Full text of DOI:10.3758/BF03195272

Memory & Cognition
002, 30 (1), 129-137
2
Generating alternatives:
A key component in human reasoning?
STEPHEN E. NEWSTEAD
University of Plymouth, Plymouth, England
VALERIE A. THOMPSON
University of Saskatchewan, Saskatoon, Saskatchewan, Canada
and
SIMON J. HANDLEY
University of Plymouth, Plymouth, England
According to mental models theory, a key aspect of deductive reasoning is the production of alterna-
tive models that can falsifyprovisional conclusions. In the present paper, the possibility is investigated
that there are individual differences in the ability to produce alternative models. The results indicate
that some people do not proceed beyond the first model when they reason with syllogismsbut that oth-
ers do. Furthermore, the ability to generate alternativescan be independently measured by asking par-
ticipants to generate different representations of pairs of premises. These findings support the predic-
tions of mental models theory and also indicate the potential importance of alternativesgeneration as
a measure of individual differences in processing style.
One of the most influential current theories in the field the inclusion of a falsification parameter did little to im-
of human reasoning is that people construct mental mod- prove the fit of their models to the syllogistic reasoning
els (Johnson-Laird & Byrne, 1991). According to this the- data.
ory, when faced with a reasoning problem, peoplewill first
Empirical evidenceis also beginningto accumulatethat
construct a representation of the premises using image- reasoners do not always search exhaustivelyfor alternative
like tokens. They then see what conclusion(if any) follows models; indeed, there is recent evidence that, in syllogistic
from this and proceed to search for alternativerepresenta- reasoningat least,peopleseldomgobeyondtheinitialmodel
tions (or models) in which that conclusionis not true. If no that they construct (Evans, Handley, Harper, & Johnson-
falsifying model can be found, the conclusion is accepted; Laird, 1999; Newstead, Handley, & Buck, 1999). Evans
otherwise, it is rejected, and the search for an alternative et al. (1999) found that people tended to accept conclu-
conclusioncontinues.The search for falsifyingmodelslies sions that were consistent with the preferred initial model
at the heart of mental models theory.
that they constructed and to reject conclusions that were
Recently, a debate has been taking place as to whether inconsistent. Errors occurred when the conclusion was
people do spontaneously search for alternative models consistent with the initial model but where there existed
(
Klauer, Musch, & Naumer, 2000; Polk & Newell, 1995). alternative,falsifying models. Participantsseemed unwill-
These theories are very similar in many respects to men- ingorunabletoconstructthesealternativemodels.TheNew-
tal models but do not assume that falsification takes place. stead et al. (1999) study involvedparticipants’constructing
Instead, it is explicitly assumed that reasoners attempt to diagrams of different colored shapes that were consistent
construct only a single model or representation of the with the premises of given syllogisms, under the assump-
premises and base their judgments of validity on that one tion that these would reflect the models constructed. The
representation.Indeed, Polk and Newell demonstratedthat results indicated that people tended to produce the same
representation over and over again (though in different
forms).
This research was funded by an operating grant from the Natural Sci-
This is not to say that people are unable to search for al-
ternative models: All the evidence suggests that they can.
For example, if the conclusion is unbelievable,people are
ences and EngineeringResearch Councilof Canada to V. A .T. The order
of authorship for the first and second authors was determined randomly.
Correspondence and reprint requests may be addressed to S. Newstead,
Department of Psychology, University of Plymouth, Drake Circus, Ply- likely to try to construct an alternative model (Newstead,
mouth PL4 8AA, England (e-mail: snewstead@plymouth.ac.uk) or to
Pollard, Evans, & Allen, 1992). Search for alternativescan
V. Thompson, Department of Psychology, 9 Campus Drive, University
also be inducedby theinstructionsgiven(Evans, Newstead,
of Saskatchewan, Saskatoon, SK, S7N 5A5 Canada (e-mail: valerie.
Allen, & Pollard, 1994), or by reducing the cognitive load
thompson@usask.ca).
—
Accepted by previous editorial team (Roberts, 2000). Furthermore, the ease with which alterna-
1
29
Copyright 2002 Psychonomic Society, Inc.

