Comparing dual process theories:

Evidence from event-related potentials

Adrian P. Banks
University of Surrey, UK

In W. De Neys (Ed.), Dual Process Theory 2.0. New York, NY: Psychology Press

1
Overview
Research on the electrophysiology of reasoning is comparatively rare, but it has the potential
to offer considerable insights into the time course of cognitive processes and contribute to a
wide range of theoretical questions such as the role of dual processes in reasoning. Although a
behavioural response to a reasoning problem can indicate a single time point at which a complex
series of cognitive events ends, event-related potentials (ERPs) can be used to examine the
timing of different events as they unfold during the reasoning process. That is, it is possible to
measure cognitive events in the window between presentation of the problem and the
behavioural response. Theories differ crucially about what occurs in this window, and ERPs
offer the potential to observe this activity. A small number of studies have been conducted with
the aim of identifying the electrophysiological correlates of reasoning on tasks that have been
used more widely to examine dual process theory. In this chapter I will review the ERP research
that has aimed to test dual process theories of reasoning, discuss the findings, and explore both
the potential and the limitations of this technique. Finally, I will discuss how the theoretical
implications of these findings support the idea that Type 1 processes are fast and automatic,
occur in parallel, and when acquired over time, can reproduce any thinking process that can be
automated, including both normatively correct logical responses and belief-based responses
within a belief bias task.

Belief bias and dual process theories

So far, ERP studies of dual process theory have examined the widely studied belief bias
paradigm. Belief bias is the tendency to judge conclusions to reasoning problems based on prior
beliefs rather than logical validity (Evans, Barston, & Pollard, 1983; Klauer, Musch, & Naumer,
2000). Although belief bias has been studied in a number of different reasoning tasks, the
seminal experimental paradigm was developed by Evans et al. (1983). In this, participants are
asked to evaluate the logical validity of a conclusion to a syllogism. The problems are constructed
to independently manipulate both logical validity and belief of the conclusion. For example:

No cigarettes are inexpensive

Some addictive things are inexpensive
Therefore, some addictive things are not cigarettes

This syllogism has a conclusion that is both valid and believable. Whereas:

2
 No addictive things are inexpensive
Some cigarettes are inexpensive
Therefore, some cigarettes are not addictive

This syllogism has a conclusion that is valid but unbelievable. Evans et al. found that
participants were more likely to accept valid conclusions, as instructed, but were also more likely
to accept believable conclusions, and interestingly there was a greater effect of belief on invalid
than valid conclusions. These key findings have been replicated many times and have been an
important line of evidence in developing and testing dual process theories (e.g. De Neys, 2006;
Evans & Curtis-Holmes, 2005; Stupple, Ball, Evans, & Kamal-Smith, 2011).
Dual process theories propose two types of thinking, one referred to as heuristic, Type 1, or
System 1 thinking and the other referred to as analytic, Type 2, or System 2 thinking. In this
chapter I will mostly use the terms Type 1 and 2 thinking, but this largely refers to the same
concept as System 1 and 2. De Neys (this volume) has characterised three types of dual process
models, and explaining the phenomenon of belief bias has played a notable part in developing
and contrasting these models. Many studies of belief bias have explained the effect using the
default-interventionist model (Evans, 2006; Evans & Stanovich, 2013). This is a serial process in
which a fast, heuristic, belief-based judgement is made initially, but analytic processes may
override this response with the logical solution under certain circumstances. Parallel models, in
contrast, propose that both of these types of thinking are engaged simultaneously from the
beginning of the reasoning process (e.g. Sloman, 1996). This allows for a conflict to be detected
if the two types of thinking lead to different responses, but raises other questions such as
whether it is maladaptive to always engage both processes when in most cases the simpler
process will provide the best response. Finally, hybrid models such as the logical intuition model
(this volume; De Neys, 2012), the three-stage dual-process model of analytic engagement
(Pennycook, Fugelsang, & Koehler, 2015), or the parallel model of Trippas and Handley (this
volume; Handley & Trippas, 2015) draw on elements of both serial and parallel models. They
propose an initial response made using Type 1 processes followed by the subsequent
engagement of analytic Type 2 processes under certain circumstances. However, they propose
that the initial Type 1 processes can occur in parallel. Furthermore, amongst the parallel Type 1
processes are those that are logical and cue the normatively correct response. This contrasts with
the assumption in other dual process models that Type 1 processes tend to cue heuristic
responses that may not be normatively correct.

3
 These three types of dual process models differ in fundamental ways in their description of
how reasoning occurs. There are two areas in which ERP research is beginning to contribute to
resolving the debate between these models. First, the temporal resolution of ERPs means that
they can be used to assess if and when reasoning occurs serially or in parallel. Second, the fact
that ERPs can be measured prior to a behavioural response means that they can be used to
examine the initial fast reasoning processes to assess if these cue biased or normatively correct
responses.

Event-related potentials and reasoning

An electroencephalogram, or EEG, is a recording of the electrical activity of the brain.

