Reasoning About Necessity and Possibility: A Test of The Mental Model Theory of Deduction

Journal of Experimental Psychology:
Learning, Memory, and Cognition

1999, Vol. 25, No. 6. 1495-1513
Copyright 1999 by the American Psychological Association, Inc.

0278-7393/99/S3.00
Reasoning About Necessity and Possibility: A Test of the Mental Model

Theory of Deduction
Jonathan St. B. T. Evans, Simon J. Handley, and
Catherine N. J. Harper
Philip N. Johnson-Laird
Princeton University
University of Plymouth
This article examined syllogistic reasoning that differs from previous research in 2 significant
ways: (a) Participants were asked to decide whether conclusions were possible as well as
necessary, and (b) every possible combination of syllogistic premises and conclusions was
presented for evaluation with both single-premise (Experiment 1) and double-premise
(Experiment 2) problems. Participants more frequently endorsed conclusions as possible than
as necessary, and differences in response to the 2 forms of instruction conformed to several
predictions derived from the mental model theory of deduction. Findings of Experiments 2 and
3 showed that some fallacies are consistently endorsed and others consistently resisted when
people are asked to judge whether conclusions that are only possible follow necessarily. This
finding was accounted for by the computational implementation of the model theory: Fallacies
are made when the first mental model of the premises considered supports the conclusion
presented.
Traditional applications of logic concern the validity of

arguments, that is, proving that some conclusion is necessary given some premises. However, in everyday reasoning
it may be just as important to decide whether some
proposition is possible in light of the given information.
Inferences of possibility occur whenever rules and regulations constrain a person's behavior rather than determine
what it must be. For example, students choosing a degree
program within a modular course structure will have many
degrees of freedom but will have to respect constraints
owing to timetabling restrictions, availability of teaching
staff, prerequisite and corequisite relations between modules, and so on. Their decision making here involves
inferring what is possible and then deciding between the
possibilities identified. Most of the rules by which people
live their lives in society, including criminal law, operate in a
similar way. It is rarely determined by such rules that people
must follow a particular course of action, but they are frequently constrained to act within a set of legal possibilities.
The psychological study of deductive reasoning has been
steadily increasing in recent years, and there are now many
hundreds of experimental studies reported in the literature
(see Evans, Newstead, & Byrne, 1993, for a review). In view
of the above remarks, however, there is a curious limitation
Jonathan St. B. T. Evans, Simon J. Handley, and Catherine N. J.

Harper, Department of Psychology, University of Plymouth, Plymouth, United Kingdom; Philip N. Johnson-Laird, Department of
Psychology, Princeton University.
This study was supported by a research grant from the Economic
and Social Research Council of the United Kingdom (R000221742).
Correspondence concerning this article should be addressed to
Jonathan St. B. T. Evans, Center for Thinking and Language,
Department of Psychology, University of Plymouth, Plymouth PL4
8AA United Kingdom. Electronic mail may be sent to
J.Evans@plym.ac.uk.
in almost all studies reported to date: They ask participants

only to decide if a conclusion is necessary given some
premises. There have been almost no studies prior to that
reported in the present article that ask participants to decide
what is or is not possible given some assumptions.
In logic, a deductive inference is valid if its conclusion
must be true given that its premises are true. Conclusions in
standard logic state what is the case, but in "modal" logic
they can in addition state what is possibly the case, or what is
necessarily the case. Most psychological experimental investigations of reasoning have considered only conclusions
about what is the case. Typically, the participants are asked
to evaluate the validity of a given conclusion (the evaluation
task) or else to generate their own valid conclusion (the
production task). Most of the studies have been based on
problems drawn from the propositional calculus, which
concerns the logic of negation and such connectives as if, or,
and and, or on classical syllogisms of the sort devised by
Aristotle. There have been few studies of modal reasoning
(but see Bell & Johnson-Laird, 1998; Galotti, Baron, &
Sabini, 1986; Osherson, 1976).
In the research reported here, we examine inferences
about what is possible as well as what is necessary within the
syllogistic reasoning paradigm. The research is motivated by
the mental model theory of deductive reasoning from which
we are able to derive some general predictions about
necessity and possibility in deductive reasoning. We start
with a brief account of that theory.
Mental Model Theory of Deduction
The reason that so much research on standard deduction
has been carried out is because many cognitive scientists
believe that the ability to reason deductively is central to
human intelligence. In fact, people make many logical errors
on these tasks and exhibit systematic biases (Evans, 1989), a
1495
1496
EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD
finding which has led to much debate about human rationality (see Evans & Over, 1996; Manktelow & Over, 1993).
However, most psychologists in the area would agree that
untrained reasoners do show some basic deductive competence: For example, they can distinguish valid inferences
from fallacies at well above chance rates. Accounting for
such competence has, however, led to a protracted, sometimes heated, and as-yet unresolved debate between two
theoretical camps. The debate is focused on whether deductive reasoning is achieved by the use of mental rules or
mental models (see Evans et al., 1993, chapter 3, for detailed
coverage of the debate).
Rule-based theories derive from so-called natural deduction, a method of formalizing logic that is intuitively
plausible and that postulates rules of inference for each
logical term. The most durable of these have proved to be the
systems devised by Rips (1983, 1994) known as PSYCOP,
and the system developed by Braine and O'Brien (Braine,
1978; Braine & O'Brien, 1991; Braine, Reiser, & Rumain,
1984). However, we provide little discussion of these
theories in the present article for two reasons. First, rule
theorists have largely confined their theoretical and experimental work to the study of propositional reasoning and
have had rather little to say about syllogistic inference (but
see some discussion of reasoning with quantifiers by Rips,
1994). Second, the rale theories as published provide only
rules for valid inferences and hence can account for necessary but not possible inferences. We defer consideration of
how rule theories might be extended to the explanation of
possible inference until the General Discussion.
The account of human deduction in terms of mental
model theory was devised by Johnson-Laird (1983; JohnsonLaird & Byrne, 1991). In this approach, people do not prove
conclusions syntactically by applying valid inference rules.
Rather, their deductions are based on grasping a semantic
principle, namely, that a conclusion is valid if there is no
model of the premises that exclude it. The model theory
includes psychological constraints that do not apply to
formal semantic methods such as truth-table analysis. In
particular, it is supposed that people focus on situations in
which the premises would be true and represent such
situations selectively.
The mental model theory was originally devised as an
account of syllogistic reasoning (Johnson-Laird & Bara,
1984) and in its general form supposes that deduction
requires three stages:
1. Model formation. The reasoner forms an initial model
from the premises.
2. Conclusion formation. The reasoner derives a putative
conclusion from the model that is informative (e.g., not a
repetition of a premise).
3. Conclusion validation. The reasoner searches for a counterexample, that is, a model in which the premises are true but
the conclusion is false. If no such model is found, the
conclusion is valid.
A large experimental program of research has been
carried out by mental model theorists who claim support for
the predictions of their theory (see Johnson-Laird & Byrne,
1991). For example, Johnson-Laird and Bara (1984) ran a
large-scale study of syllogistic reasoning using production

task methodology. All possible premise pairs of the classical
syllogistic form (see below) were presented, and participants
were asked to derive a conclusion. The authors analyzed
these premise pairs according to whether they were consistent with one, two, or three models. As predicted, error rates
were higher on multimodel problems. The psychological
principle here is that consideration of multiple models places
a strain on working-memory capacity. In particular, people
often commit fallacies because they fail to find a counterexample model at Stage 3, even though such a model exists.
Many specific findings claimed in support of the model
theory have, however, been disputed by rule theorists (see
Evans et al., 1993). Evans and Over (1996,1997) argued that
it is very difficult to decide between the two approaches in
most of these arguments as neither theory is fully specified.
For this reason, we present here an attempt to test some
predictions that follow from only the most general principles
of the model theory and that do not rely on any supplementary assumptions to account for the specific tasks. Nor do
they rest on any assumptions about how particular connectives or quantifiers are represented. These predictions concern what people will do when asked to make judgments of
necessity and possibility and are derived below.
Possible and Necessary Inference

In modal logic, a distinction is drawn among truths,
necessary truths, and possible truths. A statement is true if it
is a matter of fact; a statement is necessary if it must be true;
and a statement is possible if it may be true. Consider the
following three arguments based on universal premises:
All artists are beekeepers,
Lisa is an artist,
Therefore, it necessary that Lisa is a beekeeper.
Lisa is a beekeeper,
Therefore, it is possible Lisa is an artist.
Lisa is an artist,
Therefore, it is possible that Lisa is not a beekeeper.
(1)
(2)
(3)
Argument 1 is a modus ponens inference in which the

modal conclusion states a necessity. Given that the premises
are true, then indeed it is necessary that Lisa is a beekeeper,
and so the conclusion is valid. Argument 2 is also valid, but
any stronger conclusion, such as "It is necessary that Lisa is
an artist," would be invalid, because the premises are
consistent with a situation in which there are beekeepers
who are not artists, and so Lisa may be a beekeeper who is
not an artist. Argument 3 is invalid, because as Argument 1
shows, Lisa is necessarily a beekeeper given the premises.
In this article, we refer to these three sorts of argument as
Necessary, Possible, and Impossible. Standard deductive
reasoning instructions ask people to make judgments about
the validity of factual conclusions. If people are asked to
decide whether it is necessary that X, where X is an
assertion, then they should endorse Necessary conclusions
and reject both those that are Possible and Impossible. If
1497
NECESSITY AND POSSIBILITY
people are asked instead to decide whether it is possible that

