Eighth edition published 2020

by Routledge
2 Park Square, Milton Park, Abingdon, Oxon OX14 4RN
and by Routledge
52 Vanderbilt Avenue, New York, NY 10017
Routledge is an imprint of the Taylor & Francis Group, an informa business
© 2020 Michael W. Eysenck and Mark T. Keane
The right of Michael W. Eysenck and Mark T. Keane to be
identified as authors of this work has been asserted by them
in accordance with sections 77 and 78 of the Copyright,
Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted
or reproduced or utilised in any form or by any electronic,
mechanical, or other means, now known or hereafter
invented, including photocopying and recording, or in any
information storage or retrieval system, without permission
in writing from the publishers.
Trademark notice: Product or corporate names may be
trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
First edition published by Lawrence Erlbaum Associates 1984
Seventh edition Published by Routledge 2015
Every effort has been made to contact copyright-holders.
Please advise the publisher of any errors or omissions, and
these will be corrected in subsequent editions.
British Library Cataloguing-in-Publication Data
A catalogue record for this book is available from the British Library
Library of Congress Cataloging-in-Publication Data
A catalog record has been requested for this book

ISBN: 978-1-13848-221-0 (hbk)

ISBN: 978-1-13848-223-4 (pbk)
ISBN: 978-1-35105-851-3 (ebk)

Typeset in Times New Roman by

Servis Filmsetting Ltd, Stockport, Cheshire

Visit the companion website: www.routledge.com/cw/eysenck.

9781138482210_COGNITIVE_PSYCHOLOGY_PRE_CHAP_1.indd 6 28/02/20 2:15 PM

 Chapter

14 Reasoning and hypothesis

testing

INTRODUCTION
For hundreds of years, philosophers have distinguished between two kinds
of reasoning. One is inductive reasoning, which involves drawing general
conclusions from premises (statements) referring to particular instances.
A key feature of inductive reasoning is that the conclusions of inductively
valid arguments are probably (but not necessarily) true.
The philosopher Bertrand Russell provided the following example. A
turkey might use inductive reasoning to draw the conclusion “Each day I
am fed”, because that has always been the case in the past. However, there
is no certainty that the turkey will be fed tomorrow. Indeed, if tomorrow is
Christmas Eve, it is likely to be proven false.
Scientists very often use inductive reasoning similarly to Russell’s
hypothetical turkey. A psychologist may find across numerous experiments
that reinforcement (reward) is needed for learning. This might lead them
to use inductive reasoning to propose the hypothesis that reinforcement is
KEY TERMS essential for learning. This conclusion is not necessarily true because future
experiments may not replicate past ones.
Inductive reasoning
Forming generalisations The other kind of reasoning identified by philosophers is deductive
(that may be probable reasoning. Deductive reasoning allows us to draw conclusions that are
but are not certain) definitely or certainly valid provided other statements are assumed to be
from examples or true. For example, the conclusion Tom is taller than Harry is necessarily
sample phenomena; see
true if we assume Tom is taller than Dick and Dick is taller than Harry.
deductive reasoning.
Deductive-reasoning problems owe their origins to formal logic.
Deductive reasoning An important issue is whether the distinction between inductive and
Reasoning to a conclusion
deductive reasoning is as clear-cut in practice as it appears above. There is
from a set of premises
or statements where increasing evidence that similar processes are involved in both cases. For
that conclusion follows example, Stephens et al. (2018) asked participants to evaluate identical sets
necessarily from the of arguments after receiving inductive- or deductive-reasoning instructions.
assumption the premises With the former instructions, participants decided whether the conclusion
are true; see inductive
was plausible, strong or likely to be true. With the latter instructions, they
reasoning.
decided whether the conclusion was necessarily true.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 666 28/02/20 4:14 PM

 Reasoning and hypothesis testing 667

Stephens et al. (2018) found participants used the same processes KEY TERMS
whether instructed to reason inductively or deductively. The only major
Informal reasoning
difference was that greater argument strength was required to decide that a A form of reasoning
conclusion was necessarily true (deductive condition) than to decide it was based on one’s relevant
strong or likely to be true. knowledge and
The wide chasm between the artificial, logic-driven, deductive-­reasoning experience rather than
logic.
tasks traditionally used in the laboratory and everyday reasoning in the
form of argumentation has led to a rapid increase in research on informal Falsification
reasoning. Informal reasoning (discussed later, pp. 694–701) is based on Proposing hypotheses and
then trying to falsify them
our knowledge and experience rather than logic. It is a form of inductive
by experimental tests; the
­reasoning that resembles our everyday reasoning. logically correct means
A major consequence of this shift in research is that it is increasingly by which science should
accepted that reasoning processes often resemble those used in judgement work, according to Popper
and decision-making. For example, the Bayesian approach, according to (1968).
which our subjective probabilities (i.e., that X is dishonest) are adjusted in
the light of new information, plays a prominent role in theorising about
judgements (see Chapter 13). In a similar fashion, the Bayesian approach is
increasingly applied to reasoning (Navarrete and Mandel, 2016).

HYPOTHESIS TESTING
Karl Popper (1968) distinguished between confirmation and falsifica-
tion. Confirmation involves the attempt to obtain evidence confirming
or supporting one’s hypothesis. In contrast, falsification involves the
attempt to falsify hypotheses by experimental tests. Popper claimed we
cannot achieve confirmation via hypothesis testing. Even if all the avail-
able evidence supports a hypothesis, future evidence may disprove it.
He argued falsifiability (the potential for falsification) separates scien-
tific from u
­ nscientific activities such as religion or pseudo-science (e.g.,
psychoanalysis).
According to Popper, scientists should focus on falsification. In fact,
as discussed later, they often seek confirmatory rather than disconfirmatory
evidence when testing their hypotheses. It has also been claimed the same
excessive focus on confirmatory evidence is found in laboratory studies on
hypothesis testing – research to which we now turn.