Embedded within it is the electrical activity generated as a result of many different cognitive
processes. In order to extract from this the response to a specific event, the same event is
repeated across many trials, the same time window is extracted from each of these trials, and
then the time windows are averaged together. This process means that random noise in the
signal is averaged out across trials, leaving a clearer signal that is the ERP. The ERP is a series of
positive and negative voltage deflections, or components, that are labelled P or N to indicate a
positive- or negative-going component and numbered to indicate the relative position in the
waveform (P1 is the first positive component, P2 is the second positive component, etc.). Each
of these components relates to different cognitive processes, so finding differences in either the
amplitude or latency of components provides information about differences in the cognition
between experimental conditions. It provides a continuous measure of cognitive processing
between the presentation of the task and the behavioural response, and this is done with high
temporal resolution.
Initial work on reasoning has begun to identify the components associated with several
standard tasks. The most common component that is influenced by reasoning complexity is the
P3. This is a positive deflection that typically occurs 300 ms to 500 ms after an event and
indicates updating of working memory (Polich, 2007). It is found in conditional reasoning
(Bonnefond & Van der Henst, 2013; Bonnefond, Kaliuzhna, Van der Henst, & De Neys, 2014),
relational reasoning (Bonnefond, Castelain, Cheylus, & Van der Henst, 2014), and syllogistic
reasoning (Qiu, Li, Luo, Zhang & Tu, 2009). However, these experiments vary in their procedure
and so the full interpretation of the P3 in reasoning has not yet been settled. The N2 has also
been found in conditional reasoning (Bonnefond et al., 2014) and the Wason selection task (Cai,
et al., 2011). The N2 is a negative deflection that is typically found 200 ms to 350 ms after the

4
event and indicates the resolution of conflict (Folstein & Van Petten, 2008). Late positivity or
positive slow waves have also been found in conditional reasoning (Bonnefond & Van der
Henst, 2009) and syllogistic reasoning (Luo et al., 2012). Despite the name, this component is
still found before the behavioural response to a reasoning task (e.g. 1000 ms to 2000 ms after the
stimulus). The cognitive processes associated with this component are les d d, b
c d a da (G ,S ,L ,L ,B a ., 1996) a
a (Ga c ́a-Larrea & Cézanne-Bert, 1998). Finally, a P2
component has been found, although rarely (Bonnefond & Van der Henst, 2009). The P2
component indicates the detection of expected features (e.g. Luck & Hillyard, 1994) and in
language comprehension tasks it is greater for expected than unexpected words (Federmeier,
Mai, & Kutas, 2005).

Event-related potentials and dual process theories of reasoning

The first ERP study that relates to dual process theory investigated the effect of belief bias on
syllogistic reasoning (Luo, Yuan, Qiu, Zhang, Zhong & Huai, 2008). Syllogisms were constructed
with beliefs that were either consistent with or conflicted with the logically correct answer,
referred to by the authors as facilitatory and inhibitory beliefs. There was also a baseline
condition in which there was no connection between the two premises. Premises were presented
serially the major premise was presented, followed by the minor premise, followed by the
conclusion. Participants responded after the presentation of the conclusion. The ERPs were time
locked to the onset of the minor premise. No difference in accuracy was found between the
facilitating and inhibiting beliefs, but response times were shorter for facilitating beliefs. The
main ERP finding was an increased positivity between 300 ms and 500 ms and 1000 ms and
1600 ms for both facilitatory and inhibitory beliefs compared to baseline and an increased
positivity between 300 ms and 600 ms for inhibitory compared to facilitatory beliefs. The
authors suggest that this positive component indicates that more attentional resources were
required to activate and apply rules of inference.
However, this study has some limitations that restrict the interpretation of these findings.
First, the task differed from the typical paradigm used for studying belief bias. Whereas a typical
task requires the evaluation of a conclusion that may or may not be believable, this study used a
production task in which premise believability is manipulated to influence reasoning. As the
ERPs were time locked to the minor premise rather than the conclusion, it is the role of beliefs
in facilitating the integration of minor premise with major premise that is the focus of this study