X, then they should endorse Necessary conclusions and
Possible conclusions and reject only Impossible conclusions. The ability to discriminate Possible from Impossible
conclusions is neglected in normal psychological research
on reasoning.
Consider the three general stages of the model theory
defined earlier. In the first two stages, the reasoner constructs
a single model of the premises that supports a provisional
conclusion. This conclusion can be reported as possible
without any further reasoning. The conclusion validation
stage (search for counterexamples) only comes into play
when people are asked to prove that the putative conclusion
is necessary. Hence, according to the model theory, possible
inference should be easier than necessary inference. Our first
prediction (P) therefore is as follows:
PL People will be more willing to endorse conclusions as
possible than as necessary.
The model theory provides some further clear predictions
about the relative difficulty of different types of problems in
this paradigm. Consider the three basic types of problems
from a model theory perspective:
Necessary problems: the conclusion is true in all models of the
premises.
Possible problems: the conclusion is true in at least one model
of the premises.
Impossible problems: the conclusion is true in no model of the
premises.
inference task (see Newstead & Griggs, 1983). Experiment 2

involves full syllogistic reasoning and to our knowledge is
the first to ask people to evaluate every possible conclusion
with regard to every possible pair of premises even for
validity (Necessity) judgments. Both experiments included
groups required to make judgments of Possibility as well as
of Necessity of conclusions.
Experiments 1 and 2 were designed together and run on
the same participants in the same sessions. Owing to the
large difference in scale between the two experiments,
however, counterbalancing of the order of the experiments
seemed inappropriate. It was also undesirable to ask the
same participant to make judgments of both necessity and
possibility. The procedure adopted was as follows. Participants were allocated into Necessity and Possibility groups
and run through the short task of Experiment 1 using the
appropriate form of instruction. Each participant then performed the large task required by Experiment 2 using the
same form of instruction allocated to them in Experiment 1.
The groups were also subdivided in Experiment 2 according
to the order of terms in the conclusions they evaluated (see
Method section of Experiment 2 for details). We also report a
third experiment that replicates some of the key findings of
Experiment 2 using a much smaller subset of syllogisms and
simpler method of problem presentation. Experiment 3 also
provides a check that the results of Experiment 2 were
not influenced by the prior experience of taking part in
Experiment 1.
From this analysis we derive and test in this article two

specific predictions:
P2. It will be easier to decide that a conclusion is possible if it
is also necessary. Specifically, we predict more endorsements
of possibility for Necessary than for Possible problems.
P3. It will be easier to decide that a conclusion is not
necessary if it is also not possible. Specifically, we predict that
more Possible than Impossible problems will be endorsed as
necessary.
P2 follows because it is only necessary to discover one
model supporting the conclusion in order to infer that it is
possible. On Necessary problems, all models support the
conclusion, whereas on Possible problems there is at least
one model that does not. Hence, reasoners are at risk of
thinking first of a model that does not support the conclusion
in the latter case. P3 follows because a conclusion is not
necessary when there is at least one model that does not
support it. On Impossible problems, there is no model of the
premises that supports the conclusion, whereas on Possible
problems there is at least one that does. Hence on the latter
problems reasoners are at risk of first thinking of a model
that supports the conclusion.
We report two experiments designed to test our a priori
hypotheses and a third that addresses issues arising from the
findings of these studies. In Experiment 1, we tested the
above predictions using a simpler task than full syllogistic
reasoning. The task involves giving people a single premise
of the kind used in syllogisms from which they must draw an
inference; this is usually referred to as the immediate
Experiment 1: Immediate Inference

Syllogisms are formed by combining two premises and a
conclusion, each of which can have one of only four
quantified forms known traditionally by the abbreviations A,
E, I, and O and described as "moods." These are as follows:
A
E
I
O
Universal affirmative
Universal negative
Particular affirmative
Particular negative
All X are Y
No X are Y
Some X are Y
Some X are not Y
In Experiment 1, we looked only at the immediate

inferences that people may draw when one of these statements acts as a premise and the other as a conclusion.
Because the order of terms can be exchanged, there are
seven possible conclusions (excluding the repetition of the
premise) to be considered for each of the four possible
premises.
As with syllogisms, immediate inference problems can be
classified into three types, where there is no doubt about the
existence of entities referred to by the various terms:
Necessary. The conclusion statement must be true given that
the premise is true. Of the 28 problems, 6 are of this kind.
Possible. The conclusion statement might be true given that
the premise is true. Of the 28 problems, 12 are of this kind.
Impossible. The conclusion statement cannot be true given
that the premise is true. Of the 28 problems, 10 are of this
kind.
1498
Note that logically any conclusion that is necessary is also

possible. The problems we classify as Possible are in fact
ones in which the conclusion is possible but not necessary.
To demonstrate the method, we consider an example of
what can be inferred from the premise "All A are B." Each
of the following possible conclusions can be presented in
turn for evaluation:
1.
2.
3.
4.
5.
6.
7.
NoAareB
Some A are B
Some A are not B
All Bare A
No B are A
Some B are A
Some B are not A
Conclusions 2 and 6 are necessary; 4 and 7 are possible

(but not necessary), whereas 1, 3, and 5 are impossible.
Hence, if the participant was instructed to decide whether
each of these conclusions was necessary given the premise,
the correct answer would be yes for 2 and 6 and no for all the
others. If the instruction were to decide which were possible,
then the correct answer would be yes for 2,4,6, and 7 and no
for 1,3, and 5.
In Experiment 1, participants were presented with immediate inference problems under either Necessity or Possibility instructions. This allows us to test predictions P1-P3
identified earlier. Note that in this design there is an
inescapable confounding between the type of instruction
(Possibility or Necessity) and the kind of answer constituting a correct response. For example, because many more
conclusions are in fact possible than necessary, any general
bias to accept conclusions would lead to more correct
responding in the Possibility group compared with the
Necessity group. However, this does not present us with a
problem in testing our predictions above, which are in any
case phrased in terms of conclusions accepted rather than
correct decisions. Note also that predictions P2 and P3 both
involve comparisons within groups, in which instruction
type is held constant.
problems with a layout as in the example that follows:

GIVEN THAT
Some K are R
IS IT NECESSARY THAT
Some R are not K
The above example is in the form shown to the Necessity group.
In the Possibility group, the second header read "IS IT POSSIBLE
THAT." Underneath were two boxes showing the words YES and
NO in uppercase. Participants signaled their decision by moving a
mouse pointer into either box and clicking. The program randomly
assigned letters to the sentences excluding I and O.
Procedure. Participants were run in small groups in a laboratory containing several computers running identical software. Each
had their own workstation and worked independently at their own
pace. Participants were randomly assigned to the experimental
conditions on entering the room. As a group, they were given a
brief introduction as to what the tasks of both Experiments 1 and 2
would involve, and they were informed that their data would be
confidential and that they had a right to withdraw at any time
without penalty.
The sheets containing the instructions for Experiment 1 were
already present at the participants' computer terminals; they were
asked to read these and then to start as soon as they were ready.
They were instructed that the experiment was designed to find out
how people solve logical reasoning problems. (Because naive
reasoners are known to assume that assertions of the form "All A
are B" implicate the existence of A's and B's, we did not feel it
necessary to state this explicitly in the instructions. Our results
support our assumption.) It was explained that the problems would
be presented one at a time on separate screens. The problem layouts
were described, and participants were instructed to click in the YES
or NO box to make their response. Necessity participants were
instructed to click YES if the conclusion must follow from the
premises and NO if it did not follow. Possibility participants were
instructed to click YES if the conclusion could follow from the
premises and NO if it could not. Each participant received the 28
problems in an independently randomized order. Both answers and
response latencies1 were recorded by the program and saved to a
disk.
Results and Discussion

This study involves collection of data sets that could be
analyzed for a number of theoretical purposes other than
Method
Design. Participants were run in two experimental groups
according to whether they received Necessity or Possibility instructions. Each evaluated the 28 immediate inference problems described above in an independently generated random order. Experiment 1 was the first task administered to these participants, who
then proceeded to cany out the tasks of Experiment 2, reported
below.
Participants. A total of 120 undergraduate students at the
University of Plymouth, Plymouth, England, took part in this study
and received payment for their participation. Participants were
randomly allocated to the two experimental groups, 60 in each. All
were familiar with the use of computers including the operation of
a mouse.
Materials. The instructions for the tasks were presented separately on sheets of A4 paper. However, the reasoning problems
were presented on an Acorn computer using a custom-written
computer program. A separate screen was used for each of the 28
Although latency measures were collected in these experiments, we decided not to report them. Interpretation of the latency
data was considered problematic for two main reasons. First, in the
large experiment (Experiment 2) people had to evaluate four
conclusions in relation to the premises pair on each screen with
only one latency recorded. Hence, we cannot separate the time
taken for reasoning about each conclusion. Second, there is much
variability of responding (our main measure of interest) associated
with different types of logical problems in all of the experiments, so
that overall latencies are based on a variable mix of yes and no
responses. There is no possibility of excluding "error" trials as
with a simple cognitive task. Given the multifactorial nature of the
designs, there is also no practical method of analyzing responses to
yes and no decisions separately. In addition to these main problems,
we are also of the view that psychologists currently lack a good
theory of how response latencies map onto cognitive processes
with problems of this complexity.
those that we have space to describe here. Consequently, a

complete list of the percentage of conclusions endorsed on
each problem by each group is shown in Appendix A. We
concentrate our analysis here on the three categories of
problems identified a priori as having conclusions that were
Necessary ( = 6), Possible (n = 12), or Impossible (n = 10)
given their premises. The mean percentage acceptance of
conclusions in each category within the two instruction
groups is shown in Figure 1. Note that more conclusions
were endorsed in the Possibility group throughout and that
the order of endorsement of problem type was Necessary >
Possible > Impossible under both forms of instruction. The
latter trend is to be expected, not just by the model theory but
by any account that allows .a significant element of deductive
competence in syllogistic reasoning.
A 2 X 3 analysis of variance (ANOVA) was carried out to
test the effect of groups (Necessity vs. Possibility) and types
of inference (Necessary vs. Possible vs. Impossible). For
purposes of this analysis, the percentage of inferences
endorsed by each participant was computed for each problem type such that each participant contributed three data
points to the ANOVA. There was a highly significant effect
of group, reflecting the greater endorsement under Possibility instructions, F(l, 118) = 50.89, MSE = 0.421,p < .001.
There was also a highly significant main effect of problem
type, F(2, 236) = 283.8, MSE = 0.349, p < .001. In
addition, the interaction between the two variables was
significant, F(2, 236) = 7.07, MSE = 0.349,/? < .002. The
main effect of instructions confirms our general prediction
PI, discussed earlier. The interaction appears to reflect the
larger difference between the two groups for conclusions
that were possible. The phenomenon is not surprising,
100
20-
Necessary
Possible
Impossible
Figure 1. Percentage of conclusions accepted in Experiment 1