Wason’s 2-4-6 task

Wason (1960) devised a much-researched hypothesis-testing task (see
Evans, 2016, for a review). Participants were told three numbers 2-4-6 con-
formed to a simple relational rule. Their task was to generate sets of three
numbers and provide reasons for generating each set. After each choice,
the experimenter indicated whether the set of numbers conformed to the
experimenter’s rule. Here is the rule: “Three numbers in ascending order of
magnitude.” The participants could announce what they believed to be the
rule on any trial and were told whether it was correct.
The rule sounds (very) simple. However, only 21% of university stu-
dents were correct with their first attempt (Wason, 1960). In spite of the
emphasis in the literature on the poor levels of performance with the 2-4-6
task, Wason found 72% of participants eventually solved it.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 667 28/02/20 4:14 PM

 668 Thinking and reasoning

KEY TERM Why is performance so poor? One explanation focuses on confirma-

tion bias – most people seek information confirming their hypothesis. For
Confirmation bias
In hypothesis testing, example, participants whose original hypothesis or rule is that the second
seeking evidence that number is twice the first, and the third number is three times the first
supports one’s beliefs. number often generate sets of numbers (test triples) consistent with that
hypothesis (e.g., 6-12-18; 50-100-150).
Wason assumed participants produced test triples conforming to their
current hypothesis. However, this is an oversimplification (Evans, 2016).
It is true that participants mostly produce positive or confirmatory tests
conforming to their hypothesis and expected to receive positive feedback
(conforms to the rule).
However, participants sometimes produce negative tests, not con-
forming to their hypothesis, where they expect to receive negative feed-
back (does not conform to the rule). There are two types of negative tests
(Evans, 2016):

(1) those (e.g., 12-8-4) where participants expect to receive the answer
“No” and which are therefore confirmatory;
(2) those (e.g., 1-4-9) where participants expect to receive the answer
“Yes” and which are therefore disconfirmatory.

Findings
Participants typically engage in very few falsification attempts on the 2-4-6
task and have a low success rate. Tweney et al. (1980) enhanced perfor-
mance by telling participants the experimenter had two rules in mind and
they had to identify both. One rule generated DAX triples and the other
MED triples. They were also told 2-4-6 was a DAX triple. After generating
each test triple, participants were informed whether the set fitted the DAX
or MED rule. The DAX rule was any three numbers in ascending order and
the MED rule covered all other sets of numbers.
Over 50% of participants produced the correct answer on their first
attempt (much higher than with the standard problem). Of importance,
participants could identify the DAX rule by using positive testing to
confirm the MED rule, and so they did not have to try to disconfirm the
DAX rule.
Gale and Ball (2012) carried out a study resembling that of Tweney
et al. (1980). They always used 2-4-6 as an example of a DAX triple,
but the example of a MED triple was 6-4-2 or 4-4-4. Success in identify-
ing the DAX rule was much greater when the MED example was 6-4-2
(75%) rather than 4-4-4 (23%). The greatest difference between solvers and
non-solvers of the DAX rule was the number of descending triples they
produced. This indicates the importance of participants focusing on the
ascending/descending dimension, which was difficult to do when the MED
example was 4-4-4.
Cowley and Byrne (2005) argued people show confirmation bias because
they are loath to abandon their initial hypothesis. They suggested people
might be much better at managing to falsify a given incorrect hypothesis
if told it was someone else’s. As predicted, 62% of participants abandoned
the other person’s hypothesis compared to only 25% who abandoned their

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 668 28/02/20 4:14 PM

 Reasoning and hypothesis testing 669

own hypothesis. Cowley (2015) extended these findings. Participants were

told the hypothesis “even numbers ascending in two” was theirs or that of
an imaginary participant Peter. They generated more negative falsifying
tests when the hypothesis was Peter’s rather than their own (32% vs 7%,
respectively).
In another experiment by Cowley (2015), all participants were told
Peter’s hypothesis was “even numbers ascending in twos”. In two condi-
tions, they were also told another participant, James, had a different explicit
hypothesis (“any ascending numbers”) or implicit hypothesis (“something
else”). In the third, condition, there was no mention of James. There were
three key findings. First, there were numerous falsifying test triples (posi-
tive + negative tests) in all conditions (percentages between 43% and 54%;
predominantly negative falsifying tests). Second, Peter’s incorrect hypoth-
esis was rejected by nearly 80% of participants. Third, the correct rule was
discovered by 50% of participants given James’ explicit hypothesis and
31% given his implicit hypothesis.
In sum, “Negative falsifying is possible more often than the literature
has ever shown . . . falsification is sufficient to announce that a hypothesis
is untrue, but an explicit alternative hypothesis that explains the falsifying
result is necessary for truth discovery” (Cowley, 2015, pp. 32–33).
Performance on the 2-4-6 task involves separable processes of hypoth-
esis generation and hypothesis testing. Most research has focused on the
latter, but Cherubini et al. (2005) focused on hypothesis generation. They
argued participants try to preserve as much of the information contained in
the example triple (i.e., 2-4-6) as possible in their initial hypothesis, making
this hypothesis much more specific than the general rule.
Cherubini et al. (2005) presented participants with two initial triples
exemplifying the rule. When these triples inhibited the generation of a very
specific rule (e.g., 6-8-10; 9-14-15), participants generated more general
hypotheses than when the two triples were consistent with a specific rule
(e.g., 6-8-10; 16-18-20). The success rate was much higher in the former
condition (70% vs 30%).