5
rather than the conflict of belief and logic in the evaluation of a conclusion. Also, by time
locking to the onset of the minor premise, the recording conflates the reading process with the
reasoning process. Second, all of the problems were invalid, and the same syllogistic format was
used in all of the 160 reasoning trials (all A are B; all C are B; therefore, the relation between A
and C is indefinite). Given the high level of accuracy (96% and 87% for facilitatory and
inhibitory conditions), it is possible that the response became well practiced and responses made
through recognition rather than reasoning. Furthermore, as no valid syllogisms were used, it is
not possible to test the characteristic interaction of logic and belief that is assessed in typical
belief bias studies.
Luo, Yang, Du, and Zhang (2011) applied the same approach to studying belief bias in
conditional reasoning. Modus tollens and denial of the antecedent arguments were presented
with facilitatory or inhibitory beliefs. ERPs were time locked to the onset of the minor premise,
and greater negativity was observed in the 400-ms to 600-ms and 800-ms to 1600-ms windows
for inhibitory compared to facilitatory modus tollens, but no differences were found for denial of
the antecedent arguments. The authors suggest that the earlier negativity could be a delayed N2,
which is a component associated with detecting conflict and could indicate that a conflict
between logic and belief responses has been detected (Folstein & Van Petten, 2008). But the
timing of the negativity is much later than is typically found for an N2, which questions this
interpretation. No component is proposed to account for the later negativity. The differences
between modus tollens and denial of the antecedent arguments was explained in terms of
different strategies that may have been adopted based on the training received, but there is no
direct evidence that this is what occurred.
As in Luo, Yuan, Qiu, Zhang, Zhong, and Huai (2008), time locking to the minor premise
moves the focus of this study to integration of the premises. The same logical argument is again
repeated many times each one is presented sixty times raising the possibility that the
responses are learnt rather than reasoned. Neither Luo et al. (2008) nor Luo et al. (2011) report
ratings of the strength of belief in the problems, and weak beliefs undermine the belief bias
phenomena (Banks, 2013). The items in the paper do not seem compelling, for example, an
b b I a c , b a .S b d
a .T c . a d a b a a a c b
elicit the effect, although accuracy was significantly greater with facilitatory problems. The main
concern though is that the pattern of ERPs cannot be readily explained in terms of the
components that are evoked and the theoretical mechanisms that they might indicate.

6
 Luo et al. (2012) shift the focus to the evaluation of conclusions in syllogisms in which the
logically validity of the conclusion either conflicted or did not with the believability of the
conclusion. This study is more comparable with the majority of behavioural studies of belief bias
and therefore provides more direct evidence to differentiate between competing theoretical
accounts of belief bias and dual process theories. Luo et al. presented the major and minor
premises together, followed by the conclusion, at which point a response was made. The ERPs
were time locked to the response. This provides an indication of the reasoning process when the
response is decided upon rather than the reasoning process as the premises are integrated and
the conclusion is presented. Conflict problems were found to evoke more positivity -400 ms to -
200 ms prior to the response than no conflict problems. The authors suggest that this late
positive component (so called because it is one of the later ERP components, but it still precedes
the behavioural response) reflects cognitive control as the heuristic response is inhibited and the
analytic response made. If so, this finding provides initial support for dual process theories of
belief bias, as it demonstrates that a conflict between belief and logic must be resolved prior to
responding.
Although this study enables a stronger conclusion to be made about responses to belief bias
problems than the previous studies, there are still some methodological limitations. Only one
valid and one invalid syllogism were used in all the conditions, each repeated eighty times and
potentially leading to a learnt rather than a reasoned response. There was also no assessment of
the strength of belief of the conclusions, with some conclusions not seeming to evoke a
c b , a , T , a b a d a d b ac d
c a . T ac a a acc c aa a d c
accuracy between conflict and no conflict problems: 53% and 58%, respectively.
Overall, these studies demonstrate the potential to record ERPs during reasoning and evoke
differences between conflict and no-conflict problems (or inhibition and facilitation). But no
clear pattern emerges across the studies. With syllogisms, both Luo et al. (2008) and Luo et al.
(2012) find that inhibition or conflict is associated with early positivity although one is time
locked to the minor premise and the other to the response whereas with conditionals Luo et al.
(2011) find that inhibition is associated with early negativity, but only for modus tollens
inferences. It is hard to draw a persuasive theoretical interpretation from these contrasting
findings. This could be for a number of methodological reasons that were already highlighted
earlier. In our studies we sought to overcome these methodological issues and provide a clearer
test of the electrophysiology of reasoning.

7
 Methodologically, we (Banks & Hope, 2014) sought to develop a task that enabled ERPs to
be assessed whilst ensuring as far as possible comparability with standard tasks in this domain in
order that the findings could be used to test current theories of dual process theory.
Theoretically, our aim was to exploit the precise temporal resolution of ERPs to compare serial
and parallel dual process theories. To do this, we measured the ERPs evoked by the conclusion
rather than the premises or time locked to the response as has been the case in previous
research. The problems were designed so that it was not possible to evaluate the conclusion until
the final word was presented. ERPs were time locked to the onset of this word. We manipulated
both logic and belief in a full factorial design and used a wider range of different problem types
than in previous studies. Relational reasoning problems were used, as there are a wide range of
possible problems, they are relatively pure strategically, and they are not prone to variations in
premise interpretations (Roberts, 2000). A nonsense term was used to ensure that no premise
conflicted with prior knowledge, as this can be difficult to integrate into a coherent
a (K a a ., 2000). H a a a ca b :

Premise 1 Giraffes are bigger than elephants

Premise 2 Zoots are bigger than elephants
Premise 3 Mice are bigger than zoots
Initial conclusion Mice are bigger than
Final conclusion word Elephants