(immediate inference) in Group N (Necessity instructions) and
Group P (Possibility instructions).
1499
because these conclusions differ in their correct answer

between the two groups: it is yes for the Possible group and
no for the Necessary group.
Follow-up analyses were conducted to test predictions P2
and P3 derived from the mental model theory. As a reminder,
P2 states:
P2. It will be easier to decide that a conclusion is possible if it
is also necessary. Specifically, we predict more endorsements
of possibility for Necessary than for Possible problems.
The rationale for this prediction is that people only need to
identify one model that supports the conclusion to decide
that it is possible. On Necessary problems all models of the
premises support the conclusion, whereas on Possible problems there is at least one model that does not. Hence, it will
be easier to find a supporting model on Necessary problems.
This prediction was tested using a one-tailed related-groups t
test comparing endorsement of Possible and Necessary
problems for the group given Possibility instructions. The
prediction was strongly confirmed, ?(59) = 4.84, p < .001.
Prediction P3 states:
P3. It will be easier to decide that a conclusion is not
necessary if it is also not possible. Specifically, we predict that
more Possible than Impossible problems will be endorsed as
necessary.
The rationale here is that reasoners need only identify one
model of the premises that does not contain the conclusion in
order to decide that the conclusion is not necessary. This will
be easier on Impossible problems because no model of the
premises supports the conclusion than on Possible problems
in which at least one model will support it. This prediction
was tested with a one-tailed t test comparing acceptance
rates for Possible and Impossible problems under Necessity
instructions and was again supported, f(59) = 17.04, p <
.001.
All of the main predictions have then been strongly
supported in our analysis of conclusions accepted on the
immediate inference task of Experiment 1. An additional
aspect of the frequency data is also worthy of a note.
Whereas Possible problems are less frequently endorsed
than Necessary problems overall in both instruction groups,
the data under Necessity instructions seem to follow a
bimodal distribution. Of the 12 Possible problems, the 7
most frequently endorsed ranged from 41% to 91% and the
remaining 5 from 3% to 11%. The first group averages
75%slightly more than the mean acceptance for Necessary problems with these instructions; the second group
averages 6.5%a little less than the mean for Impossible
problems. Although retrospective, this observation suggests
that reasoners often fail to search for counterexamples even
when instructed to make decisions of necessity. The data are
consistent with the hypothesis that some Possible problems
suggest an initial model that supports the conclusion
leading to endorsement of the fallacywhereas others do
not. We return to this issue in discussing the results of
Experiment 2 and providing the rationale for Experiment 3,
both involving syllogistic reasoning proper.
1500
Experiment 2: Syllogistic Reasoning

Experiment 2 extends the investigation reported in Experiment 1 to syllogistic inference, so we first provide some
brief discussion of the logic and psychology of syllogisms. A
syllogism has two end terms, which we refer to as A and
Cin order of mention in the premisesand a middle or
linking term which we call B. The conclusion can be in
either order A-C or C-A. Either premise or the conclusion
may take each of the four forms (moods) discussed earlier
and investigated in the immediate inference task of Experiment 1. For example, a possible syllogism is
All A are C
Some B are not C
Therefore, some A are not C
In classical syllogistic logic, a distinction is drawn
between four figures of syllogism, which are defined by the
order of the terms in the premises with reference to the order
in the conclusion. Classical syllogisms are also defined in a
way that fixes only one order of premises. In this study we
wish to consider all possible ways of presenting these
problems with conclusions. Hence, we follow the practice of
Johnson-Laird and Bara (1984) in defining figure by the
order of terms in the premises without regard to the
conclusion. There are four such possible orderings:
Figure 1
Figure 2
A-B
B-C
B-A
C-B
Figure 3
A-B
C-B
Figure 4
B-A
B-C
In an evaluation task, as used in the present study, each of

the above figures can be presented with a conclusion in
either AC or CA order. The number of possible problems
increases considerably in the evaluation paradigms because
the two possible conclusion orders must be multiplied by the
four possible moods. This results in 512 distinct problems.
All of these were presented for evaluation in Experiment 2.
To keep the task within reasonable bounds, however,
participants evaluated either the AC or CA order but not both. In
addition, they evaluated four conclusions on the same screen for
a given premise pair, so that they effectively evaluated 256
syllogisms each, presented on 64 separate screens.
There are numerous theories of syllogistic inference to be
found in the psychological literature (see Evans et al., 1993).
The details of most of these need not concern us here. We
note briefly that the theories fall into two broad types. One
tradition emphasizes nonlogical processes in which the form
of the premises biases the choice of conclusion without
intervention of deductive processes. Although this approach
has been out of fashion for many years, a new nonlogical
theory based on probability heuristics has been developed
recently by Chater and Oaksford (1999). With regard to
deductive approaches, a provisional account of reasoning
with quantifiers has been offered by Rips (1994) within the
mental logic tradition. A detailed account of syllogistic

reasoning has been developed over the years by mental
model theorists (see Johnson-Laird & Byrne, 1991) and has
been implemented in a working computer program (JohnsonLaird & Byrne, 1994). A significant development is the
verbal reasoning model of Polk and Newell (1995), which
is also model based but differs in some significant respects from the theory presented by Johnson-Laird and his
colleagues.
As stated earlier, our objective in this research is to test
predictions derived from general principles of the model
theory that were independent of implementation details. For
this reason, we omit a description of the detailed account of
syllogistic reasoning offered by Johnson-Laird and Byrne
(1991), which is not required to follow the rationale and
analysis of the present study. The experiment permits us to
test the same three predictions PI, P2, and P3 about the
difference between necessary and possible inference that
were assessed on the simpler immediate inference task in
Experiment 1.
Method
Design. Participants were run in four experimental groups
immediately following the presentation of the immediate inference
task reported as Experiment 1. Participants received the same kind
of experimental instruction (Necessity or Possibility) as they were
given in Experiment 1 but were subdivided according to whether
they received conclusions to evaluate in AC or CA order (see
introduction to this experiment). Each participant was shown 64
separate computer screens, one for each possible premise pair. For
each pair they were required to indicate whether each of four
possible conclusions was either necessary or possible according to
instruction group. The four possible conclusions were those in each
of the four moods and in an order (AC or CA) dependent on the
subgroup to which participants were assigned.
Participants. The same 120 students who participated in
Experiment 1 served as participants for Experiment 2, administered
immediately afterward.
Materials. Problems were presented in a similar form to those
of Experiment 1, except that there were two premises and a list of
four conclusions to be evaluated. An example screen layout is as
follows:
GIVEN THAT
Some T are P
No P are G
IS IT NECESSARY THAT
AllTareG
Some T are G
No T are G
Some T are not G
YES
D
NO
The above example would be given to the Necessity-AC group.

Possibility participants were shown a second header that read, IS IT
POSSIBLE THAT, and those in CA groups would have received
the conclusions reversed: All G are T, and so on. Letters were
assigned randomly and independently to logical forms by the
program on each problem. The order of the four conclusion types
(moods) was fixed as above as A, I, E, O for all problems.
Procedure. On completion of Experiment 1, participants were
presented with further written instructions titled TASK 2. They
were told that there would be another series of logical problems but
that this time there will be two statements and that they would be
asked, given that these statements were true, to decide which of the
following statements must be true (Necessity group) or could be
true (Possibility group). They were shown an example of the screen
layout and were told to signal their answers by clicking the mouse
in the YES or NO box next to eaeh of the following statements.
They were told that they could decide about the statements in any
order but that a response once made could not be changed.
Whenever participants clicked in a box, it was filled on the
screen. When all four responses were made to a screen, a box
appeared underneath labeled CLICK HERE FOR THE NEXT
PROBLEM. Thus participants were self-paced and could rest
between problems in the long sequence as required. The program
recorded both the responses and the latencies and saved them to
disk. The 64 screens were presented in an independently randomized order for each participant.

A list of response frequencies to all individual problems is
given in Appendix B. Of the 256 premise-conclusion
combinations evaluated by each participant, 24 were Necessary (conclusion must follow from premise), 208 were
Possible (conclusion could follow from premises), and 24
were Impossible (conclusion could not be true given the
premises). These numbers reflect the fact that where, for
example, premises support a conclusion of the form "All G
are T," they also support "Some G are T" and "Some T are
G," given that naive reasoners take universal assertions to
imply the existence of the relevant entities and that "some"
is taken to mean "at least some, and possibly all." The mean
acceptance rates for each problem type in each instruction
group are shown in Figure 2 averaged over the AC and CA
subgroups (this factor had no overall effect). The graph is
extraordinarily similar to that for Experiment 1 (see Figure
Necessary
Possible
Impossible

(syllogistic inference) in Group N (Necessity instructions) and
Group P (Possibility instructions).
1501
1) when one considers the additional complexity of the two

premise syllogistic format. A 2 X 2 X 3 ANOVA was
conducted again on aggregated data with each individual
contributing one data point for each of the three types of
problems. The additional (group) factor to that of Experiment 1 was conclusion order.
The trends in the ANOVA were similar to those observed
in Experiment 1. There was a significant main effect of
instruction group, F(l, 116) = 64.7, MSE = 0.374, p <
.001, reflecting higher acceptance rates under Possibility
instructions. There was also a significant main effect of
problem type, F(2, 232) = 446.6, MSE = 0.218, p < .001,
reflecting the order Necessary > Possible > Impossible.
Again, as in Experiment 1, the interaction between these two
factors was significant, F(2,232) = 6.03, MSE = 0.218,/? <
.01, reflecting the fact that instructions had the largest effect
on problems of Possible type. There was, however, no effect
of conclusion order.
As in Experiment 1, prediction PI is confirmed as a main
effect in the ANOVA, but P2 and P3 require follow-up tests.
These were assessed using one-tailed related-groups t tests,
with separate tests being carried out for the AC and CA
subgroups {dfs, = 29). P2 predicts that participants given
Possibility instructions will accept more Necessary than
Possible conclusions. This was confirmed for both AC
(f = 6.11, p<.00l)
and CA (f = 5.19, p < .001) subgroups. P3 predicts that participants given Necessity instructions will accept more Possible than Impossible conclusions.
This also was confirmed for both AC (t = 12.3, p < .001)
andCA(f = 9.2, p < .001) subgroups.
This experiment provides the opportunity to look at some
other issues of interest using the syllogistic evaluation task.
An important claim of the model theory is that people will be
less inclined to draw valid conclusions from syllogistic
premises consistent with multiple mental models, as opposed to a single model. The point here is that when the
conclusion is validas on our Necessary problemsthen
any model of the premises will support the conclusion.
Hence, if people consider only one model of the premises,
there should be no difference in performance on single and
multimodel problems. If, on the other hand, people make an
effort at deduction by trying to prove the conclusion in all
models of the premises, then multimodel problems should
be more difficult. To our knowledge, this prediction has not
previously been tested on the syllogistic evaluation task,
although evidence for it was adduced by Johnson-Laird and
Bara (1984) using the production task. To maximize comparability with their study, we eliminated syllogisms with valid
but weak conclusions from this analysisfor example, ones
in which the conclusion to be evaluated was "Some A are C"
even though the conclusion "All A are C" was supported by
the premises. Such conclusions are rarely generated on a
production task.
Of the 256 problems presented to each participant, just 36
are valid with strong conclusions. For this subset, we
examined the difference in acceptance rates between single
and multiple-model problems. Under Necessity instructions,
we would expect a similar trend to that observed by
Johnson-Laird and Bara (1984)that is, fewer acceptances
1502
of the valid conclusions on multiple-model problems. Under