1
30
NEWSTEAD, THOMPSON, AND HANDLEY
tives can be brought to mind appears to be a critical fac- generationwas the only measure that served to predict be-
tor. There is a wealth of evidence in conditionalreasoning lievability effects: A range of other measures, including
that the availability of alternative antecedents influences need for cognition,open-mindedness,and intelligence,all
performance in causal reasoning (Cummins, 1995), the failed to serve as predictors. Newstead, Handley, Harley,
Wason selection task (Handley, Feeney, & Harper, in Wright, and Farrelly (2002) found that alternativesgener-
press), and everyday conditionalreasoning (Byrne, 1989; ation also predicted performance on abstract versions of
Thompson, 1995, 2000). Similarly, the availability of al- the Wason selection task and on certain kinds of condi-
ternative routes to the consequent also appears to be im- tional reasoning tasks.
portant (Cummins, 1995; Thompson, 1995, 2000).
Our main hypothesisis that reasoners vary in their abil-
There is thus ample evidence that task variables can af- ity and inclinationto construct alternativerepresentations
fect the production of alternative models. In addition, and that these individualdifferences are systematicallyre-
there may also be individualdifferences in the willingness lated to performance on syllogistic reasoning tasks. The
to search for alternatives, and the ability to construct al- basic prediction is that people who generate more alter-
ternative representations might be systematically linked natives will be more likely to solve certain types of syllo-
toreasoningperformance.Thisisa centralthemeofthepres- gisms correctly.
ent paper. The possibility of such individual differences
has long been acknowledged by mental models theorists alternatives and syllogistic reasoning is open to an alter-
e.g., Johnson-Laird, 1983) and has recently been demon- nativeexplanation,since both might be mediatedby a third
strated empirically by Bucciarelli and Johnson-Laird variable such as need for cognition(e.g., Klaczynski, Fauth,
1999).Theseauthorsused a taskin which theirparticipants & Swanger, 1998) or general intelligence(e.g., Stanovich
were asked to dress stick figures. For example, they might & West, 1998). One way to rule out this possibilityis to in-
However, a correlation between the ability to search for
(
(
be given the syllogism
clude measures of these variables in the experimental de-
sign. By using this strategy, both Torrens et al. (1999) and
Newstead et al. (2002) found that their measure of alterna-
tives generation predicted deductive reasoning perfor-
Some of the chefs are musicians
All of the painters are musicians
and then were asked to dress the stick figures with chefs’ mance even after accountingfor the effects of intelligence
hats, guitars,and artists’palettesin such a way that made the and the need for cognition.
conclusionSome ofthechefsare paintersnotpossible.Their
An alternativeapproach,and the one adoptedinthe pres-
participants were able to do this, which once again illus- ent study, is to use materials that on a priori groundsshould
trates that people are capable of falsifying conclusions if discriminate between reasoners who search for alterna-
the task demands it. More important from the present per- tives and those who do not. Thus, we used problems whose
spective were the large individual differences in perfor- solutionsrequired a search for alternatives. On such prob-
mance. Their best participant produced 95% correct re- lems, a correlation between the search for alternativesand
sponses,theirworst just25%(the participantswere students reasoning performance would providestrong evidencethat
at Princeton University).
Further evidence for the existence of individual differ- reasoning ability that is mediating performance.
ences in alternatives generation stems from the work of
In our experiment, we presentedreasoners with pairs of
it is the search for alternativesitself and not a more global
Markovits (e.g., 1984). He measured participants’ ability premises and a conclusion to evaluate. The conclusion
to think of alternatives by simply asking them the follow- was either weak or strong (Evans et al., 1999), a distinc-
ing question: When David has homework to do, he gets tion that is perhaps best illustratedby an example. The syl-
into a bad mood. I saw David after school today and he logism No A are B; All B are C is a multiple model syllo-
was in a bad mood. Can you imaginewhat wouldhave put gism, having,accordingto Johnson-LairdandByrne (1991)
David in a bad mood? Those participants who indicated the following models:
possibilitiesother than homework were superior in condi-
tionalreasoningperformance to those who mentionedjust
the one possibility.