Theoretical analysis
Most hypotheses are sparse or narrow (applying to under half the possi-
ble entities in any given domain: Navarro & Perfors, 2011). For example,
Perfors and Navarro (2009) asked people to generate all the rules and
hypotheses applying to numbers in a given domain (numbers 1 to 1,000).
The key finding was that 83% of the rules (e.g., two-digit numbers; prime
numbers) applied to fewer than 20% of the numbers.
With sparse hypotheses, positive testing is optimal “because there
are so many ways to be wrong and so few to be right”. In such circum-
stances, the learner will discover “the world has a bias towards saying
‘no’, and asking for ‘yes’ is the best way to overcome it” (Perfors &
Navarro, 2009, p. 2746). Thus, positive testing is typically successful. In
contrast, the 2-4-6 task penalises positive testing because the target rule is
so general.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 669 28/02/20 4:14 PM

 670 Thinking and reasoning

Evaluation
Wason’s 2-4-6 task has been “the classic test-bed reasoning task for investi-
gations of hypothesis falsification for over forty years” (Cowley, 2015, p. 2),
with research having clarified the strengths and limitations of human induc-
tive reasoning. The processes involved in the 2-4-6 task are of relevance to
understanding scientists’ hypothesis testing.
What are the limitations of Wason’s approach? First, his task differs
from real-life hypothesis testing. Participants given the 2-4-6 task receive
immediate accurate feedback but are not told why the numbers they pro-
duced attracted a “yes” or “no” response. In the real world (e.g., scientists
testing hypotheses), the feedback is much more informative, but is often
delayed in time and sometimes inaccurate.
Second, the correct rule or hypothesis in the 2-4-6 task (three numbers
in ascending order of magnitude) is very general because it applies to a
fairly high proportion of sets of three numbers. In contrast, most rules
or hypotheses apply to only a smallish proportion of possible objects or
events. Positive testing works poorly on the 2-4-6 task but not with most
other forms of hypothesis testing.
Third, Wason argued most people show confirmation bias and find a
falsification approach very hard to use. However, there is much less con-
firmation bias but more evidence of falsification when testing someone
else’s hypothesis (Cowley, 2015; Cowley & Byrne, 2005). This is consistent
with scientists’ behaviour. For example, at a conference in 1977 on the
levels-of-processing approach to memory (see Chapter 6), nearly all the
research presented identified limitations with that approach.

Hypothesis testing: simulated and real research

environments
According to Popper (1968), a crucial feature of all truly scientific theories
is falsifiability. Scientists should focus on falsification rather than confir-
mation because the latter cannot be achieved. His arguments possess merit
but are oversimplified (see below). Suppose a scientist obtains findings
apparently inconsistent with their hypothesis. The findings may mean the
hypothesis is incorrect. However, they may reflect problems with the exper-
imental design or the accuracy of the data. Of relevance, a recent study
(Open Science Collaboration, 2015; see Chapter 1) found that attempts to
replicate 100 findings in psychology were successful only 36% of the time.
Dunbar (1993) found evidence of confirmation bias using a simulated
research environment. Participants had to explain how genes are controlled
by other genes using a computer-based molecular genetics laboratory. This
problem is so difficult that those solving it in real life (Jacques Monod
and François Jacob) received the Nobel prize! The participants were led to
focus on the hypothesis that the gene control was by activation whereas it
was actually by inhibition.
Participants who simply sought data consistent with their activation
hypothesis failed to solve the problem. In contrast, the 20% of partici-
pants solving the problem tried to explain the discrepant findings. Most
participants started with the general hypothesis that activation was the key

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 670 28/02/20 4:14 PM

 Reasoning and hypothesis testing 671

IN THE REAL WORLD: HYPOTHESIS TESTING BY SCIENTISTS

What actually happens in real-life science? Uchino et al. (2010) analysed numerous research articles
in psychology. The great majority (77%) sought confirmation by testing the hypothesis favoured
by the researcher(s) and 91% supported an existing theory. Only 22% discussed other hypoth-
eses. These findings suggest, “Someone must be wrong: either scientists are going about their
business incorrectly or Popper was mistaken about how science progresses” (Sanbonmatsu et al.,
2015, p. 2).
Sanbonmatsu et al. (2015) proposed a solution based on a distinction between absolute or uni-
versal hypotheses and non-absolute hypotheses. Absolute hypotheses claim a given phenomenon
always occurs, whereas non-absolute hypotheses claim a phenomenon occurs only in some condi-
tions. Popper (1968) assumed scientific theories are absolute or universal. On that assumption, the
optimal approach involves falsification or disconfirmation – a single negative observation would
disprove an absolute theory. In contrast, a confirmatory approach is generally more informative
than a disconfirmatory one with non-absolute hypotheses.
What do scientists actually do? Sanbonmatsu et al. (2015) found 96% of researchers in psychol-
ogy indicated their research was mostly driven by non-absolute hypotheses. Nearly all (96%) said
they generally used a confirmatory approach. With absolute hypotheses, 81% of researchers would
use a disconfirmatory approach. With non-absolute hypotheses, in contrast, only 9% would use a
disconfirmatory approach with 91% favouring a confirmatory approach.
Feist (2008) argued a useful heuristic in science is “confirm early – disconfirm late”. Scientists
should initially seek confirmatory evidence for a theory; when they have such evidence, they should
focus more on disconfirming the theory and discovering its breadth of application. Eighty-three
percent of scientists were most likely to use a confirmatory approach early in a research pro-
gramme, and 87% were most likely to use a disconfirmatory approach subsequently (Sanbonmatsu
et al., 2015).
In sum, Popper adopted an excessively black-and-white approach. The reality is messy: research
rarely provides a definitive falsification of a theory. Instead, theories are modified as their limita-
tions become increasingly apparent (Lakatos, 1978).
Scientists’ strategies when engaged in hypothesis testing approximate to maximising the infor-
mativeness of the evidence obtained. Thus, describing their typical approach as “confirmation
bias” (see Glossary) is misleading. Kane and Webster (2013) suggested using the term “confirma-
tion heuristic” to refer to appropriate focus on a confirmation strategy, limiting the use of “con-
firmation bias” to situations where scientists refuse to use the disconfirmatory strategy even when
optimal.

controlling process, focusing on one gene after another as the potential

activator. Only after every activation hypothesis had been disconfirmed
did some participants focus on explaining data inconsistent with activation
hypotheses.