Possibility instructions, we should not find such a trend,
however, because in this case only a single model need be
found supporting the conclusion for the participant to justify
a YES response. The trends in our data accord with this
hypothesis. Of the 36 problems, 18 are of a kind classified as
single model, and 18 as multiple model by Johnson-Laird
and Bara. Under Necessity instructions, 81% of single
model problems had their conclusions accepted, compared
with 59% of multiple-model problems. As expected, this
trend was weaker under Possibility instructions: 88% acceptance of single-model problems compared with 77% acceptance on multiple-model problems. To test the significance
of this interaction, we computed the difference between
acceptance rates of single- and multiple-model problems for
all participants. These difference scores were then compared
on a one-tailed between-groups t test, comparing the 60
participants who received Necessity instructions with the 60
in the Possibility group. The test was comfortably significant, r(118) = 2.89, p < .005.
The above finding supports the model theory claim that
reasoners look for counterexamples, at least on this restricted set of 36 syllogisms. However, looking at the data as
a whole, it is clear that any such facility must be weak, in
view of the very high rates of endorsement of fallacies
reported in this study as well as elsewhere in the literature.
Nevertheless, it appears that the lower rate of acceptance of
Possible than Necessary problems under Necessity instructions could be due to some searching for counterexamples
that successfully leads to refutation of some of the potential
fallacies. However, there is an alternative interpretation of
the lower rate of acceptance of Possible conclusions that we
need to consider. Suppose that people are willing to accept
as necessary any conclusion supported by the first model of
the premises they think of. Suppose also that the premises of
Possible problems consistently suggest a model that either
does or does not support the conclusion. The result would be
that some Possible problems would be endorsed as frequently as Necessary problems and others as infrequently as
Impossible problems. We have already noted that this was
the case for the immediate inference task in Experiment 1.
Inspection of the data for Possible problems under Necessity
instructions in the syllogistic inference task of Experiment 2
suggested a similar trend was present. Some fallacies were
strongly endorsed, as for example:
All A are B
Some B are not C
Ninety percent of participants judged the above conclusions to follow under instructions of Necessity, even though
it is fallacious. Judgments of Possibility were also 90%,
suggesting that (a) people easily imagine situations (mental
model) in which premises and conclusion are true but (b)
they fail to seek counterexamples to prove necessity. By
contrast, consider the following syllogism:
Some A are B
All C are B
Therefore, all C are A
Only 3% of participants endorsed this syllogism under

instructions for Necessity, but 43% did so when given
Possibility instructions. This suggests that participants do
not initially think of a model of the premises that supports
the conclusion in this case and hence reject it as unnecessary.
However, in this case it appears that some effort to find an
alternative and supporting model does occur when people
are asked if the conclusion is possible. A trend to seek
alternative models when proving possibility rather than
necessity could account for the interaction between instructions and logical classification of problems reported above.
Recall that this interaction reflects the fact that the increase
in acceptance of conclusions under Possibility instructions
was larger for Possible syllogisms than either Necessary or
Impossible ones.
Experiment 3 was designed to establish further the
reliability of the finding that some fallacies are consistently
made and others withheld and to investigate the processes
that underlie this effect.
Experiment 3: Possible Strong Versus Possible

Weak Syllogisms
As noted earlier, Experiments 1 and 2 have strongly
supported the a priori predictions that we derived about
possible and necessary inference from the mental models
account. However, there is also an observation that fits less
comfortably with this framework. It appears that when
Possible problems are assessed under Necessity instructions,
some are accepted with very high rates comparable with
those of Necessary problems, whereas others have very low
acceptance rates comparable with those associated with
Impossible problems (see above for examples of the two
kinds of syllogism). A possible interpretation is that people
are not searching for counterexample models, but rather are
simply saying yes if the first model suggested by the
premises contains the conclusion.
Experiment 2 is to our knowledge the only study to have
examined people's evaluations to all possible conclusions
with all possible premise types, and certainly the first to have
done so with judgment of Possibility as well as Necessity
required. Although it should prove useful to theorists of all
persuasions to have such a database available (Appendix B),
there is a disadvantage inherent in such a comprehensive
approach, namely, that a very large number of problems
have to be administered to each participant. This enforced
the use of a method in which 64 premise pairs were
presented with four conclusions to be evaluated, as the
alternative would have required processing of 256 separate
computer screens for each participant, even keeping conclusion order as a between-participant variable.2
Another possible problem with the procedure of Experiment 2

is that the conclusions to be evaluated are logically related to each
other. For example, if a participant has decided that the conclusion
"All A are C" is necessary given the premises, then they might
conclude that "Some A are not C" is impossible without returning
1503
Table 1
Syllogisms Selected on the Basis of Inference Rates in Experiment 2 for Further
Investigation in Experiment 3
B-A
C-B
A-B
B-C
Syllogism
Necessary
Impossible
Possible strong
Possible weak
A-B
C-B
B-A
C-B
ac
ca
ac
ca
ac
ca
AAa
AEe
IAi
IEo
AEa
AEi
IAe
IE a
AOo
Hi
OAo
OOo
He
Ole
OEa
OOa
AEe
IAi
EAo
EIo
AAe
AEa
AEi
IAe
Ali
10 o
OAo
01 o
AOa
II a
IEa
OOa
Ali
AEo
IEo
EAe
AAe
Ale
IEa
EAa
AOo
IAi
II o
OAo
Ala
AOa
OAa
OOa
AAa
Ali
EAe
EIo
AAe
AAo
EAa
El a
Hi
10 o
OAo
OIo
AOa
IEa
EOa
OOa
AEe
IEo
EAe
OAo
AEa
AEi
EAa
OAa
AIo
IAi
IOo
OIo
Ala
IAe
II a
OEa
AEe
AOo
EAe
EIo
AEa
AOa
EAa
El a
Ali
IAi
IOo
OAo
IAa
IEa
EOa
Ola
ac
ca
Ali
AEo
IEo
EAe
AAe
Ale
IEa
EAa
AOo
IAi
II 0
OAo
Ala
AOa
OAa
OOa
AAa
Ali
EAe
EIo
AAe
AAo
EAa
El a
II i
IOo
OAo
OIo
AOa
IEa
EOa
OOa
Note. Uppercase letters represent premise pairs and lowercase represent conclusion. Key to
premise and conclusion types: A = All X are Y; E = No X are Y; I = Some X are Y; O = Some X are
notY.
Experiment 3 involves presenting selected syllogisms

such that each participant receives only 64 problems in total
with each presented on a single screen, similar to the
immediate inference problems of Experiment 1. In selecting
syllogisms, our main interest was in the hypothesis that there
are two kinds of Possible problems that we define as follows:
dorsed Possible problems in each figure. Similarly, Possible

Weak problems were the least frequently endorsed. Finally,
the least frequently endorsed Impossible problems were
selected. This method provides a strong test of whether
Possible Strong problems are endorsed as often as Necessary
problems, and Possible Weak problems as rarely as Impossible problems. With Necessity instructions, the levels were
Possible Strong: problems whose conclusions are possible comparable in both cases (see the black bars in Figure 3).
but not necessary given their premises, but that are frequently
When the same problems were considered under Possibility
endorsed as having necessary conclusions.
Possible Weak: problems whose conclusions are possible instructions, however, Possible Weak problems had a somebut not necessary given their premises, and that are infrewhat higher acceptance rate than Impossible problems (see
quently endorsed as having necessary conclusions.
Figure 4). Although the Necessity data of Experiment 2 were
used as the basis for problem selection, participants received
In other words, Possible Strong syllogisms are the fallathe
same syllogisms in Experiment 3 whether responding to
cies that people tend to make (under instructions for
Possibility
or Necessity instructions.
Necessity), and Possible Weak syllogisms are the fallacies
they tend to avoid. The identification of these two categories
Method
has been necessarily retrospective in the two experiments
reported to date. Hence, we decided to choose examples of
Design. There were four groups of participants (as with
Possible problems with high and low endorsement rates in
Experiment 2) according to conclusion order (AC or CA) and the
Experiment 2 and submit them to independent replication in
type of instruction (Necessity or Possibility). Each participant was
Experiment 3. The actual syllogisms selected are shown in
presented with 64 separate syllogisms generated by two withinTable 1. Note that the 64 problems presented to AC groups
participant variables on four levels each: the type of problem
were not necessarily the same as those presented to CA
(Necessary, Impossible, Possible Strong, and Possible Weak),
groups. This is because problems were selected strictly on
syllogistic figure, and a repetition of four examples in each cell.
the basis of performance in Experiment 2 under Necessity
Participants. Eighty undergraduate students from the Univerinstructions. The Necessary syllogisms selected were the
sity of Plymouth took part in this experiment and received either
payment or course credits for their participation. None had
four most frequently endorsed associated with each figure of
participated in Experiment 1 and 2 or in other similar studies. They
premises and conclusion order. Possible Strong syllogisms
were randomly allocated to the four experimental conditions, in
were produced by choosing the four most frequently ento the premises. Any possible distortion of responses resulting from
this will be eliminated by the procedure in Experiment 3, in which
only one conclusion is evaluated at a time with respect to a
particular pair of premises.
groups of 20.
Materials and procedure. The computer program randomly
allocated letters (excluding I and O) for the end and middle terms
used in the syllogisms that were previously selected as shown in
Table 1. The screen layout used was similar to that of Experiment 1
except that two premises were given with a single conclusion to be
tions on screen depending on the experimental condition to which

they were assigned. Instructions were similar to those used in
Experiment 2 except for the requirement to evaluate a single
conclusion on each screen. Before the syllogisms were randomly
presented, participants were given practice with the mouse to get
used to the manner of signaling Yes and No responses.

