidence, and Bench et al., 1993; Carter et al., 1995, for neuroimaging
evidence of the TC). Even non-word stimuli containing lexical information (e.g., letter
strings) can interfere because they are readable (Klein, 1964; Sharma and McKenna,
1998). The more word-like the stimulus, the more interference it produces (Monsell
et al., 2001). Thus, as evident from this distinction, the genuine marker of automaticity
is the TC whereas the IC amplifies the interference from the irrelevant task. That
is, in order to argue that the automatic reading process can be controlled one should
actually show that what is controlled is the TC.
List-wide proportion congruent effect and the conflict adaptation account
The more frequent the incongruent trials are in an experiment, the smaller the Stroop
effect (Logan and Zbrodoff, 1979; Logan, 1980; Tzelgov et al., 1992). This is known
as the list-wide proportion congruent effect because the proportions are manipulated
at the list level. The list-wide proportion congruent effect is considered to be a
marker of control since it demonstrates the modulation of the magnitude of the Stroop
effect, and as such, is interpreted in terms of conflict adaptation. According to
the conflict-monitoring framework (Botvinick et al., 2001; De Pisapia and Braver,
2006; Braver, 2012), an increased proportion of incongruent trials results in higher
conflict at the response level, which triggers the control system. The control process
includes two stages: conflict detection and control exertion. Referring to our previous
discussion, in order to claim the automatic reading process can be controlled, the
TC should be the target of both stages of the control process. However, according
to our analysis, this is not the case. In fact, the TC is only a target of the control
exertion stage. According to the models within the conflict-monitoring framework,
conflict reduction is achieved through adjusting the weights of the two tasks, thereby
minimizing the TC. However, the conflict detection stage is centered on response competition,
which requires the TC to be amplified by the IC. When there is no IC, that is, no
competing color-concept activation by reading, no competing color-response can be
activated. Focusing on response competition (and thereby on IC) by Botvinick et al.
and later models (De Pisapia and Braver, 2006; Blais et al., 2007; see also Verguts
and Notebaert, 2008, for a model integrating cognitive control and reinforcement learning)
leads to the conclusion that the detection of conflict, and therefore triggering of
the entire control process, requires the IC being present (see Kalanthroff et al.,
2013b, for evidence inconsistent with this assumption). There is no “path” in these
architectures allowing TC to be monitored without the presence of the IC (Figure 1).
That is, the theory behind these architectures in their current state does not allow
an unequivocal claim that reading, as an automatic process, can be controlled.
Figure 1
The proposed architectures within the conflict-monitoring framework. (A) All models
share the same core architecture introduced by Cohen et al. (1990) in their explanation
of Stroop effect performance. This includes the definition of conflict as response
competition, implying an aggregated contribution of task conflict and informational
conflict. The assumption that conflict is controlled solely by adjustment of task
representation weights implies that only the task conflict can be directly controlled.
(B) Botvinick et al.'s (2001, 2004) model added a conflict-monitoring unit thereby
generating a control loop for adjusting the task representation weights, while (C)
Blais et al. (2007) proposed that task weights can be adjusted differentially for
specific items. (D) De Pisapia and Braver's (2006) architecture captures the distinction
between reactive and proactive control. (E) Verguts and Notebaert's (2008, 2009) model
suggests control is modulated through conflict-based Hebbian learning. Note the models
are depicted in a very schematic way, with no reference to the nature and direction
of the existed connections, their specific weights, etc. The detailed information
can be found in the original articles (see references). R, red; G, green; C, color;
W, word; ACC, anterior cingulate cortex; LC, locus coeruleus; MFC, medial frontal
cortex; PFC, prefrontal cortex; DLPFC, dorsolateral prefrontal cortex.
Item-specific proportion congruent effect: Conflict adaptation vs. learning account
The conflict adaptation account has been challenged by manipulating the proportions
of incongruent trials at the item level, revealing an item-specific proportion congruent
effect (Jacoby et al., 1999, 2003). In the item-specific paradigm, list-wide proportion
congruence is held at 50%, and specific words are paired in most of the trials with
a specific color, creating mostly congruent (i.e., not associated with conflict) or
mostly incongruent (i.e., strongly associated with conflict) stimuli. The item-specific
proportion congruent effect refers to a smaller interference for mostly incongruent
items than for mostly congruent items.
In order to determine whether a word causes conflict, the word should be read, which
contradicts the main assumption of the models in the conflict-monitoring literature,
including those specifically adapted to explain the item-specific findings (Blais
et al., 2007), that control operates proactively to prevent the initiation of the
reading process. Consequently, it has been proposed (Bugg et al., 2008, 2011), and
supported by empirical data (Bugg and Hutchison, 2012; Abrahamse et al., 2013), that
item-specific control may be based on reactive control. This idea, however, is inconsistent
with the assumption that reading, as an automatic process, is ballistic (Bargh, 1989),
that is, difficult to stop once started. Stopping a ballistic reading process seems
especially unlikely given that it is completed extremely quickly (Sereno et al., 1998;
Cohen et al., 2000; Pulvermuller et al., 2001). More important, as the previous section
illustrates, the conflict adaptation account can only explain the pattern obtained
for incongruent (i.e., producing IC) items in mostly incongruent and mostly congruent
conditions. However, manipulating the proportions at the item level seems also to
affect the congruent (i.e., producing no IC) items, as evidenced (in our view) by
the results of Jacoby et al. (2003). In that study, a 50/50 condition in which the
number of presentations of words in each color was equal for congruent or incongruent
stimuli was included in addition to the mostly congruent and mostly incongruent conditions.