Confirmation bias: analysis and interpretation

We have seen scientists generally adopt a confirmatory approach during
hypothesis testing. We now focus on confirmation bias when scientists
analyse and interpret their findings (see Nuzzo, 2015, for a review).
Fugelsang et al. (2004) studied professional scientists working on issues
in molecular biology relating to how genes control and promote replication

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 671 28/02/20 4:14 PM

 672 Thinking and reasoning

in bacteria, parasites and viruses. Of 417 experimental results, over half

(223) were inconsistent with the scientists’ predictions. They responded to
88% of these inconsistent findings by blaming problems with their methods.
In only 12% of cases did the scientists modify their theories. Thus, the sci-
entists showed considerable reluctance to change their original theoretical
position.
Approximately two-thirds of the inconsistent findings were followed
up, generally by changing the methods used. In 55% of cases, the incon-
sistent findings were replicated. The scientists’ reactions were very different
this time – in 61% of cases, they changed their theoretical assumptions.
How defensible was the scientists’ behaviour? Note that almost half the
inconsistent findings were not replicated when a second study was carried
out. Thus, it was reasonable for the scientists to avoid prematurely accept-
ing possibly spurious findings. Overall, these findings suggest the scientists’
exhibited only a modest tendency towards confirmation bias.
John et al. (2012) asked over 2,000 psychologists to provide anony-
mous information about their questionable research practices. There was
substantial evidence of such practices. For example, John et al. estimated
78% of respondents had selectively reported studies that “worked”, 62%
had excluded data (typically those inconsistent with their hypotheses), and
36% had stopped data collection after achieving the desired result.
Here are two other examples of confirmation bias. First, Bakker and
Wicherts (2011) found in an analysis of statistical analyses in research jour-
nals that over 10% of p values were incorrect. In the great majority of
cases where such errors changed the statistical significance of the results,
the change was from non-significant to significant.
Second, researchers often expect their meta-analyses (see Glossary)
to support their existing hypotheses. As Watt and Kennedy (2017, p. 1)
argued, “Decisions about studies to be included [in a meta-analysis], statis-
tical analyses, and moderating factors are made after the analysts know the
outcomes of the studies. These retrospective decisions provide high poten-
tial for [confirmation] bias.”
How can we reduce confirmation bias in the analysis and interpretation
of data? Several answers have been proposed (Hamlin, 2017). First, more
openness or transparency is required by experimenters so other research-
ers can see precisely what has been done. Second, there is blind analysis –
all relevant statistical analyses are completed before the experimenter(s)
is aware of the outcomes of such analyses. Third, and most importantly,
there is pre-registration – experimenters announce the rationale, hypoth-
eses, design and proposed methods of data analysis before conducting a
piece of research.

DEDUCTIVE REASONING
In deductive reasoning, conclusions can be drawn with certainty. In this
section, we will mostly consider conditional and syllogistic reasoning prob-
lems based on traditional logic. In the next section, we consider general
theories of deductive reasoning. As we will see, theory and research increas-
ingly focus on the non-logical strategies and processes used when people
solve deductive-reasoning problems.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 672 28/02/20 4:14 PM

 Reasoning and hypothesis testing 673

Conditional reasoning KEY TERM

Conditional reasoning (basically, reasoning with “if”) had its origins in Conditional reasoning
propositional logic, in which logical operators such as or, and, if . . . then, if A form of deductive
reasoning based on if . . .
and only if are included in sentences or propositions. In this system, symbols then propositions.
represent sentences and logical operators are applied to them to reach con-
clusions. Thus, we might use P to stand for the proposition “It is raining”
and Q to stand for “Nancy gets wet”, and then use the logical operator
if . . . then to relate these two propositions: if P then Q.
The meanings of words and propositions in propositional logic differ
from their natural language meanings. For example, propositions can have
only one of two truth values: true or false. If P stands for “It is raining”,
then P is true (in which case it is raining) or P is false (it is not raining).
Propositional logic does not admit any uncertainty about the truth of P,
such as when it is so misty you could almost call it raining.
Many people produce incorrect answers when given certain condi-
tional-reasoning problems. Consider the following (affirmation of the
consequent):

Premises
If Nancy is angry, then I am upset.
I am upset.
Conclusion
Interactive exercise:
Therefore, Nancy is angry. Conditional reasoning

Many people accept the above conclusion as valid. However, it is not valid
because I may be upset for some other reason (e.g., my football team has
lost).
Here is another problem in conditional reasoning:

Premises
If it is raining, then Nancy gets wet.
It is raining.
Conclusion
Nancy gets wet.

This conclusion is valid. It illustrates the rule of inference known as modus

ponens: “If P, then Q” and also given “P”, we can validly infer Q.
Another major rule of inference is modus tollens: from the premise “If
P, then Q” and the premise “Q is false”, the conclusion “P is false” neces-
sarily follows. Here is an example:

Premises
If it is raining, then Nancy gets wet.
Nancy does not get wet.
Conclusion
It is not raining.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 673 28/02/20 4:14 PM

 674 Thinking and reasoning

People consistently perform much better with modus ponens than modus
tollens: many people argue incorrectly that the conclusion to the above
problem is invalid.
Another inference involves denial of the antecedent:

Premises
If it is raining, then Nancy gets wet.
It is not raining.
Conclusion
Therefore, Nancy does not get wet.

Many people argue the above conclusion is valid although it is invalid.

It does not have to be raining for Nancy to get wet (e.g., she might have
jumped into a swimming pool).
Traditionally, research on conditional reasoning was limited in three
ways. First, unlike everyday life, it focused on disinterested reasoning
(goals and preferences are irrelevant). For example, denial of the anteced-
ent is invalid in traditional logic. In natural language, however, “If P, then
Q” often means “If and only if P, then Q”. If someone says to you “If you
mow the lawn, I will give you five dollars”, you are likely to interpret it to
imply, “If you don’t mow the lawn, I won’t give you five dollars.”
Second, traditional research typically involved instructions indicat-
ing that background knowledge was irrelevant. Nowadays, participants
are generally not instructed to disregard their relevant knowledge with
­conditional-reasoning problems.
Third, traditional research required participants to draw definite con-
clusions (true or false). In contrast, participants nowadays are often asked
to assess the probability of the conclusion being true. This change is desir-
able because we often assign probabilities to conclusions in everyday life
(Singmann et al., 2016).