(Ex 2) under Necessity instructions broken down into four problem
types: Necessary (N), Possible Strong (PS), Possible Weak (PW),
and Impossible (I), together with replication data from Experiment
3 (Ex 3).
evaluated by pressing the mouse in a YES or NO box. The 64

syllogisms were presented on separate screens in an independently
randomized order for each participant.
Initially the participants were given an overview of the structure
of the experiment and informed of their right to withdraw and
aspects of confidentiality. They were then presented with instruc-
100 -,
The frequency of acceptance of different logical types

under Necessity instructions is shown in Figure 3 together
with the corresponding data from Experiment 2. It can be
seen that Experiment 3 provides a close replication, dispelling any possibility that the difference between Possible
Strong and Possible Weak problems was magnified by
retrospective choice of these categories (e.g., due to statistical regression). It remains the case in Experiment 3 that
acceptance rates of Necessary and Possible Strong problems
are very similar. Possible Weak problems are also accepted
only with slightly higher frequency than Impossible problems. Figure 4 compares acceptance frequencies for the
different problem types under Possibility instructions, again
comparing the data of Experiments 2 and 3. A good
replication is again seen, but the tendency for Possible Weak
problems to be more highly accepted than Impossible
problems appears more marked in the replication.
Mean percentage acceptance of conclusions in Experiment 3 is shown in Table 2, broken down by figure, problem
type, conclusion order, and instruction type. Each cell
represents the mean of 20 participants' responses with four
replications of syllogisms. The use of replications permitted
us to run a four-way ANOVA on the frequency of Yes
responses, with instruction and conclusion order as betweengroups variables and figure and problem type as withingroup variables. There was a significant main effect of
instruction, F(l, 76) = 18.51, MSE = 0.127,/? < .001, such
that more conclusions were accepted under Possibility
(57%) than Necessity (49%) instructions and as might be
80-
Table 2
Percentage of Conclusions Accepted in Experiment 3,
Broken Down by Problem Type, Figure, Conclusion Order,
and Instructional Group
Instruction and
problem type
20-

under Possibility instructions broken down into four problem
types: Necessary (N), Possible Strong (PS), Possible Weak (PW),
and Impossible (I), together with replication data from Experiment
3 (Ex 3).
Necessity
Necessary
Impossible
Possible strong
Possible weak
M
Possibility
Necessary
Impossible
Possible strong
Possible weak
M
A-B
B-C
B-A
A-B
B-A
C-B
C-B
B-C
ac
ca
ac
ca
ac
ca
ac
ca
81
9
84
18
48
79
21
83
23
51
70
25
80
24
50
84
20
81
19
51
76
13
74
19
46
86
15
81
15
49
84
24
70
13
48
79
13
78
19
47
80
17
79
19
89
31
88
59
67
78
15
93
26
53
86
30
86
51
64
86
5
93
21
51
74
21
90
45
58
83
14
91
26
54
86
30
79
43
59
85
19
79
29
53
83
21
87
38
expected a very highly significant effect of problem type,

F(3,228) = 412.00, MSE = 0.090, p < .001.
In Experiments 1 and 2, problem type (defined on three
rather than four levels) interacted significantly with instructions, such that relatively more Possible conclusions were
accepted under instructions for Possibility. The two variables also interacted significantly in this analysis, F(3,
228) = 4.85, MSE = 0.903, p < .005. With the breakdown
of the Possible category into Strong and Weak, however, it
appears from the means (right-hand column of Table 2) that
the major difference lies in the acceptance rates of Possible
Weak problems that are accepted at a rate of 38% under
Possibility instructions compared with only 19% under
Necessity instructions. This difference is statistically significant, (78) = 4.48, p < .01, and is theoretically important
because it suggests that under instructions for Possibility,
reasoners do seek alternative models of the premises if the
first fails to support the premises, rather than simply
responding No. This contrasts with the very high acceptance
rates on Possible Strong problems under Necessity instructions, which suggest that rather little search for alternative
models occurs when the first model identified confirms the
conclusion.
There were several other significant effects. First, there
was a significant if small main effect of figure, F(3, 228) =
3.04, MSE = 0.026, p < .05. Conclusions were most often
accepted in Figure 1 (55%) and least often in Figure 4
(52%). There was no main effect of conclusion order, but
this variable did interact significantly with instruction, F(l,
76) = 7.88, MSE = 0.127, p < .01. Inspection of Table 3
suggests that this was due to greater acceptance of AC
conclusions in the Possibility group only. There is no
obvious explanation for this trend. The remaining trends
were interactions involving problem type. This interacted
significantly with figure and conclusion order separately,
and all three variables significantly if weakly interacted
together, F(9,684) = 2.12, MSE = 2.12, p < .05.
Finally, it should be noted that the strong replication of
similar findings for the same syllogisms presented in Experiment 3 when compared with Experiment 2 counters two
possible criticisms of the design of the previous experiment.
First, it might be suggested that the prior experience of
participating in Experiment 1 might have had some effect on
performance in Experiment 2; Experiment 3 had no such
prior task. Second, it could be argued that participants
became fatigued or bored while processing the large number
of syllogisms used in Experiment 2, which might add noise
to their data. We think this was unlikely to be true in any
case, because of the self-pacing provided and the fact that
the analysis showed very clear and systematic trends. The
fact that the data were so similar on corresponding syllogisms in Experiment 3, in which only one quarter as many
were used, confirms this view.
General Discussion
Our experiments have shown that logically untrained
individuals are able to draw modal conclusions both to
single quantified assertions and to syllogistic premises. The
1505
results also confirm our three main predictions based on the

mental model theory. First, the participants were more likely
to endorse conclusions about what is possible than what is
necessary (Prediction PI). Possibility calls for only a single
model of the premises to support the conclusion, whereas
necessity calls for all the models of the premises to support
the conclusion. Second, the participants were more likely to
endorse a possible conclusion if the premises supported its
necessity as opposed to its mere possibility (Prediction P2).
In the case of necessity, the conclusion holds in all the
models of the premises, and so no matter which model of the
premises reasoners construct, it will support the conclusion.
In contrast, if the premises only support a possible conclusion, then reasoners need to search for the model of
the premise in which the conclusion holds, and they need
to ignore those models that are counterexamples to the
conclusion.
Our third hypothesis was that participants would be more
likely to decide that a conclusion was not necessary if it was
impossible (Prediction P3). A conclusion that is impossible
holds in no models of the premises, whereas a conclusion
that is merely not necessary holds in at least one model of the
premises, and so reasoners are liable to go wrong in the latter
case if they focus on a model in which the conclusion holds.
In other words, they have to search for a counterexample to
the content of the conclusion to infer that it is not necessary.
The search is easier where all the models of the premises are
counterexamples than where at least one is not. The three
predictions were corroborated in Experiment 1 in which the
participants evaluated immediate inferences from a single
quantified premise to a quantified conclusion, in Experiment
2 in which they evaluated all possible modal conclusions
from syllogistic premises, and in Experiment 3 in which they
evaluated a subset of modal conclusions from syllogistic
premises.
One initially unexpected effect was that problems supporting possible conclusions fell into two categories. Some of
them were regularly taken to imply necessary conclusions,
whereas others were rarely taken to imply necessary conclusions and indeed sometimes were not even taken to imply
possible conclusions. This phenomenon was first detected in
Experiment 1. It also occurred in Experiment 2, and it was
corroborated by Experiment 3, which we designed to
contrast the two sorts of problems, which we termed
Possible Strong and Possible Weak. It is particularly striking
in Figure 3 that Possible Strong syllogisms are endorsed as
frequently as Necessary ones and that Possible Weak
syllogisms are endorsed as infrequently as Impossible ones.
This suggests that any search for counterexample models is
weak in the present study and that most participants are
basing their conclusions on the first model that occurs to
them (we return to this issue later). On this argument, the
fallacies are sharply divided into our two categories because
some premise pairs consistently suggest an initial model that
supports the conclusion (Possible Strong), whereas others
consistently suggest a model that negates the conclusion
(Possible Weak).
It is clear that such a finding cannot be accounted for at the
1506
very general level of our other predictions. However, the

computer program implementation of the model theory
reported by Johnson-Laird and Byrne (1994) does generate
models in specific orders, so we were able to examine the
output of the program for our two categories of syllogisms to
try to confirm our intuitive account. The results provided
strong corroboration. All premise pairs studied in the two
Possible conditions of Experiment 3 were run through the
program. In 29 out of the 32 cases, the Possible Strong
syllogisms had an initial model that supported the given
conclusion, whereas none of the 32 Possible Weak syllogisms had an initial model that supported the given
conclusion.
To see how this arises, consider an example of Possible
Strong syllogisms whose fallacious conclusion is endorsed
by most participants:
Some A are not B
All B are C
The initial model of the premises, as constructed by the
program, is
a
-,b
-,b
[b] c
[b] c
where . denotes negation and a, b, and c are tokens

representing exemplars for the classes A, B, and C. The
square brackets around b in the latter models indicate that b's
are exhaustively represented with respect to c's. In other
words, there cannot be b's found in models that do
not contain c's. This model supports the conclusion presented, and 75% of the participants in Experiment 3 inferred
(wrongly) that it was a necessary conclusion whereas
85% of the participants inferred (correctly) that it was a
possible conclusion. In fact, as the program also shows,
there is an alternative model of the premises that refutes the
conclusion:
a -^b
a -.b
[b]
[b]
c
c
c
c
Hence, the conclusion is possible but not necessary. An