A recentstudybyTorrens,Thompson,andCramer(1999)
produced a similar picture. In a study of conditional syl-
logistic reasoning,participantswere first asked to draw as
(
1)
(2)
(3)
[
[
a]
a]
[a]
[a]
[a]
[a]
c
c
c
[
[
[b] c]
[b] c]
[[b] c]
[[b] c]
[[b] c]
[[b] c]
many diagrams as possible to represent the premises of In this notation, square brackets indicate that the entity is
two quantified syllogisms; the number of different dia- exhaustivelyrepresented. Thus, in model (1), there are no
grams producedwas the alternativesgenerationscore. This as that are either bs or cs; bs are exhaustively represented
score correlated positively with logical performance on with respect to cs; and there are no other cs that are either
the conditionalsyllogisms and negatively with the effects as or bs (though there may be other cs that are neither as
of conclusionbelievability. In otherwords, the participants nor bs). Now, if this is the first and onlymodelconstructed,
who produced more alternativerepresentationswere more participants would probably indicate that the conclusion
likely to give logically correct answers and less likely to No A are C (or that No C are A) necessarily followed, and
respond on the basis of belief. Furthermore, alternatives indeed these two conclusionsare very common (see Evans

GENERATING ALTERNATIVES AND HUMAN REASONING
131
et al., 1999). If, on the other hand, participants construct came to mind. These tasks were both presumed to indi-
only model (2), they would reject these two conclusions. cate the initialmodel constructed, and we were lookingfor
A strong conclusion is one that is consistent with the multiple model syllogisms that produced the same initial
first model constructed,whereas a weak conclusionis one model both across participants and across tasks.
that is not consistent with the initial model but is consis-
In addition to providing information about the initial
tent with an alternative representation. We believe that models, this study had a second aim in that it allowed us to
many reasoners will accept a conclusion that is consistent examinethe relationshipbetween thetwo tasks. Since both
with the initial model, but reject a conclusion that is in- were designed to determine the initial model constructed,
consistentwith this model (Polk & Newell, 1995). In other and since mental models theory assumes that initial con-
words, they will accept strong conclusionsand reject weak clusionsare derived from initialmodels, thistheory would
conclusions.However, other reasoners will not be content presumably predict that the two tasks should correlate.
with the outcome of this initial decision process and will
search for alternative representations. Those who do so
should be more likely, for example, to correctly reject in-
valid strong conclusions(i.e., conclusionsthat are consis-
tent with the initial model but not with other possible rep- credit. The mean age of the students was approximately 19 years,
resentations). We hypothesize that individual differences and approximately 60% were female.
in reasoning style will mediate whether or not a person
Method
Participants. The 111 participants were introductory psychology
students at the University of Saskatchewan who participated for course
Materials. According to Johnson-Laird and Byrne (1991) there
are 17 valid multiple model syllogisms, all of which were used in the
will search for alternative representations and hence in-
present study. Two booklets were prepared containing these syllo-
fluence performanceon specifictypesof conclusions.This
gisms in a single random order.
The first booklet contained instructions that asked the participants
is a very strong test of our hypothesis. In contrast, there
are othercases in which the ability to constructalternative to draw the first diagram that came to mind when they combined the
representations should not be correlated with performance two premises. An example was given: All of the jeans are blue and
all of the jeans are denim (a multiple model syllogism). The instruc-
tions explained that the task required them to draw a diagram show-
sincethe correct answer will emerge from the initial model
constructed (Handley, Dennis, Evans, & Capon, 2000).
ing how jeans, blue things, and denim are related to each other. A
A full set of predictionsis presented in the introduction
sample diagram was given in which the first premise was drawn as
to Experiment2. Experiment 1 was designed to determine
which is the preferred initial model for the syllogisms we
a circle representing jeans; this was contained within (i.e., was a sub-
set of) a larger circle representing blue things. A further diagram
were using.Thesefindingsare interestingintheirown right indicated that the second premise could be combined with this dia-
and are thus reported in some detail.
gram by drawing a circle representing denim things so that it con-
tained the circle representing jeans and overlapped with the circle
representing blue things. It was stressed that although there may be
other ways of representing the premises, it was important to draw
the first diagram that came to mind.
The instructions in the second booklet asked the participants to
write down the first conclusion that came to mind. The same exam-
ple premise pair was given, and the instructions asked the partici-
EXPERIMENT 1
In order to determine the initial model constructed, we
presented the participants with syllogisms; in one condi-
tion, we asked them to draw diagrams to represent the re-
lationshipsand,inthe other, to give the first conclusionthat pants to think of a conclusion that linked the two nonrepeated items
Figure 1. Possible Euler circle relationships between two sets A and C.