The main prediction of this experiment contrasts serial and parallel dual process models. If
the initial evaluation of the conclusion is based only on belief or only on logic, as predicted by
serial models, the ERP elicited by the conclusion will only be influenced by one factor (e.g. the
believability but not the logical validity of the conclusion). If the initial evaluation of the
conclusion is based on both belief and logic, as predicted by parallel models, the ERP elicited by
the conclusion will be influenced by the interaction of both factors (e.g. problems where logic
and belief conflict will differ from problems where there is no conflict). The main finding of this
study was a larger P3 component for both conditions in which logic and belief conflict (logically
valid but unbelievable and logically invalid but believable conclusions) than for both conditions
in which logic and belief do not conflict (logically valid and believable and logically invalid and
unbelievable conclusions). This supports the parallel dual process models, as it indicates that
both belief and logic are processed at the same time, and very rapidly after the presentation of
the conclusion within the 300-ms to 500-ms time window. Furthermore, there was no

8
difference in the latency of the peak of the P3 component between the four conditions. This also
indicates that the initial evaluation of the conclusion is influenced by both belief and logic in
parallel rather than serially.
Our interpretation of these findings is that participants are constructing a representation, a
mental model, of the relations between the items in the problem as they read the premises. This
is a dynamic process in which the mental model is updated to incorporate new information as
each word of the problem is read. Participants typically seek to construct a single, coherent,
isomeric representation of relational reasoning problems in which new items and relations are
incorporated and represented efficiently (Schaeken, Van der Henst, & Schroyens, 2007). As the
conclusion is read, it, too, is represented in some form. In the no-conflict condition this
conclusion is coherent with the current representation, but in the conflict condition the
conclusion is inconsistent and requires the representation of the problem in working memory to
be updated. As the P3 is a component that is associated with updating of working memory, the
P3 amplitude is greater in the conflict condition. Although further research is required to explore
exactly how dual processes influence working memory updating during reasoning, the empirical
findings as they stand provide support for models in which the initial evaluation of the
conclusion is influenced by both belief and logic rapidly and in parallel rather than serially.
Our second study sought to examine the time course of belief and logic-based processes
further (Banks & Dunne, in prep.). A typical belief bias task is based on a paradigm in which the
logical task is rather difficult, often a syllogism with low levels of accuracy, and a belief that is
compelling. From this paradigm it is concluded that the belief judgement is fast and occurs by
default and the logic judgement is slow and may not occur at all. There is a wide range of
evidence supporting this serial or default-interventionist model (Evans & Stanovich, 2013). More
recently, evidence has been presented for a parallel processes interpretation of this phenomenon.
These theories suggest that both belief- and logic-based processing occur in parallel but that they
vary in time course according to factors such as the complexity of the judgement. That is,
different, less complex, logical problems might be faster to evaluate than the belief judgement.
This research implies the apparently default nature of belief judgements and that slower
intervention of logical analysis is a consequence of the type of problems that have been studied
historically rather than an inherent property of logic and belief.
Handley, Newstead, and Trippas (2011) provided the first evidence for this. They introduced
a manipulation in which participants evaluated the believability of logical problems as well as the
more common logical validity of belief-based problems. They found that logic influenced the
evaluations of belief more than belief influenced the evaluation of logic. This is not consistent

9
with the application of a belief heuristic by default. In this case logic-based processing appears to
have been applied b d a , c a c ba . S c d, d d d ca
problems to those typically employed in belief bias research. They chose problems with simpler
logical forms than is typical: conditionals and disjunctions. They suggest that these simpler forms
may lead to inferences that are made automatically because of the familiarity to participants. As a
result, the inferences are made without intentional control and conflict with the belief-based
judgement. Trippas, Thompson, and Handley (2017) provide further evidence for parallel logic
and belief-based processes that vary in their time course. They manipulated the complexity of the
logical problems and were able to show that when the processing of complex logical problems
was greater than the processing of belief judgements, belief interfered with logic. But with
simpler logical problems the processing was less than the processing of belief judgements, and
then logic interfered with belief. It is the relative complexity of the logic and belief judgement
task that determines the interaction, rather than an inherent property of logic or belief.
These behavioural studies introduce new ways of examining the dual process theories by
manipulating instructions and the logical complexity. Our second study again exploited the
temporal resolution of ERPs to provide direct evidence for the time course of logic and belief-
based processing. We used the instructional manipulation of Handley et al. (2011) in which
participants judged either the believability or the logical validity of the problems in order to
examine belief as well as logical processing. We used a simpler logical form in order to examine
familiar inferences that may occur more automatically than the complex task used in our
previous study. However, we did introduce some methodological changes in order to assess
ERPs adequately. As before, we studied relational reasoning (three term series) and time locked
to the onset of the final word of the conclusion. It was not possible to evaluate the conclusion
until this word was presented. A typical problem is:

Premise 1 Sheep are bigger than toads

Premise 2 Toads are bigger than hippos
Conclusion Sheep are bigger than hippos

We also adapted the belief judgement task in order to more closely parallel the logical task. In
each case there were either two believable statements and one unbelievable statement or one
believable statement and two unbelievable statements. The belief judgement task was to assess if
the statements were mostly believable or mostly unbelievable. This could only be resolved when
the final statement was read. Therefore, much like the logical task, each statement had to be read