Theories
Here we briefly discuss theories of conditional reasoning. More general
theories of deductive reasoning are discussed later. Klauer et al. (2010)
proposed a dual-source model of conditional reasoning. There is a
­knowledge-based process influenced by premise content where the subjec-
tive probability of the conclusion depends on individuals’ relevant know­
ledge. There is also a form-based process influenced only by the form of the
premises.
Verschueren et al. (2005) also proposed a dual-process model (other
more general dual-process models are discussed later, pp. 683–690). They
focused on individual differences in conditional reasoning more than
Klauer et al. (2010). Some reasoners use a relatively complex counter­
example strategy: a conclusion is considered invalid if the reasoner can
find a counterexample to it (this process is discussed later in the section
on mental models, pp. 681–683). The other process is an intuitive
statistical strategy based on probabilistic reasoning triggered by relevant
knowledge.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 674 28/02/20 4:14 PM

 Reasoning and hypothesis testing 675

Findings
Singmann et al. (2016) tested the dual-source model using many condition-
al-reasoning problems. For each problem, participants rated the likelihood
of the conclusion being true on a probability scale running from 0% to
100%. Knowledge-based processes were more important than form-based
ones. Overall, the model accounted very well for participants’ performance.
De Neys et al. (2005) obtained evidence relevant to Verschueren
et al.’s (2005) dual-process model. On some trials, participants were pre-
sented with few or many counterexamples conflicting with valid conclu-
sions (modus ponens and modus tollens). According to classical logic, these
counterexamples should have been ignored. In fact, however, participants
were more likely to decide wrongly the conclusions were invalid when there
were many counterexamples.
Markovits et al. (2013) tested the dual-process model using problems
involving affirmation of the consequent (where the conclusion is invalid).
Here are two examples:

(1) If a rock is thrown at a window, then the window will break.

A window is broken. Therefore, a rock was thrown at the window.
(2) If a finger is cut, then it will bleed. A finger is bleeding. Therefore, the
finger was cut.

Reasoners using the statistical strategy were influenced by the fact that
the subjective probability that “If a finger is bleeding, it was cut” is greater
than the probability that “If a window is broken, it was broken by a rock”.
As a result, such reasoners accepted the invalid conclusion more often in
problem (2) than (1).
In contrast, reasoners using the counterexample strategy accepted the
conclusion if no counterexample came to mind. They also accepted the
invalid conclusion more often in problem (2). This was because it was
easier to find counterexamples with respect to the conclusion of problem
(1) (the window might have been broken by several objects other than a
rock) than the conclusion of problem (2).
According to the model, the counterexample strategy is more cogni-
tively demanding than the statistical strategy. Accordingly, Markovits
et al. (2013) predicted it would be used less often when participants had
limited time. This prediction was supported: that strategy was used on 49%
of trials with unlimited time but only 1.7% of trials with limited time.
Markovits et al. (2017) gave their participants modus ponens inferences
which are always valid. Each problem was presented with additional infor-
mation indicating the relative strength of evidence supporting the inference
(50%; 75%; 99%; or 100%). Markovits et al. compared groups of partici-
pants previously identified as using counterexample or statistical strategies.
Markovits et al. (2017) found clear differences between the two groups
(see Figure 14.1). Statistical reasoners were strongly influenced by relative
strength when deciding whether to accept modus ponens inferences. In con-
trast, counterexample reasoners showed a sharp reduction in acceptance
when some evidence failed to support the inference (e.g., the 99% and 75%
conditions). These findings were as predicted.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 675 28/02/20 4:14 PM

 676 Thinking and reasoning

Figure 14.1
Mean number of MP 3

Mean number of MP inferences accepted

(modus ponens) inferences
accepted (out of 3) as a 2.5
function of relative strength
of the evidence and 2
strategy.
From Markovits et al. (2017). 1.5
Counterexample

1 Statistical

0.5

0
100% 99% 75% 50%
Relative strength

Summary
Research on conditional reasoning has become much more realistic. For
example, reasoners are encouraged to use their relevant knowledge on rea-
soning tasks. They also assign probabilities to the correctness of conclu-
sions rather than simply deciding conclusions are valid or invalid.
Theoretically, it is assumed reasoners are influenced by the form of the
premises. More importantly, however, their relevant knowledge and expe-
rience lead them to engage in probabilistic reasoning (dual-source model)
or to try to find counterexamples to the stated conclusion (dual-process
model).

Wason selection task

The Wason selection task has been studied intensively by researchers inter-
ested in deductive reasoning. However, it is more accurately described as a
Research activity: task involving hypothesis testing using a conditional rule.
Deductive reasoning In the standard version of the task, four cards lie on a table (R, G, 2, 7;
see Figure 14.2). Each card has a letter on one side and a number on the
other, and there is a rule applying to the four
cards (e.g., “If there is an R on one side of
the card, then there is a 2 on the other side of
the card”). The participants’ task is to select
only those cards needing to be turned over
to decide whether or not the rule is correct.
What is your solution? Most people select the
R and 2 cards. If you did the same, you are
wrong!
You need to see whether any cards fail
to obey the rule when turned over. From this
Figure 14.2 perspective, the 2 card is irrelevant: if there
The Wason selection task. Rule: If there is an R on one side of is an R on the other side, this indicates only
the card, then there is a 2 on the other. that the rule might be correct. If there is any