1
32
NEWSTEAD, THOMPSON, AND HANDLEY
in the premises. The sample conclusion Some of the blue things are duced in the diagram drawing task were assumed to indi-
denim was given. It was explained that all answers needed to contain cate the first model produced by the participants and, as
one of the standard quantifiers (all, some, some. . . not, and no) and
can be seen in Table 1, there was considerableconsistency
that the important thing was to give the first conclusion that came to
in the preferences shown. One would expect that the con-
mind.
clusions given in the second condition would reflect this,
Each booklet contained the 17 pairs of premises in random order.
since they are presumablydrawn from this first model. Be-
cause this is a between-subjectsdesign, there is no way of
examining this relationship directly, but it can be investi-
gated indirectly. A fairly straightforward prediction is that
the conclusion generated for a syllogism will, by and
large,be consistentwiththepreferred modelinthediagrams
task. For those problems where the exclusion diagram is
preferred, one would expect conclusionsto be either No A
The terms were all names of occupations or hobbies. Here is an ex-
ample of one of the syllogisms used:
All of the piano players are doctors
and
None of the piano players are golfers.
Procedure. The participants were run in small groups, were ran-
domly allocated to conditions (i.e., diagram or conclusion booklet),
and were allowed as much time as they wanted to complete the task.
There were 54 participants in the diagrams condition and 57 in the are C or No C are A. For those syllogisms where the over-
conclusion-production task.
lap relationship is preferred, one would expect some and
some. . . not conclusions in either direction.
Results and Discussion
In order to examine this, the prediction was tested that
On the diagrams task, the participantsproduced a wide the participants would tend to produce more conclusions
variety of different diagrams that we inductively placed that were consistent with the preferred model than would
into 42 categories covering all the depicted relationships be expectedbychance.Fortheexclusionrelationship,which
between the three sets (the size or shape of the diagrams was the most common in the diagrams task, only two of
used to depict the sets was deemed to be irrelevant). The the eight conclusions (those using no) were consistent
4
2 diagramsproducedcanbe collapsedandclassifiedsolely with this; hence chance expectancyis .25. When the over-
according to the set relationship between the two end lap relationship was preferred, half of the conclusions
terms, A and C. The are just five possible relationships: (those using some and some. . . not) were consistent with
identity, A as a subset of C, C as a subset of A, overlap, and this; hence, the chance probabilityis .5. The preferred di-
exclusion (see Figure 1). The frequency with which these agram and the conclusionsconsistentwith it are shown in
diagrams were given for each syllogismis presented in the bold in Table 1. It can be seen that for 14 of the 17 syllo-
left-hand part of Table 1. The conclusions given by the gisms, conclusions consistent with the preferred model
participants in the conclusion-production task are pre- occurred more often than by chance, a difference that is
sented on the right-hand side of Table 1.
significant on a sign test ( p < .01).
Although different participantswere involvedin the di-
Despite this general tendency, however, there are clearly
agram and conclusion-productiontasks, one would expect some syllogisms where the conclusions were not consis-
some relationship between the tasks. The diagrams pro- tent with the preferred diagram. The two most obviousex-
Table 1
Diagrams and Conclusions Generated in Experiment 1
Diagram
Identity
Conclusion
Syllogism Exclusion
Overlap
AsubsetC
CsubsetA
Aac
Aca
Iac
Ica
Oac
Oca
Eac
Eca
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
Aab,Ocb
Iab,Ebc
Iab,Ecb
Eab,Abc
Eab,Ibc
Eab,Icb
Oab,Acb
Aba,Abc
Aba,Ebc
Aba,Obc
Aba,Ecb
Iba,Ebc
Iba,Ecb
Eba,Abc
Eba,Ibc
Eba,Icb
Oba,Abc
73
68
76
81
75
73
80
20
73
30
76
78
75
76
72
74
26
20
30
22
17
21
21
18
58
20
59
24
18
21
18
19
23
67
0
0
0
0
0
0
0
14
0
0
0
0
0
2
0
0
0
2
2
2
0
0
2
0
6
6
2
0
4
4
0
0
0
0
4
0
0
2
4
4
2
2
0
9
0
0
0
4
9
2
2
0
0
0
0
2
0
0
73
6
0
2
0
2
0
0
2
4
2
0
0
0
2
0
4
0
0
2
4
0
0
7
0
0
2
32
22
14
10
11
7
44
22*
6
30
7
15
11
7
23
0
7
2
7
14
11
2*
0
9
2
2
2
0
4
6
13
18
37*
30*
4
9
2
20*
0
2
2*
7*
25*
13
0
0
2
2
0
2
28
33
60
46
34
2
2
68
2
43
35
13
43
21
28
4
2
13
14
21
17
18
2
0
11
0
33
17
43
33
28
15
2
22*
0
9*
54*
7*
31*
21*
2
13
4
6
1
1
1
1
1
1
1
1
9
9*
26*
36*
42*
9
9
27
Notes—The logically correct conclusion is indicated by an asterisk. Entries in bold indicate the most popular diagram in the diagram construction
task and the conclusions consistent with this in the conclusion task. A, all; E, no; I, some; O, some. . . not. Syllogistic figure is indicated by lower
case letters, for example Aab,Ocb indicates All A are B, Some C are not B. The figures indicate the percentage of participants selecting each dia-
gram or conclusion.