10
in turn, a representation of them held in working memory, and a belief-based evaluation of all
three made after they have been presented. This is similar to a process of logical reasoning in
which the premises are read and held in working memory and a logical evaluation is made when
the third statement, the conclusion, is presented.
As in our earlier study, we aimed to compare serial and parallel dual process models by
examining the timing of the influence of logic and belief on conclusion evaluation. Serial models
predict an initial influence on ERPs of just one factor (e.g. belief) whereas parallel models predict
that both factors influence the initial evaluation of the conclusion, causing a difference between
no conflict problems (where the two factors lead to the same evaluation) and conflict problems
(where they lead to different evaluations).
Under logic instructions we found that the judgements accurately reflected the logical validity
of the problems but were not influenced by their believability. However, under belief
instructions we found that the judgements accurately reflected the believability of the problems
and that responses were influenced by their logical validity. That is, logic influenced belief-based
judgements, but that belief did not influence logic-based judgements. This replicates the
behavioural findings of Hand a. a c ad a .
The key question though is the influence that the task instructions have on the ERPs. The
main finding was an interaction between task instructions and conflict such that the difference in
P2 amplitude between no-conflict and conflict conditions was greater under belief task
instructions than logic task instructions. This difference was found in the 125-ms to 200-ms time
window an earlier stage of the process than our first experiment that used a more complex
reasoning task. P2 is a component that in sentence comprehension tasks is larger for expected
words than unexpected words (Federmeier et al., 2005). Our interpretation of these findings is
that the expectation of the final concluding word can be influenced by both belief and logic.
Where expectations from each of these factors are the same (i.e. in no-conflict problems), the
final word is more expected than in conflict problems where different expectations are generated
by belief and logic. Therefore under belief instructions, both belief and logic are influencing the
conclusion evaluation in parallel because in this condition the P2 was greater for no-conflict than
conflict problems. However, interestingly, no difference between no-conflict and conflict
problems was found under logic instructions. In this condition the believability of the conclusion
did not influence the judgements of logical validity.
These findings provide further support for parallel models, as logic and belief can influence
conclusion evaluation at the same time on presentation of the conclusion. Also, as there was an
influence of logic under belief instructions, they provide further evidence that with simpler logic

11
problems, automatic logical inferences are made that can influence judgements by default and
without intentional control.

Parallel processing and automatic inferences: evidence from ERPs

The ERP studies described earlier provide a new line of evidence for investigating dual
process theory. Empirically, there are several new findings. The main finding that was present in
some form in all of the studies was that conflict (or inhibitory belief) problems differed from no-
conflict (or facilitatory belief) problems. Although conflict detection has repeatedly been found
in behavioural studies, these ERP data add to the findings by providing more precise
information about when the cognitive processing of conflict and no-conflict problems begins to
differ. This occurred early in the evaluation (typically within the first 1000 ms), and Banks and
Hope directly assessed the latencies and showed no difference between belief and logic
conditions. The findings show that there is not an initial difference between conditions based on
belief followed by a later conflict with logic. Instead, the initial difference contrasts conflict with
no-conflict problems that both logical and belief-based evaluations have been made in parallel.
The second set of empirical findings relates to the ERP components elicited by reasoning,
and this pattern of findings is less clear. The two studies that time-lock to the minor premise
(Luo et al., 2008; Luo et al., 2011) find opposite effects increased positivity and increased
negativity, respectively, for conflict problems. There are also some concerns about the methods
used in these studies, and resolving these will lead to clearer findings about the integration of the
minor premise in reasoning. Examining conclusion processing, Luo et al. (2012) and Banks and
Hope (2014) both used a typical belief bias task and found increased positivity in conflict
problems late positivity for Luo et al. and a P3 for Banks and Hope. These components
indicate that the application of attentional resources is elicited by the conflict, either exerting
cognitive control in inhibiting a response or updating a representation required by the task.
These studies used complex logical problems, and increased attentional resources are required to
resolve the conflicting responses from difficult logic and simple belief. In contrast, the simpler
logical problems tested by Banks and Dunne (in prep.) evoked differences in an earlier
component, the P2, which indicates an expectation for a final concluding word. There was no
difference in the P3. This indicates that differences in the cognitive processing of conflict and
no-conflict problems emerge prior to the updating of working memory representations. This
suggests more automatic, less working memory dependent processes for the simpler three-term
series problem. Interestingly, this effect was greater under belief instructions than logic