example of a Possible Weak syllogism has the following
form:
All A are B
Some C are B
Therefore, all A are C
The initial model of the premises is:
[a]
[a]
This model does not support the given conclusion, but

instead the conclusion "Some A are C" (or its converse). But
this conclusion can also be refuted by an alternative model
of the premises:
[a] b
b c
[a] b
b c
To grasp that the given conclusion is at least possible,

reasoners need to construct another model of the premises:
[a] b c
[a] b c
c
As this account predicts, few participants in Experiment 3

wrongly inferred that the given conclusion was necessary
(15%), and only a minority inferred that it was even possible
(40%). It is important to note that discovery of such Possible
Weak syllogisms was made in this experiment because of the
methodology involved whereby participants were asked to
evaluate every possible conclusion for every possible premise
pair. With a production task method in which people are
instructed to produce a single, necessary conclusion, the
premise pairs involved would normally be associated with
the production of some alternative conclusion, as illustrated
in the above example. An interesting possibility for future
research would involve asking people to produce conclusions that are possible rather than necessary.
The program implementing the model theory works by
constructing separate models of the two premises and then
joining them together by way of the tokens representing the
middle term. Thus, additional steps are required to add
further tokens to the model (as in the example above of a
Possible Strong syllogism) or to split an individual, such as
[a] b c, into two distinct individuals: [a] b and b c (as in the
example above of a Possible Weak conclusion). These steps
yield models that are constructed after the initial models (see
Johnson-Laird and Byrne, 1994, for a detailed description of
the operations of the program).
Although the above analysis is very encouraging for the
model theory, we should note that alternative explanations
can be offered for the difference between Possible Strong
and Possible Weak syllogisms. In particular, the probability
heuristics model of Chater and Oaksford (1999) can account
for the difference in terms of the "min-heuristic." The
min-heuristic specifies that the conclusion drawn from
syllogistic premises will match the form of least informative
premise, in which informativeness has the following order:
all > some > no > some . . . not. Of the 32 Possible Strong
syllogisms shown in Table 1,30 have conclusions that match
the less informative premise, whereas the remaining 2 have
conclusions that are less informative than either premise. For
the 32 Possible Weak problems, all but 1 have a conclusion
that is more informative than at least one premise.
The equally strong account that both the model program
and the min-heuristic provide the difference between Possible Strong and Weak syllogisms is both surprising and
interesting. The min-heuristic is offered as part of an entirely
nondeductive account of syllogistic reasoning by Chater and
Oaksford (1999). The effect that it has is to ensure that
arguments are not supported when their conclusions are
more informative than the premises. This property is true not

only of all valid arguments (deduction cannot add information) but of many fallacies also. Hence, the theory accounts
(as it was designed to do) for the fact that in syllogistic
reasoning experiments, participants endorse most valid
arguments but also many fallacies. It is clear that the model
program, based on a quite different approach, has the effect
of first generating models whose provisional conclusions
conform with the min-heuristic. This coincidence has been
discussed recently in terms of the "atmosphere effect"
whose predictions are similar to those of the minheuristicby Shaw and Johnson-Laird (1998). These authors also examined the order with which the mental models
program produces models and observed, as we do here, that
errors often result from fallacious conclusions being supported by the initial models of the premises considered.
Let us return now to the question of whether reasoners
actually search for alternative models as envisaged by the
third stage of the model theory. This issue is theoretically
important because the emphasis placed on this stage by Polk
and Newell (1995) in their verbal reasoning theory is much
less than in the mental models account of, for example,
Johnson-Laird and Byrne (1991). It is also the aspect of the
model theory required to account for deductive competence.
We say above that the evidence for such searching is weak in
the present study. We do not say that it is absent because
other aspects of our dataas well those of other studies in
the literaturesuggest that at least some search for alternative models is occurring. First, a robust result in standard
syllogistic reasoning is that one-model problems are reliably
easier than multiple-model problemsevery study in the
literature, including those carried out by critics of the model
theory, has corroborated this result (see, e.g., Ford, 1995).
This difference is manifest on valid syllogisms taken separately in which any model of the premises supports the
conclusion. Thus if people were deciding on the basis of the
first model alone, there should be no difference.
The present study provides some important new evidence
for this claim. First, Experiment 2 showed that the effect
occurs when participants were instructed to judge necessity
of conclusionsequivalent to the validity judgments used in
conventional reasoning experiments. Second, for those participants who were instructed to judge the possibility of
conclusions, die effect was reliably reduced. This interaction
strongly suggests that people understand that a judgment of
necessity requires finding a conclusion in all models,
whereas judgment of possibility does not. This requires the
further assumption, however, that people realize that there
are multiple models but are not confident of their ability to
check them all.
Also pertinent is the significant interaction between
instruction type and logical classification reported in Experiments 1 and 2. Inspection of Figures 1 and 2 reveals that this
was a result of the difference in endorsement rates between
Possibility and Necessity instructions that was most marked
on Possible syllogisms (potential fallacies), which were not
subdivided at this stage of our report. Although a general
increase in acceptance rates under Possibility instructions
could be due to response bias, such as a caution effect, the
1507
interaction suggests that at least some people are searching

for alternative models and finding them. However, the
analysis of Experiment 3, in which the two kinds of Possible
problems were separated, indicates that this difference was
mostly due to a quite marked increase in acceptance rate of
Possible Weak syllogisms under Possibility instructions.
Thus we have clear evidence that people search for alternative models to prove the possibility of conclusions that are
not supported by the first model considered, but in this case
less clear evidence that people search for counterexamples
to establish the necessity of conclusions tiiat are supported
by the first model considered.
The issue of whether people search for alternative models
has also been investigated in studies reported elsewhere,
with mixed conclusions depending on the methodology
used. Newstead, Handley, and Buck (1999) used two
process-tracing methods to address this question. The first
method involved asking participants to indicate any other
conclusion they had considered immediately after responding, and the second involved drawing Euler circle diagrams.
Neither method produced evidence that people were considering alternative models. A ratiier different picture emerges
from the study of Bucciarelli and Johnson-Laird (1998).
They asked their participants to construct external models
(cut-out shapes) as they drew conclusions from syllogistic
premises, and they videotaped the resulting sequences of
models. Their results showed that reasoners consider a
variety of different interpretations of the premises, as
claimed by Polk and Newell (1995). However, the participants tended not to reencode the premises for one-model
syllogisms (in contrast with the verbal reasoning), and they
did generate sequences of models for multiple-model syllogisms. Whether they were searching for counterexamples
was less clear. However, another phenomenon of syllogistic
reasoning does support a search for counterexamples. Byrne
and Johnson-Laird (1990) tested their participants' ability to
recognize conclusions that they had drawn earlier to syllogistic premises. As a search for counterexamples predicts,
participants often falsely recognized conclusions supported
by an initial model of the premises when, in fact, they had
responded correctly that nothing followed from the premises. This result suggests that they had fleetingly considered the erroneous conclusion only to reject it as a result of a
counterexample.
In view of this somewhat ambiguous evidence, a cautious
conclusion might be that people can search for alternative
models but do not necessarily do so spontaneously. This
might explain why evidence for such search seems to
depend on the experimental method used. We are aware of
two factors that seem to influence strongly whether such a
search occurs. The first is the so-called belief bias effect that
might better be described as a debiasing effect. Although
commonly described as a tendency for people to endorse
fallacies whose conclusions are believable, it might be more
accurately described as a tendency to suppress fallacies
when their conclusions are unbelievable. Most problems
used in this literature are of the type we term Possible Strong
in which people will normally endorse the fallacy when the
conclusion is abstract or neutral. Those few studies that have
1508
included neutral as well as believable and unbelievable

conclusions have reported that the endorsements of neutral
conclusions are at least as high as for believable ones, and
that belief bias is due to suppression of endorsement of
unbelievable conclusions (Evans & Handley, 1997; Evans &
Pollard, 1990; Newstead, Pollard, Evans, & Allen, 1992).
This finding supports the mental model account of the
belief bias effect (see, e.g., Oakhill, Johnson-Laird, &
Gamham, 1989). When a putative conclusion is unbelievable, then people are motivated to search for counterexamples to disprove it. Also consistent with the view that
people may require some stimulus to engage in model
searching is the effect of instructions. Evans, Allen, Newstead, and Pollard (1994) showed that strong instructional
emphasis on logical necessity significantly reduced belief
bias and also significantly suppressed the tendency to
endorse possible but not necessary conclusions. If the
natural mode of thinking is inductive rather than deductive
(as argued by Evans & Over, 1996) then it makes sense that
deductive competence will be most apparent under conditions in which people are clearly motivated to make an effort
at deduction.
Finally, might there be an alternative explanation for our
main findings in terms of a theory based on formal rules?
Our principal predictions are independent of the details of
the model theory, and they cannot be made by current
versions of formal rule theories that describe only a mechanism for necessary inference. Osherson's (1976) rules for
modal inferences concern only a limited set of inferences
based on prepositional connectives and make no predictions
about quantified modal reasoning. Our findings clearly
provide a strong challenge for mental logicians as well as
those psychologists who believe that syllogistic reasoning
can be explained without any reference to an effort at
deduction.
References
Bell, V. A., & Johnson-Laird, P. N. (1998). A model theory of
modal reasoning. Cognitive Science, 22, 25-52.
Braine, M. D. S. (1978). On the relation between the natural logic
of reasoning and standard logic. Psychological Review, 85, 1-21.
Braine, M. D. S., & O'Brien, D. P. (1991). A theory of If: A lexical
entry, reasoning program, and pragmatic principles. Psychological Review, 98, 182-203.
Braine, M. D. S., Reiser, B. J., & Rumain, B. (1984). Some
empirical justification for a theory of natural prepositional logic.
In G. H. Bower (Ed.), The psychology of learning and motivation
(Vol. 18, pp. 317-371). New York: Academic Press.
Bucciarelli, M., & Johnson-Laird, P. N. (1998). Strategies in
syllogistic reasoning. Unpublished manuscript, Princeton
University.
Byrne, R. M. J., & Johnson-Laird, P. N. (1990). Remembering
conclusions we have inferred: What biases reveal. In J.-P.
Cavemi, J.-M. Fabre, & M. Gonzalez (Eds.), Advances in
psychology: Vol. 68. Cognitive biases (pp. 109-120). Amsterdam: North-Holland.
Chater, N., & Oaksford, M. (1999). The probability heuristics
model of syllogistic reasoning. Cognitive Psychology, 38, 191
258.
Evans, J. St. B. T. (1989). Bias in human reasoning: Causes and
consequences. Hillsdale, NJ: Erlbaum.
Evans, J. St. B. T, Allen, J. L., Newstead, S. E., & Pollard, P.