The analysis carried out by the authors showed that the larger the proportion of incongruent
items was (from mostly congruent to 50/50 to mostly incongruent), the lesser the Stroop
effect obtained. However, the 50/50 condition can be defined not just as a condition
including more incongruent items than the mostly congruent condition, but also as
a neutral condition where the conflict cannot be predicted by reading. Jacoby et al.'s
data reveal that in comparison to the “neutral” (50/50) condition, incongruent items
in the mostly incongruent condition were 32 ms faster. Similarly, and surprisingly,
congruent items in the mostly congruent condition also showed a 21 ms reaction time
(RT) reduction. Identical information regarding the conflict is provided by reading
congruent words in the 50/50 and mostly congruent conditions, and yet RT in the latter
condition is faster. This pattern contradicts the conflict adaptation account since
congruent items do not produce IC, which according to our analysis, is the basis for
control modulation. Schmidt et al. (2007; Schmidt and Besner, 2008) proposed a contingency
learning account to explain Jacoby et al.'s (2003) finding without assuming conflict
adaptation. It postulates that pairing a word with a specific color creates an association
between that word and a specific response. The mechanism of contingency learning functions
by lowering the threshold of the most frequently encountered response to the word,
and does not lower the thresholds of other possible responses. Since according to
the contingency learning account it does not matter if the word is paired mostly with
congruent or incongruent colors, the facilitative effect of learning predicted by
the contingency learning account is consistent with the results of Jacoby et al. (2003).
To prove the independence of the contingency learning mechanism of conflict, Schmidt
and Besner (2008) demonstrated that the effects of contingency learning and congruency
(i.e., IC) are additive by reanalyzing Jacoby et al.'s (2003) data. This evidence,
however, is somewhat problematic because the rearrangement of the cells in the design
manipulating proportion congruency still has the (congruency) confound, and the effect
of contingency learning cannot be validly evaluated in such an analysis. In order
to test directly whether contingency learning depends on the presence of conflict,
Schmidt et al. (2007) (also Schmidt and Besner, 2008) conducted an experiment in which
they eliminated IC by using neutral (i.e., color-unrelated) words only as stimuli
in a color naming task. Their results demonstrated that the contingency learning effect
does not require a stimulus to be a color-related (i.e., conflicting) word, supporting
the idea that contingency learning is independent of the presence of IC. However,
as suggested by MacLeod and MacDonald (2000), even neutral words are conflicting with
respect to TC. Thus, the contingency learning effect might be independent of IC, but
not of TC. Although such dependency would not weaken the ability of the contingency
learning to explain the item-specific proportion congruent effect, it would suggest
that this account might actually represent another control-like adaptive mechanism
activated by (task) conflict. In fact, such evidence would dissipate the core controversy
(i.e., control vs. learning) around the interpretation of the conflict adaptation
effect, by incorporating the contingency learning into the category of control mechanisms.
Another potential problem with the contingency learning account is that it assumes
that the association learned refers to a particular response in the sense of the button
that should be pushed, but not in the sense of the correct color. This claim, supported
by the results of their Experiment 4, is explicitly stated by Schmidt et al. (2007):
“… pairings of stimuli do not simply form semantic connections… but instead directly
cause changes in our behavior …” It is also evident in the architecture of the proposed
parallel episodic processing (PEP) model (Schmidt, 2013) where the response generation
layer consists of representations of the buttons the responses are mapped to, but
not of the response set colors. If so, then it posits the questions of what would
happen, and how contingency learning would express itself when instead of pushing
the keys on a keyboard, responses are made vocally. When the response requires naming
the color, there is no other way contingency learning can proceed but through linking
the word with a specific color-concept because the latter is necessary for making
a verbal response. That is, with vocal responses, contingency learning is predicted
to affect the informational and not the response level of representations. However,
if the words already have a strong semantic association with the color concept (i.e.,
congruent condition) then the contribution of the contingency learning process should
be minimal, if at all. Therefore, with respect to the current discussion, the congruency
of the item, or in other words, informational conflict or its absence, in some situations,
might matter even for the contingency learning process.
Summary
We do not pretend that the distinction between task and information conflict can solve
the ongoing argument regarding the mechanism behind the “flexibility” of the Stroop
effect, as reflected by the proportion effect. We do believe that the awareness of
the fact that only one of two components contributing to the Stroop effect is a genuine
marker of the automaticity of reading, would undoubtedly help in further developing
existing control models, and probably new ones, that would be able to answer the question
regarding controllability of reading. Distinguishing between two types of conflict
can also be valuable with respect to the “control vs. learning” debate. For now, the
proposed learning mechanism (i.e., contingency learning), as an alternative explanation
for some of the proportion congruent effects, has only been proven to be independent
from the IC. However, as mentioned, the TC is what really matters. Hence, in order
to be considered as an independently standing mechanism that is not part of the control
system, the contingency learning should also be evident when no TC is produced by
stimuli.
Conflict of interest statement
The authors declare that the research was conducted in the absence of any commercial
or financial relationships that could be construed as a potential conflict of interest.