12
instructions, indicating that the automatic expectations generated by the logical structure of the
problem had a greater effect on belief judgements than automatic beliefs had on logical
judgements. Under belief instructions the simple three-term series apparently generated an
automatic logical inference that conflicted with the belief judgements, despite no instructions to
reason logically. That is, the normatively correct logical response was generated by an automatic
process.
These conclusions are still tentative as the small number of studies and range of
methodological approaches used mean that there is not yet a solid body of replicated empirical
evidence that firmly establishes the ERP components that are found when reasoning.
Nonetheless, it is possible to consider the theoretical implications of the findings to date.
It is difficult to reconcile these findings with a serial, default-interventionist model in which
an initial judgement is made based solely on heuristic processes followed under certain
circumstances by a later logical judgement based on analytic processes. None of the studies
found a time window prior to the behavioural response in which conditions differed only by
belief-based heuristic, with an interaction with logic intervening later. They all indicate an
evaluation of logic and belief in parallel. Several different parallel models have been proposed,
and the ERP studies further discriminate between them. One traditional theory is that analytic
and heuristic processes are engaged from the onset of the problem (Sloman, 1996). Although
this is a possibility, it is notable that the components influenced by conflict are comparatively
early (i.e. 200 to 300 ms after conclusion presentation), whereas this model has typically not
emphasised fast logical processing; that is, in particular, logical responding for a typical belief bias
task is generally expected to be analytic and slow. If the conflict arises because of the
simultaneous engagement of slow analytic processing and fast heuristic ones, it should be
expected to arise much later (i.e. when the slow logical process has completed its response
generation) than the 300-ms to 500-ms time window that was found by Banks and Hope (2014).
The findings are easier to reconcile with more recent parallel models (or so-ca d b d
models, De Neys, this volume), in particular the logical intuition model (De Neys, 2012) and the
three-stage dual process model of analytic engagement (Pennycook et al., 2015), or the parallel
model of Trippas and Handley (Handley & Trippas, 2015). These models predict an initial stage
involving only fast, Type 1 processes that operate in parallel followed by a later stage in which
analytic processes generate a response if the Type 1 processes conflict. These models account for
the ERP data well, explaining the differences between conflict and no-conflict problems that are
found soon after the presentation of the conclusion.

13
 Different ERP components are associated with different cognitive processes, so the ERP
studies not only provide data about the timing of dual processes, but also the components
elicited provide information about the type of cognitive process evoked by the different
problems. The most interesting distinction is between the complex logical problems tested by
Luo et al. (2012) and Banks and Hope (2014) and the simpler logical problems tested by Banks
and Dunne (in prep.). The former tasks were more typical of tasks used to study belief bias and
conflict evoked components associated with working memory updating and cognitive control.
The simpler logical problems, however, evoked differences in an earlier component associated
with expectations generated automatically as the problem is read. Future research on the
different components evoked by different reasoning tasks will generate further insight into the
different elements and stages of the reasoning process as it unfolds.
The ERP evidence presented here demonstrates that the technique is best suited to examining
fast, Type 1 processes that are not directly measurable with behavioural paradigms. It is well
suited to further test the theoretical accounts of Type 1 processes. One possibility is that
automaticity is the basis for Type 1 logical intuitions (e.g. De Neys, 2012), and the
complementary theory has also been proposed, namely that working memory involvement is the
basis of Type 2 analytic processes (Evans & Stanovich, 2013). Extensive research on
automaticity on the acquisition and qualities of automatic cognition has identified different
mechanisms for automatic cognition, including a shift towards direct memory retrieval (Logan,
1988), the proceduralisation of an algorithm after repeated practice (Anderson, 1982), or the
consistent mapping between stimulus and response (Schneider & Shiffrin, 1977). If these
mechanisms apply to reasoning, what forms of reasoning could be automated? A three-term
series problem could be readily automated, either because it is more frequently encountered or
because the simple form means that there is a more consistent mapping of the response to the
stimulus, or both. But it may be that only some reasoning can be automated because the
mapping between the stimulus and response is highly variable, providing a limit on what can and
cannot become a Type 1 process. For example, this is likely to be the case when reasoning about
moral dilemmas, as the utilitarian, Type 2 process will generate different responses to different
dilemmas depending on what the alternative is. If, indeed, some Type 1 processes are automated
inferences, the further implication is that a dual process theory requires more than an accurate
account of the underlying cognitive architecture, but also an account of the interaction of the
architecture with different properties of the reasoning task such as the consistency of mapping
between premises and conclusion. Answering these questions requires further research on the
speed of processing and level of attentional control for different problems. The studies described

14
earlier show how ERPs can be used to measure the latency of cognitive processes, and the
amplitude of different components can be used to index different cognitive processes, such as
the allocation of attention. ERP research can contribute to answering these theoretical questions.