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 676 28/02/20 4:14 PM

 Reasoning and hypothesis testing 677

other letter on the other side, we have discovered nothing about the rule’s KEY TERMS
validity. The correct answer is to select the R and 7 cards, an answer given
Matching bias
by only about 10% of university students. The 7 is necessary because it The tendency on the
would definitely disprove the rule if it had an R on the other side. Wason selection task to
How can we explain performance on the Wason selection task? select cards matching the
Several factors are involved. First, performance is worse with abstract items explicitly mentioned
in the rule.
versions of the task (as above) compared to concrete versions referring
to everyday events (e.g., “Every time I travel to Manchester, I travel by Deontic rules
train”). Ragni et al. (2018) found in a meta-analysis that the percentage Rules relating to
obligation and
of correct answers increased from 7% with abstract versions to 21% with
permissibility.
concrete versions.
Second, there is matching bias (the tendency to select cards matching
the items named in the rule). Thompson et al. (2013a) obtained strong evi-
dence for matching bias on the selection task. In addition, cards named in
the rule were selected faster than other cards and produced greater feelings
of rightness.
Third, the logical solution to the Wason selection task conflicts with
everyday life (Oaksford, 1997). According to formal logic, we should test
the rule “All swans are white” by searching for swans and non-white birds.
However, this would be extremely time-consuming because only a few
birds are swans and the overwhelming majority of birds are non-white. It
would be preferable to adopt a probabilistic approach based on the likely
probabilities of different kinds of events or objects.
The problem of testing the above rule resembles the Wason selection
task, which has the form “If p, then q”. We should choose q cards (e.g., 2)
when the expected probability of q is low but not-q cards when q’s expected
probability is high to maximise information gain. As predicted, far more
q cards were selected when the percentage of q cards was low (17%) than
when it was high (83%) (Oaksford et al., 1997).
Fourth, motivation is important. People are more likely to select the
potentially falsifying card (7 in the original version of the task) if moti-
vated to disprove the rule. Dawson et al. (2002) gave some participants the
rule that individuals high in emotionality lability experience an early death.
The four cards showed high emotional lability, low emotional lability,
early death and late death, with the correct answer involving selecting the
first and last cards. Of participants led to believe they had high emotional
lability (and so motivated to disprove the rule), 38% solved the problem
(versus 9% of control participants).
Motivation is also involved with deontic rules (rules concerned with
obligation or permission). Sperber and Girotto (2002) used a deontic rule
relating to cheating. Paolo must decide whether he is being cheated when
buying things through the internet: the answer is to select the “item paid
for” and “item not received” cards (selected by 68% of participants). This
unusually high level of performance was achieved because the motiva-
tion to detect cheating led participants to select the “item not received”
card. In a meta-analysis, Ragni et al. (2018) found the correct answer was
selected by 61% of participants with deontic versions but 7% using abstract
versions.
Marrero et al. (2016) proposed a general motivational approach
accounting for the above findings. Individuals concerned about potential

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 677 28/02/20 4:14 PM

 678 Thinking and reasoning

KEY TERMS costs focus on disconfirming evidence whereas those concerned about
potential benefits focus on confirming evidence.
Syllogism
A type of problem used Fifth, Ragni et al. (2018) evaluated 15 theories of relevance to Wason’s
in deductive reasoning; selection task. In a large-scale meta-analysis, they found Johnson-Laird’s
there are two statements (1983) mental model theory (discussed shortly, pp. 681–683) best predicted
or premises and a performance. In essence, this theory assumes that selections on Wason’s
conclusion that may or
selection task depend on two processes:
may not follow logically
from the premises.
(1) There is an intuitive process producing selections matching the rea-
Belief bias
soners’ hypothesis (e.g., selection of R in the version of the task
In syllogistic reasoning,
the tendency to accept shown in Figure 14.2).
invalid but believable (2) There is a more deliberate process producing selections of potential
conclusions and reject counterexamples to the hypothesis (e.g., selection of 7 in the same
valid but unbelievable version).
ones.

The extent to which reasoners search for counterexamples depends on

factors discussed above including task content, instructions and so on.

Syllogistic reasoning
Syllogistic reasoning has been studied for over 2,000 years. A syllogism
consists of two premises or statements followed by a conclusion. Here is
an example: “All A are B; all B are C. Therefore, all A are C”. A syllogism
contains three items (A, B and C), with one (B) occurring in both premises.
The premises and conclusion all contain one of the following quantifiers:
all; some; no; and some . . . not.
When presented with a syllogism, you must decide whether the conclu-
sion is valid assuming the premises are valid. The validity (or otherwise)
of the conclusion depends only on whether it follows logically from the
premises – the conclusion’s truth or falsity in the real world is irrelevant.
Consider the following example:

Premises
All children are obedient.
All girl guides are children.
Conclusion
Therefore, all girl guides are obedient.

The conclusion follows logically from the premises. Thus, it is valid regard-
less of your views about children’s obedience.

Findings
Various biases cause errors in syllogistic reasoning. Of special impor-
tance is belief bias, the tendency to accept invalid conclusions as valid
if ­believable and to reject valid (but unbelievable) conclusions as invalid
(theoretical explanations are discussed later). Klauer et al. (2000) investi-
gated belief bias thoroughly. The conclusions of half their syllogisms were
­believable (e.g., “Some fish are not trout”) whereas the others were unbe-
lievable (e.g., “Some trout are not fish”). Half the syllogisms were valid

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 678 28/02/20 4:14 PM

 Reasoning and hypothesis testing 679

and half invalid; however, some participants

were told only one-sixth of the syllogisms
were valid whereas others were told five-
sixths were.
What did Klauer et al. (2000) find? First,
there was a base-rate effect: syllogistic rea-
soning performance was influenced by the
perceived probability of syllogisms being
valid (see Figure 14.3). Second, there was
strong evidence for belief bias. Third, there
was a belief-by-logic interaction. Performance
on syllogisms with valid conclusions was
better when those conclusions were believ­
able, whereas performance on syllogisms with
invalid conclusions was worse when those
conclusions were believable. In sum, reason-
ers’ performance was influenced by factors
irrelevant to logic.
Stupple and Ball (2008) found with syl-
logistic reasoning that unbelievable premises
were processed more slowly than believable
ones. This finding suggests people experi-
enced conflict between their beliefs and what
they were asked to assume and resolving this
conflict was time-consuming.
Some problems in syllogistic reasoning
occur because of differences in the meanings
of expressions in formal logic and everyday
life. For example, we often assume “All As
are Bs” means “All Bs are As” and “Some
As are not Bs” means “Some Bs are not As”.
Ceraso and Provitera (1971) spelled out such
premises unambiguously (e.g., “All As are
Bs, but some Bs are not As”). This greatly
enhanced reasoning performance. In similar
fashion, “some” means “some but not all” in
everyday usage but “at least one and possibly Figure 14.3
Percentage acceptance of conclusions as a function of
all” in formal logic. Schmidt and Thompson perceived base rate validity (low vs. high), believability of
(2008) found syllogistic reasoning improved conclusions and validity of conclusions.
when the meaning of “some” in formal logic Based on data in Klauer et al. (2000). © American Psychological
was made explicit. Association.
Finally, people’s syllogistic reasoning per-
formance is influenced by whether the con-
clusion matches the premises in surface or superficial features. Here is an
example of matching: no A are not B; no B are not C; therefore, no C are
not A; and an example of non-matching: all A are B; all B are C; therefore
no A are not C. In spite of the irrelevance of matching vs non-matching
to formal logic, people are more likely to accept conclusions matching the
premises (Stupple et al., 2013).