(1994). Debiasing by instruction: The case of belief bias.
European Journal of Cognitive Psychology, 6, 263285.
Evans, J. St. B. T, & Handley, S. J. (1997). Necessary and possible
inferences: A test of the mental model theory of reasoning
(Report to the Economic and Social Research Council, Grant
R000221742). Plymouth, United Kingdom, University of
Plymouth.
Evans, J. St. B. T, Newstead, S. E., & Byrne, R. M. J. (1993).
Human reasoning: The psychology of deduction. Hove, England:
Erlbaum.
Evans, J. St. B. T, & Over, D. E. (1996). Rationality and reasoning.
Hove, England: Psychology Press.
Evans, J. St. B. T , & Over, D. E. (1997). Rationality in reasoning:
The case of deductive competence. Current Psychology of
Cognition, 16, 3-38.
Evans, J. St. B. T., & Pollard, P. (1990). Belief bias and problem
complexity in deductive reasoning. In J. P. Caverni, J. M. Fabre,
& M. Gonzales (Eds.), Cognitive biases (pp. 131-154). Amsterdam: North-Holland.
Ford, M. (1995). Two modes of mental representation and problem
solution in syllogistic reasoning. Journal of Experimental Psychology: General, 115, 16-25.
Galotti, K. M., Baron, J., & Sabini, J. P. (1986). Individual
differences in syllogistic reasoning: Deduction rules or mental
models? Journal of Experimental Psychology: General, 115,
16-25.
Johnson-Laird, P. N. (1983). Mental models. Cambridge, England:
Cambridge University Press.
Johnson-Laird, P. N., & Bara, B. G. (1984). Syllogistic inference.
Cognition, 16, 1-62.
Johnson-Laird, P. N., & Byrne, R. (1991). Deduction. Hove,
England: Erlbaum.
Johnson-Laird, P. N., & Byrne, R. M. J. (1994). Models, necessity
and the search for counter-examples. A reply to Martin-Cordero
& Gonzalez-Labra and to Smith. Behavioral and Brain Sciences,
17, 775-777.
Manktelow, K. I., & Over, D. (1993). (Eds.). Rationality. London:
Routledge.
Newstead, S. E., & Griggs, R. A. (1983). Drawing inferences from
quantified statements: A study of the square of opposition.
Journal of Verbal Learning and Verbal Behavior, 22, 535-546.
Newstead, S. E., Handley, S. J., & Buck, E. (1999). Falsifying
mental models: Testing the predictions of theories of syllogistic
reasoning. Journal ofMemory and Language, 27, 344354.
Newstead, S. E., Pollard, P., Evans, J. St. B. T, & Allen, J. (1992).
The source of belief bias in syllogistic reasoning. Cognition, 45,
257-284.
Oakhill, J., Johnson-Laird, P. N., & Garnham, A. (1989). Believability and syllogistic reasoning. Cognition, 31, 117-140.
Osherson, D. N. (1976). Logical abilities in children: Vol. 4.
Reasoning and concepts. Hillsdale, NJ: Erlbaum.
Polk, T. A., & Newell, A. (1995). Deduction as verbal reasoning.
Psychological Review, 102, 533-566.
Rips, L. J. (1983). Cognitive processes in propositional reasoning.
Psychological Review, 90, 38-71.
Rips, L. J. (1994). The psychology ofproof. Cambridge, MA: MIT
Press.
Shaw, V., & Johnson-Laird, P. N. (1998). Dispelling the "atmosphere" effect on reasoning. In A. C. Quelhas (Ed.), Cognition
and context (pp. 169-200). Lisbon: Instituto Superior de Psicologia Aplicado.
1509
Appendix A
Percentage Endorsements of Conclusions in Experiment 1
Premise
Conclusion
Logic
Necessity
Possibility
Some A are B
No A are B
Some A are not B
All B are A
Some B are A
No B are A
Some B are not A
N
I
I
P
N
I
P
65
3
10
57
72
17
42
82
12
7
85
80
32
52
Some A are B
All A are B
No A are B
Some A are not B
All Bare A
Some B are A
No Bare A
Some B are not A
P
I
P
P
N
I
P
3
8
92
5
82
22
85
53
55
97
53
95
50
97
No A are B
All A are B
Some A are B
Some A are not B
AH Bare A
Some B are A
No Bare A
Some B are not A
I
I
N
I
I
N
N
3
7
77
7
20
67
65
2
5
80
25
35
90
87
Some A are not B
All A are B
Some A are B
No A are B
AH B are A
Some B are A
No Bare A
Some B are not A
I
P
P
P
P
P
P
2
87
5
7
70
12
88
3
97
48
38
95
53
95
All A are B
_
Note.
N = necessary; I = impossible; P = possible.
(Appendix B follows)
1510
Appendix B
Percentage Endorsement of Conclusions in Experimen 12 Under Necessity (N) and Possibility (P) Instructions
Cone.
Fig. order
Premises
A-C
All A are B
All B are C
C-A
All A are B
All B are C
A-C
All A are B
Some B are C
C-A
All A are B
Some B are C
A-C
Some A are B
All B are C
C-A
Some A are B
All B are C
A-C
Some A are B
Some B are C
C-A
Some A are B
Some B are C
A-C
No A are B
All B are C
C-A
No A are B
All B are C
A-C
No A are B
Some B are C
C-A
No A are B
Some B are C
A-C
Some A are not B

All B are C
C-A
Some A are not B

All B are C
A-C
Some A are not B

Some B are C
C-A
Some A are not B

Some B are C
Conclusion
Fig.
Cone.
order
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
N
N
I
I
P
N
I
P
P
P
P
P
P
P
P
P
P
N
I
P
P
N
I
P
P
P
P
P
P
P
P
P
P
P
P
P
I
P
P
N
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
73
27
10
23
77
47
80
53
27
13
83
63
27
40
30
83
33
87
33
90
20
80
23
90
20
87
37
90
20
83
33
100
50
93
47
87
50
93
20
17
83
73
23
27
80
57
27
60
77
80
20
50
63
67
27
87
37
93
10
87
43
90
20
93
43
80
33
83
53
83
A-C
All A are B
No B are C
C-A
All A are B
No B are C
A-C
All A are B
Some B are not C
C-A
All A are B
Some B are not C
A-C
Some A are B
No B are C
C-A
Some A are B
No B are C
A-C
Some A are B
Some B are not C
C-A
Some A are B
Some B are not C
A-C
NoAareB
No B are C
C-A
No A are B
No B are C
20
10
87
13
80
10
83
10
67
7
87
10
80
13
83
7
67
13
80
3
73
3
77
7
67
10
10
77
40
13
23
80
47
13
40
40
53
13
33
43
57
7
63
13
83
7
60
17
83
13
77
3
70
7
50
23
83
Premises
A-C
No A are B
Some B are not C
C-A
No A are B
Some B are not C
A-C
Some A are not B

No B are C
C-A
Some A are not B

No B are C
A-C
Some A are not B

Some B are not C
C-A
Some A are not B

Some B are not C
Conclusion
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No C are A
Some C are not A
I
I
N
N
I
I
N
N
P
P
P
P
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
3 10
7 13
83 100
27 57
7 10
7 13
83 87
40 50
10 10
67 80
13 37
90 90
6 27
43 87
23 23
77 93
0
13
30 63
37 73
83 93
3 20
50 47
40 77
60 70
13 37
67 97
7 53
77 93
7 30
57 93
17 47
77 100
17 30
27 40
60 77
40 73
17 37
30 40
53 80
40 70
13 43
40 83
40 67
50 87
17 30
40 60
40 67
63 77
3 23
63 80
27 53
67 90
10 37
47 67
27 47
73 73
7 20
77 93
17 40
87 93
0 23
57 87
17 37
63 93
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
1511
Cone.
Fig. order
2
A-C
C-A
A-C
C-A
Premises
All Bare A
All C are B
All Bare A
All C are B "
All Bare A
Some C are B
All Bare A
Some C are B
A-C
Some B are A
All C are B
C-A
Some B are A
All C are B
A-C
Some B are A
Some C are B
C-A
Some B are A
Some C are B
A-C
No B are A
All C are B
C-A
No Bare A
All C are B
A-C
No Bare A
Some C are B
C-A
No B are A
Some C are B
A-C
Some B are not A

All C are B =
C-A
Some B are not A

All C are B
A-C
Some B are not A

Some C are B
C-A
Some B are not A

Some C are B
Appendix B
(continued)
Conclusion
Fig.
Cone.
order
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
P
N
I
P
N
N
I
I
P
N
I
P
P
N
I
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
I
I
N
N
I
I
N
N
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
63 83
43 80
3
7
33 30
80 87
33 70
10 17
7 20
3 33
87 87
7 27
73 87
10 27
83 87
13 20
53 80
17 37
83 97
10 20
50 87
10 43
77 77
7 20
63 77
3 53
77 97
7 57
83 87
13 40
83 100
3 57
70 97
7 17
13 20
73 97
33 70
10 13
10 17
77 83
30 60
10 27
40 50
40 77
53 60
3 10
43 63
33 70
73 80
7 27
70 87
10 53
87 77
3 23
57 83
13 27
80 97
7 37
63 87
10 47
73 90
10 30
57 90
17 43
80 87
A-C
C-A
A-C
C-A
Premises
All Bare A
No C are B
All Bare A
No C are B
All Bare A
Some C are not B
All A are B
Some B are not C
A-C
Some B are A
No C are B
C-A
Some B are A
No C are B
A-C
Some B are A
Some C are not B
C-A
Some B are A
Some C are not B
A-C
No Bare A
No C are B
C-A
No B are A
No C are B
A-C
No B are A
Some C are not B
C-A
No B are A
Some C are not B
A-C
Some B are not A