ERP methodology

Although ERP studies offer new lines of evidence, they also bring new methodological
challenges. All of the experiments described earlier have introduced some changes to the typical
presentation of reasoning problems, and this is true of all EEG studies of reasoning. It is a
necessary response to the constraints of ERP methodology and worth evaluating to what extent
these changes limit the interpretation of the findings (Roser et al., 2015). The studies described
earlier differ from the typical behavioural paradigms in two ways. First, there is a very large
number of trials. The large number of trials is a potential limitation, as it is possible that
responses are learnt rather than reasoned. If so, then the ERPs would reflect the recognition of a
problem within a given experimental paradigm rather than the intended cognitive process. We
mitigated this issue by using a wider range of problems which were constructed so that it was not
possible to predict the conclusion before it was presented. Our second study also used a number
of filler problems to reduce the predictability of the trials. We also found the expected
behavioural effects in the responses of participants, indicating that the study elicited the
phenomena of interest. Future ERP studies should continue to minimise this methodological
constraint. Second, information is presented serially (one sentence or word at a time) rather than
in parallel (the entire problem presented at once). This has been found to influence reasoning
performance, thereby reducing belief bias (Morley, Evans, & Handley, 2004). It is a necessary
constraint in order to minimise eye movement artefacts that render the EEG too noisy to
analyse and to identify an exact point of interest to time-lock the ERPs to.
Whereas the former issue is a methodological problem, the latter raises interesting theoretical
questions. When a dual process theory describes two processes occurring in parallel, at what
point does this begin? Even when using a parallel presentation format, when reading a problem,
the information is not actually absorbed in parallel but sequentially as each word is read. For
example, a typical base-rate neglect problem presents base-rate information about the population
followed by information about a specific individual. One set of information is used to reason
normatively and the other heuristically. Differences in time at which both of these reasoning
processes end (measured by a behavioural response) may reflect not just how fast each process
is, but also the point in time at which the reasoning began. If one process begins before the

15
other, then it may finish first and trigger a behavioural response, but it is not necessarily a faster
process overall. More precise timing of reasoning can help distinguish if the cognitive processes
truly are occurring in parallel or if strategic online processing as the problem is read or
interleaving of cognitive processes is giving rise to that impression.

Conclusion

The main contribution of ERP research at present lies in the investigation of fast, Type 1
processes that are difficult to tap using behavioural paradigms. The studies described here
illustrate different ERP components that are evoked by different logical problems, indicating the
timing and the type of cognition involved prior to the behavioural response. The early
interaction of belief and logic in all of the problems suggests that these factors are evaluated in
parallel at the onset of the conclusion rather than serially. The simpler logical problems differed
from the more complex logical problems and evoked an automatic logical inference that
interfered with belief judgements, suggesting that within a belief bias task, the Type 1 processes
could be logic based as well as belief based.
Together, these findings provide evidence about the initial cognitive response to reasoning, an
area where dual process theories disagree. These findings are most consistent with parallel Type
1 processes that may generate belief- or logic-based responses. The data support the idea that
these responses are automatic and will be acquired when a person has repeated exposure to
consistent patterns, be they logical forms or other patters of inference. Examining this idea
requires further work on the conditions under which automatic inferences can be acquired and
the timing and cognitive processes that index their formation. Developing reasoning tasks for
ERP studies will help in answering these questions.

16
References
Anderson, J. R. (1982). Acquisition of cognitive skill. Psychological Review, 89, 369 406.
Bago, B., & De Neys, W. (2017). Fast logic? Examining the time course assumption of dual
process theory. Cognition, 158, 90 109.
Banks, A. P. (2013). The influence of activation level on belief bias in relational reasoning.
Cognitive Science, 37, 544 577.
Banks, A. P., & Dunne, L. (in prep.). Logic and belief instruction in belief bias: Evidence
from event-related potentials. Manuscript in preparation.
Banks, A. P., & Hope, C. (2014). Heuristic and analytic processes in reasoning: An event-
related potential study of belief bias. Psychophysiology, 51, 290 297.
Bonnefond, M., Castelain, T., Cheylus, A., & Van der Henst, J. B. (2014). Reasoning from
transitive premises: An EEG study. Brain and Cognition, 90, 100 108.
Bonnefond, M., Kaliuzhna, M., Van der Henst, J. B., & De Neys, W. (2014). Disabling
conditional inferences: An EEG study. Neuropsychologia, 56, 255 262.
B d, M., & Va d H , J. B. (2009). W a b ind an inference? An EEG study
with conditional arguments. Neuropsychologia, 47, 3125 3133.
Bonnefond, M., & Van der Henst, J. B. (2013). Deduction electrified: ERPs elicited by the
processing of words in conditional arguments. Brain and Language, 124, 244 256.
Cai, X., Li, F., Wang, Y., Jackson, T., Chen, J., Zhang, L., & Li, H. (2011).
Electrophysiological correlates of hypothesis evaluation: Revealed with a modified
Wa c a . Brain Research, 1408, 17 26.
De Neys, W. (2006). Dual processing in reasoning: Two systems but one reasoner. Psychological
Science, 17, 428 433.
De Neys, W. (2012). Bias and conflict: A case for logical intuitions. Perspectives on Psychological
Science, 7, 28 38.
Evans, J. St. B. (2006). The heuristic-analytic theory of reasoning: Extension and evaluation.
Psychonomic Bulletin & Review, 13, 378 395.
Evans, J. S. B., Barston, J. L., & Pollard, P. (1983). On the conflict between logic and belief in
syllogistic reasoning. Memory & Cognition, 11, 295 306.
Evans, J. S. B., & Curtis-Holmes, J. (2005). Rapid responding increases belief bias: Evidence
for the dual-process theory of reasoning. Thinking & Reasoning, 11, 382 389.
Evans, J. S. B., & Stanovich, K. E. (2013). Dual-process theories of higher cognition:
Advancing the debate. Perspectives on Psychological Science, 8, 223 241.