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 679 28/02/20 4:14 PM

 680 Thinking and reasoning

IN THE REAL WORLD: INVALID DEDUCTIVE REASONING

So far we have focused on laboratory-based examples of invalid deductive reasoning. However,
such reasoning is also very common in everyday life. Here are a few examples starting with the
politician’s syllogism, which is along the following lines:

Premises
We must do something to save the country.
Our policy is something.
Conclusion
Our policy will save the country.

Invalid conditional reasoning occurs frequently in everyday life. Earlier we discussed the logical
fallacy known as denial of the antecedent. Here is a real-world example:

Premises
If you have got nothing to hide, you have nothing to fear.
You have nothing to hide.
Conclusion
You have nothing to fear.

The above argument is invalid because it implies that people are only interested in privacy
because they have something to hide. In fact, of course, we all have a basic human right to
privacy. Ironically, the authorities who argue strongly in favour of greater surveillance of the public
are often notoriously reluctant to provide information about their own activities!
Finally, we consider an everyday example of the logical fallacy known as affirmation of the con-
sequent (discussed earlier, p. 673).

Premises
If the Earth’s climate altered throughout pre-human history, this was due to natural climate

change.
The Earth’s climate is currently altering.
Conclusion
Natural climate change is occurring currently.

This is clearly an invalid argument. The fact that past climate change was not due to humans
does not necessarily mean that current climate change is not due to human intervention.
In sum, many groups in society (e.g. politicians; climate change deniers) are strongly motivated
to persuade us of the rightness of their beliefs. This often leads them to engage in invalid forms of
reasoning. The take-home message is that we all need to be sceptical and vigilant when exposed
to their arguments.

THEORIES OF “DEDUCTIVE” REASONING

Here we will focus on two very influential theoretical approaches to
deductive reasoning. First, there is Johnson-Laird’s mental model theory,
which represents a relatively “traditional” approach. Second, we turn our

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 680 28/02/20 4:14 PM

 Reasoning and hypothesis testing 681

attention to the increasingly popular dual-process approach. The word KEY TERMS
deductive in the title of this section has been put in quotation marks to indi-
Mental models
cate that individuals presented with deductive-reasoning problems often fail An internal representation
to use deductive processes when trying to solve them. of some possible situation
or event in the world
having the same structure
Mental models as that situation or event.

Johnson-Laird (e.g., 1983; Johnson-Laird et al., 2018) argues that reasoning Principle of truth
involves constructing mental models. What is a mental model? According The notion that assertions
are represented by
to Johnson-Laird et al. (2015, p. 202), a mental model is “an iconic rep-
forming mental models
resentation of a possibility that depicts only those clauses in a compound concerning what is true
assertion that are true. The mental models of a disjunction, ‘A or B but while ignoring what is
not both’ accordingly represent two possibilities: possibly (A) and possibly false.
(B)”. It is iconic because its structure corresponds to what it represents.
Here is a concrete example of a mental model:

Premises
The lamp is on the right of the pad.
The book is on the left of the pad.
The clock is in front of the book.
The vase is in front of the lamp.

Conclusion
The clock is to the left of the vase.

According to Johnson-Laird (1983), people use the information contained

in the premises to construct a mental model like this:

book pad lamp

clock vase

The conclusion the clock is to the left of the vase clearly follows from the
mental model. The fact we cannot construct a mental model consistent
with the premises (but inconsistent with the conclusions) (i.e., we cannot
construct a counterexample) indicates the model is valid.
Here are the theory’s main assumptions:

●● A mental model describing the given situation is constructed and the

conclusions that follow are generated.
●● An attempt is made to construct alternative models to falsify the conclu-
sion by finding counterexamples to the conclusion. If a c­ ounterexample
model is not found, the conclusion is deemed valid.
●● The construction of mental models involves the limited resources of
working memory (see Chapter 6).
●● Reasoning problems requiring the construction of several mental
models are harder than those requiring only one mental model because
the former impose greater demands on working memory.
●● The principle of truth: “Mental models represent what is true, but not
what is false” (Khemlani & Johnson-Laird, 2017, p. 16). This mini-
mises demands on working memory.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 681 28/02/20 4:14 PM

 688 Thinking and reasoning

KEY TERM (1) The more intelligent reasoners had greater difficulty resolving con-
flict when providing belief-based responses rather than logic-based
Meta-reasoning
Monitoring processes responses.
that influence the time, (2) The less intelligent reasoners exhibited the opposite pattern.
effort and strategies used
during reasoning and These findings suggest more intelligent individuals generate logic-based
problem solving.
responses faster than belief-based ones, whereas less intelligent individuals
generate belief-based responses faster.
What conclusions should we draw? First, rather than arguing belief-
based responses involve fast Type 1 processing whereas logic-based
responses involve slow Type 2 processing, we need to consider individ-
ual differences. Second, “If responses (belief- or logic-based) can be gen-
erated either quickly and effortlessly or slowly and deliberately, perhaps
these responses merely differ on a single dimension, namely, complexity”
(Newman et al., 2017, p. 1165).

What causes Type 2 processing?