No C are B
C-A
Some B are not A

No C are B
A-C
Some B are not A

No C are B
C-A
Some B are not A

Some C are not B
Conclusion
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
17 20
17 30
23 80
27 67
10 27
20 37
70 80
47 60
7 17
70 93
10 53
93 93
3 17
53 87
20 23
70 90
10 20
27 47
57 83
50 83
0 27
20 63
60 70
67 77
10 30
60 100
7 47
73 93
3 33
57 87
7 43
77 93
17 30
27 30
53 93
33 67
23 37
23 47
53 83
37 63
10 40
53 80
27 67
53 77
0 17
47 53
27 83
53 77
10 27
47 60
30 73
53 70
10 23
30 53
37 63
43 57
3 27
80 90
13 43
80 90
0 20
67 97
7 50
77
87
1512
Appendix B
Cone.
Fig. order
3
A-C
C-A
Premises
All A are B
All C are B
All A are B
All C are B
A-C
All A are B
Some C are B
C-A
All A.are B
Some C are B
A-C
Some A are B
All C are B
C-A
Some A are B
All C are B
A-C
Some A are B
Some C are B
C-A
Some A are B
Some C are B
A-C
NoAareB
All C are B
C-A
No A are B
All C are B
A-C
NoAareB
Some C are B
C-A
No A are B
Some C are B
A-C
Some A are not B

All C are B
C-A
Some A are not B

All C are B
A-C
Some A are not B

Some C are B
C-A
Some A are not B

Some C are B
(continued)
Conclusion
Fig.
Cone.
order
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
I
I
N
N
I
I
N
N
P
P
P
P
I
P
P
N
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
60
27
20
23
67
47
17
17
3
67
23
80
10
80
23
70
10
83
0
67
3
70
13
57
3
57
10
70
10
63
7
63
10
17
80
37
3
10
83
40
10
40
43
63
7
37
40
60
3
77
7
90
7
50
27
80
7
67
13
80
3
60
13
63
83
63
23
30
77
60
23
30
30
70
53
93
23
90
33
80
30
97
37
83
43
90
30
87
40
90
47
87
33
93
53
93
10
20
93
70
17
13
87
57
13
53
87
80
3
57
63
80
17
87
40
90
17
80
43
90
30
97
40
93
27
93
30
96
A-C
C-A
Premises
All A are B
No C are B
AllAareB
No C are B
A-C
AllAareB
Some C are not B
C-A
AllAareB
Some C are not B
A-C
Some A are B
No C are B
C-A
Some A are B
No C are B
A-C
Some A are B
Some C are not B
C-A
Some A are B
Some C are not B
A-C
NoAareB
No C are B
C-A
No A are B
No C are B
A-C
No A are B
Some C are not B
C-A
NoAareB
Some C are not B
A-C
Some A are not B

No C are B
C-A
Some A are not B

No C are B
A-C
Some A are not B

Some C are not B
C-A
Some A are not B

Some C are not B
Conclusion
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
NoCareA
Some C are not A
I
I
N
N
I
I
N
N
P
P
P
P
I
P
P
N
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
3
7
7 20
83 87
33 63
7
13
7
17
80 80
37 60
10 17
67 80
13 47
80 87
0
10
57 73
17 43
80 83
13 17
37 63
33 80
63 77
0
17
37 53
43 80
40 73
10 23
63 97
13 57
80 90
7 43
63 87
7 57
73 83
10 60
27 63
43 77
3 77
17 57
20 57
40 77
27 67
13 27
50 63
33 80
57 73
3 13
43 60
30 77
50 70
3 27
33 80
27 50
53 83
7 30
47 63
30 70
67 67
13 40
53 93
13 57
73 100
3 43
57 100
13 33
60 87
1513
Appendix B
Fig.
Cone.
order
A-C
All Bare A
All B are C
C-A
All Bare A_
All B are C
A-C
All Bare A
Some B are C
C-A
All B are A
Some B are C
A-C
Some B are A
All B are C
C-A
Some B are A
All B are C
A-C
Some B are A
Some B are C
Premises
C-A
Some B are A
Some B are C
A-C
No Bare A
All B are C
C-A
No Bare A
All B are C
A-C
No B are A
Some B are C
C-A
No B are A
Some B are C
A-C
Some B are not A

All B are C
C-A
Some B are not A

All B are C
A-C
Some B are not A

Some B are C
C-A
Some B are not A

Some B are C
Note.
Conclusion
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All C are A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
P
N
I
P
P
N
I
P
P
N
I
P
P
N
I
P
P
N
I
P
P
N
I
P
P
P
P
P
P
P
P
P
P
P
P
P
I
P
P
N
P
P
P
P
I
P
P
N
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
{continued)
Fig.
Cone.
order
70
27
10
10
60
47
10
17
13
90
7
70
10
83
10
60
7
77
7
73
13
73
10
67
7
57
17
60
20
57
17
57
0
17
63
53
10
27
63
33
7
37
37
50
3
40
27
57
13
67
17
87
7
60
13
87
3
67
3
87
3
53
13
67
87
70
27
43
87
70
27
37
37
93
27
90
43
90
27
37
33
87
23
87
53
87
30
77
47
90
60
90
37
90
53
83
17
20
83
60
7
17
87
60
13
47
83
80
10
57
77
77
23
83
50
87
40
80
23
90
40
83
40
93
27
80
60
83
A-C
All Bare A
No B are C
C-A
All Bare A
No B are C
A-C
All Bare A
Some B are not C
C-A
All A are B
Some C are not B
A-C
Some B are A
No B are C
C-A
Some B are A
No B are C
A-C
Some B are A
Some B are not C
Premises
C-A
Some B are A
Some B are not C
A-C
No B are A
No B are C
C-A
No Bare A
No B are C
A-C
No Bare A
Some B are not C
C-A
No Bare A
Some B are not C
A-C
Some B are not A

No B are C
C-A
Some B are not A

No B are C
A-C
Some B are not A

No B are C
C-A
Some B are not A

Some B are not C
Conclusion
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No C are A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
No A are C
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
AllAareC
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
All A are C
Some A are C
NoAareC
Some A are not C
All Care A
Some C are A
No Care A
Some C are not A
I
P
P
N
P
P
P
P
I
P
P
N
P
P
P
P
I
P
P
N
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
7
20
63
30
7
20
73
37
7
63
13
87
0
53
17
83
7
30
30
70
7
30
47
53
20
17
93
70
13
27
90
67
30
87
40
83
20
80
27
87
23
60
80
83
20
60
77
83
33
87
43
90
33
87
47
93
60
50
80
77
53
57
80
73
43
70
67
83
23
57
63
77
37
60
67
70
23
57
70
83
40
93
60
97
43
93
53
90
70
7
80
0
57
13
67
20
17
47
23
23
30
43
27
13
40
33
53
13
37
43
50
13
43
33
50
13
37
40
47
13
57
13
53
3
53
17
60
Fig. = figure; Cone. = conclusion; L = logic (under which N = necessary; I = impossible; P = possible).
Received May 12, 1998

Revision received April 27, 1999
Accepted April 30,1999

Reasoning About Necessity and Possibility: A Test of The Mental Model Theory of Deduction

Uploaded by

Copyright:

Available Formats

Reasoning About Necessity and Possibility: A Test of The Mental Model Theory of Deduction

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Reasoning About Necessity and Possibility: A Test of The Mental Model Theory of Deduction

Uploaded by

Copyright:

Available Formats

Journal of Experimental Psychology:

Learning, Memory, and Cognition

Copyright 1999 by the American Psychological Association, Inc.

Reasoning About Necessity and Possibility: A Test of the Mental Model

Traditional applications of logic concern the validity of

Jonathan St. B. T. Evans, Simon J. Handley, and Catherine N. J.

in almost all studies reported to date: They ask participants

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

large-scale study of syllogistic reasoning using production

Possible and Necessary Inference

Argument 1 is a modus ponens inference in which the

NECESSITY AND POSSIBILITY

people are asked instead to decide whether it is possible that

inference task (see Newstead & Griggs, 1983). Experiment 2

From this analysis we derive and test in this article two

Experiment 1: Immediate Inference

In Experiment 1, we looked only at the immediate

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

Note that logically any conclusion that is necessary is also

Conclusions 2 and 6 are necessary; 4 and 7 are possible

problems with a layout as in the example that follows:

Results and Discussion

NECESSITY AND POSSIBILITY

those that we have space to describe here. Consequently, a

Figure 1. Percentage of conclusions accepted in Experiment 1

because these conclusions differ in their correct answer

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

Experiment 2: Syllogistic Reasoning

In an evaluation task, as used in the present study, each of

mental logic tradition. A detailed account of syllogistic

The above example would be given to the Necessity-AC group.

NECESSITY AND POSSIBILITY

Results and Discussion

Figure 2. Percentage of conclusions accepted in Experiment 2

1) when one considers the additional complexity of the two

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

of the valid conclusions on multiple-model problems. Under

Only 3% of participants endorsed this syllogism under

Experiment 3: Possible Strong Versus Possible

Another possible problem with the procedure of Experiment 2

NECESSITY AND POSSIBILITY

Experiment 3 involves presenting selected syllogisms

dorsed Possible problems in each figure. Similarly, Possible

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

tions on screen depending on the experimental condition to which

Results and Discussion

Figure 3. Percentage of conclusions accepted in Experiment 2

evaluated by pressing the mouse in a YES or NO box. The 64

The frequency of acceptance of different logical types

Figure 4. Percentage of conclusions accepted in Experiment 2

NECESSITY AND POSSIBILITY

expected a very highly significant effect of problem type,

results also confirm our three main predictions based on the

EVANS, HANDLEY, HARPER, AND JOHNSON-LAIRD

very general level of our other predictions. However, the

where . denotes negation and a, b, and c are tokens

Hence, the conclusion is possible but not necessary. An

This model does not support the given conclusion, but

To grasp that the given conclusion is at least possible,