17
Federmeier, K. D., Mai, H., & Kutas, M. (2005). Both sides get the point: Hemispheric
sensitivities to sentential constraint. Memory & Cognition, 33, 871 886.
Folstein, J. R., & Van Petten, C. (2008). Influence of cognitive control and mismatch on the
N2 component of the ERP: A review. Psychophysiology, 45, 152 170.
Ga c ́a-Larrea, L., & Cézanne-Bert, G. (1998). P3, positive slow wave and working memory
load: A study on the functional correlates of slow wave activity. Electroencephalography and
Clinical Neurophysiology/Evoked Potentials Section, 108, 260 273.
Gevins, A., Smith, M. E., Le, J., Leong, H., Bennett, J., Martin, N., McEvoy, L., Du, R., &
Whitfield, S. (1996). High resolution evoked potential imaging of the cortical dynamics of
human working memory. Electroencephalography and Clinical Neurophysiology, 98, 327 348.
Handley, S. J., Newstead, S. E., & Trippas, D. (2011). Logic, beliefs, and instruction: A test of
the default interventionist account of belief bias. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 37, 28 43.
Handley, S. J., & Trippas, D. (2015). Chapter two-dual processes and the interplay between
knowledge and structure: A new parallel processing model. Psychology of Learning and
Motivation, 62, 33 58.
Klauer, K. C., Musch, J., & Naumer, B. (2000). On belief bias in syllogistic reasoning.
Psychological Review, 107, 852 884.
Logan, G. D. (1988). Toward an instance theory of automatization. Psychological Review, 95,
492 527.
Luck, S. J., & Hillyard, S. A. (1994). Electrophysiological correlates of feature analysis during
visual search. Psychophysiology, 31, 291 308.
Luo, J. L., Liu, X., Stupple, E. J., Zhang, E., Xiao, X., Jia, L., Yang, Q., Li, H., & Zhang, Q.
(2013). Cognitive control in belief-laden reasoning during conclusion processing: An ERP
study. International Journal of Psychology, 48, 224 231.
Luo, J. L., Yang, Q., Du, X. M., & Zhang, Q. L. (2011). Neural correlates of belief -laden
reasoning during premise processing: An ERP study. Neuropsychobiology, 63, 112 118.
Luo, J. L., Yuan, J. J., Qiu, J., Zhang, Q. L., Zhong, J., & Huai, Z. C. (2008). Neural correlates
of the belief-bias effect in syllogistic reasoning: An event-related potential study.
Neuroreport, 19, 1075 1079.
Morley, N. J., Evans, J. S. B., & Handley, S. J. (2004). Belief bias and figural bias in syllogistic
reasoning. Quarterly Journal of Experimental Psychology Section A, 57, 666 692.
Pennycook, G., Fugelsang, J. A., & Koehler, D. J. (2015). What makes us think? A three -stage
dual-process model of analytic engagement. Cognitive Psychology, 80, 34 72.

18
Polich, J. (2007). Updating P300: An integrative theory of P3a and P3b. Clinical
Neurophysiology, 118, 2128 2148.
Qiu, J., Li, H., Luo, Y., Zhang, Q., & Tu, S. (2009). The neural basis of syllogistic reasoning:
An event-related potential study. Brain Research, 1273, 106 113.
Roberts, M. J. (2000). Strategies in relational inference. Thinking and Reasoning, 6, 1 26.
Roser, M. E., Evans, J. S. B., McNair, N. A., Fuggetta, G., Handley, S. J., Carroll, L. S., &
Trippas, D. (2015). Investigating reasoning with multiple integrated neuroscientific
methods. Frontiers in Human Neuroscience, 9, 41.
Schaeken, W., Van der Henst, J., & Schroyens, W. (2007). The mental models theory of
a a a : C c a , a d c c . I W. Sc a , A.
Va d d c , W. Sc , & G. d Yd a (Ed .), The mental model theory of reasoning:
Refinements and extensions (pp. 129 150). Mahwah, NJ: LEA.
Schneider, W., & Shiffrin, R. M. (1977). Controlled and automatic human information
processing: I. detection, search, and attention. Psychological Review, 84, 1 66.
Sloman, S. A. (1996). The empirical case for two systems of reasoning. Psychological Bulletin,
119, 3 22.
Stupple, E. J., Ball, L. J., Evans, J. S. B., & Kamal-Smith, E. (2011). When logic and belief
collide: Individual differences in reasoning times support a selective processing model.
Journal of Cognitive Psychology, 23, 931 941.
Trippas, D., Thompson, V. A., & Handley, S. J. (2017). When fast logic meets slow belief:
Evidence for a parallel-processing model of belief bias. Memory & Cognition, 45, 539-552.

19