What determines whether reasoners’ responses are based on analytic (Type
2) processing or whether they reflect only intuitive (Type 1) processes?
The theories discussed above address this issue. Traditional serial models
assume that Type 2 processes monitor the output of Type 1 processes and it
is this monitoring process that determines whether reasoning performance
is based on Type 2 processes. In similar fashion, many parallel models
assume Type 2 reasoning is triggered when conflict monitoring (involving
Type 2 processing) leads to conflict detection. What is puzzling about all
these theories is that, “They assume that Type 2 processing is effectively
caused by itself” (Pennycook et al., 2015, p. 36).
Ackerman and Thompson (2017, p. 607) provided a detailed account
of meta-reasoning: “The processes that monitor the progress of our rea-
soning and problem-solving activities and regulate the time and effort
devoted to them” (see Figure 14.6). Monitoring processes assess the proba-
bility of success before, during and after performing a reasoning task. The
most important monitoring feature is the feeling of rightness: “the degree
to which the first solution that comes to mind feels right” (p. 608). Only
when this feeling is weak do reasoners engage in substantial Type 2 or
analytic processing.
Thompson et al. (2011) studied the role of feeling-of-rightness ratings
on the use of Type 2 processes with syllogistic and conditional-reasoning
tasks. Participants provided an initial answer immediately after reading
each problem (intuitive, or Type 1, answer) followed by an assessment
of that answer’s correctness (feeling of rightness). Participants then had
unlimited time to reconsider their initial answer and provide a final ana-
lytic, or Type 2, answer. As predicted, participants spent longer reconsider-
ing their intuitive answer and were more likely to change it when they had
low feelings of rightness.
What determines feeling-of-rightness ratings? Thompson et al. (2013b)
addressed this issue in a study on syllogistic reasoning. Participants produced
the first response that came to mind, provided a feeling-of-­rightness rating,
and then produced a slower, more deliberate response. Feeling-of-rightness

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 688 28/02/20 4:14 PM

 Reasoning and hypothesis testing 689

Figure 14.6
Reasoning Meta-reasoning The approximate time
courses of reasoning and
Meta-cognitive Meta-cognitive meta-reasoning processes
monitoring control during reasoning and
Time-
Identifiy components Assessment of Think? Search problem solving.
line
and goal knowledge and memory? Change From Ackerman & Thompson
strategies strategy? Stop? (2017).

Initial judgement Engage in solving?

of solvability Give up?
Generate an initial,
autonomous, response
Feeling of Provide the initial
rightness response?
Reconsider?
Analytic processing
Intermediate Provide current
confidence response? Try
another strategy?

Answer choice Provide chosen

answer?
Final Feeling Final judgement
“Don’t know”?
confidence of error of solvability
Seek help?

ratings were higher when the first response was produced rapidly rather
than slowly, indicating the importance of response fluency.
Most research indicates response fluency is a fallible measure of response
accuracy. For example, people give higher feeling-of-rightness ratings to
reasoning problems having familiar rather than unfamiliar content even
when the former problems are harder (Ackerman & Thompson, 2017).

Evaluation
What are the strengths of the contemporary dual-process approach? First,
dual-process theories have become increasingly popular and wide-ranging.
For example, they provide explanations for syllogistic reasoning and con-
ditional reasoning (Verschueren et al, 2005, discussed earlier, pp. 674–675).
In addition, such theories account for findings in problem solving (see
Chapter 12), judgement and decision-making (see Chapter 13). Second,
“Dual-process theory . . . provides a valuable high-level framework within
which more specific and testable models can be developed” (Evans, 2018,
p. 163).
Third, there have been increasingly sophisticated attempts to clarify
the relationship between Type 1 and Type 2 processes (e.g., whether they
are used serially or in parallel). Fourth, we have an enhanced understand-
ing of meta-reasoning processes (especially those involved in monitoring).
Fifth, recent theory and research are starting to take account of the flexi-
bility of processing on reasoning problems due to the precise form of the
problem and individual differences (Thompson et al., 2018).
What are the limitations with the dual-process approach? First, the pro-
cesses used by reasoners vary depending on their abilities and preferences,

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 689 28/02/20 4:14 PM

 690 Thinking and reasoning

their motivation and their task requirements. Melnikoff and Bargh (2018;
see Chapter 13) identified two ways many dual-process theories are over-
simplified: (1) they often imply that Type 1 processes are “bad” and error-
prone, whereas Type 2 processes are “good”; and (2) they assume many
cognitive processes can be assigned to just two types.
Second, “The absence of a clear and general definition of a Type 1
or Type 2 response does create difficulty for researchers wishing to test
[dual-process] theories” (Evans, 2018, p. 163). For example, it is often
assumed theoretically that fast responses reflect Type 1 processing and are
error-prone whereas slow responses reflect Type 2 processing and are gen-
erally accurate. However, we have discussed various studies disconfirming
those assumptions.
Third, there has been a rapid increase in the findings that require to be
explained theoretically, and theories have not kept pace with this increase.
For example, meta-reasoning often plays an important role in influenc-
ing reasoners’ processing strategies and performance. As yet, however, no
theorists have integrated meta-reasoning processes into a comprehensive
dual-process theory of reasoning.

BRAIN SYSTEMS IN REASONING

In recent years, there has been increased research designed to identify
the brain regions associated with deductive reasoning. Prado et al. (2011)
reported a meta-analytic review of 28 neuroimaging studies on deductive
reasoning (see Figure 14.7). They obtained evidence for a core brain system
Interactive feature:
centred in the left hemisphere involving frontal and parietal areas. Specific
Primal Pictures’
3D atlas of the brain brain areas activated during deductive reasoning included the inferior
frontal gyrus, the medial frontal gyrus, the precentral gyrus and the basal
ganglia.
Coetzee and Monti (2018) reported findings broadly consistent with
those of Prado et al. (2011). They used fMRI to assess brain activation
while participants performed deductive-reasoning tasks. There were two
key findings. First, core regions (left rostrolateral cortex, in BA10; medial
prefrontal cortex, in BA8) were more strongly activated with complex
(rather than) simple deductive reasoning. Second, the main language areas
Figure 14.7
Brain regions most
consistently activated
across 28 studies of
deductive reasoning.
PG = precentral gyrus;
MFG = middle frontal
gyrus; PPC = posterior
parietal cortex;
IFG = inferior frontal gyrus;
BG = basal ganglia;
MeFG = medial frontal
gyrus.
From Prado et al. (2011).
© Massachusetts Institute of
Technology, by permission of
the MIT Press.

9781138482210_COGNITIVE_PSYCHOLOGY_PART_4.indd 690 28/02/20 4:14 PM