GENERATING ALTERNATIVES AND HUMAN REASONING
133
amplesare Syllogisms1 and 7 in Table 1. On bothof these, ble in the initial model but possible in other models that
the exclusion relationship was the preferred representa- might be constructed). In this case, one would expect that
tion, but conclusions involving no were given by just 4% those who generate alternatives should be more accurate
of the participants!The other major discrepancy is on Syl- than those who do not, since the former will be more likely
logism 8, where the preferred diagram was overlap but to find the model in which the conclusion is possible.
only 24% of the participants gave either some or some. . .
not conclusions. The most popular conclusion was All A tives will show a different pattern of respondingthan those
are C, given by 73% of people.
who do not.Under necessityinstructions,those who search
In sum, we predicted that those who search for alterna-
The degree of consistency between the diagram and for alternativesshould be more likely to reject strong con-
conclusion-production tasks suggests that there are com- clusions; and under possibility instructions, they should
mon processes mediating performance, and mental mod- be more likely to accept weak conclusions. Performance
els provide a plausible account of what these processes on necessary and impossible conclusions was not pre-
might be. However, there were sufficient discrepancies to dicted to differ among reasoners.
warrant caution in this conclusion. These discrepancies
Although the major focus was on the measure of alter-
might well reflect task-specific demand characteristics.In natives generation used by Torrens et al. (1999), we also
the diagram task, for example, reasoners have a clear pref- used other measures that may be related to the generation
erence for exclusion relations, which might be because of alternatives. Many tests of creativity ask people to gen-
these are conceptually the easiest to represent diagram- erate alternative responses, and we were interested to
matically. In contrast, a different set of biases, such as at- know whether one such test, the uses of objects test, corre-
mosphere bias, may operate in the conclusion-production lated with eitherthe alternativesgenerationtask or syllogis-
task. It seems reasonable to concludethat performance on tic reasoning performance. In addition, we also explored
thesetasksresultsfrom a combinationofrational(i.e.,model the possibilityof any correlationsbetween these measures
based) and heuristic processes.
and a widely used measure of thinking style, the Rational
Althoughthe actualresponsesproducedin thisstudyare Experiential Inventory (REI). This measure (and one of
of interest in theirown right,our main purpose in thisstudy its precursors, need for cognition) has been claimed by
was to find syllogismsthat consistentlyproduced thesame some (e.g., Klaczynski et al., 1998) to be related to rea-
initialmodelwiththispopulationofparticipants.We selected soning performance.
eight such syllogisms in Experiment 2, the aim of which
was to investigate differences in performance as a func-
tion of individual differences in alternatives generation.
Method
Participants. There were 151 participants in this study, 64 males
and 87 females, whose mean age was 19.4 years. They were all in-
troductory psychology students at the University of Saskatchewan
who took part for course credit.
EXPERIMENT 2
Materials. We used the REI as described by Pacini and Epstein
1999). This is a 40-item self-report inventory that measures two
main dimensions, rationality (the extent to which a person enjoys
and engages in problem solving) and experientiality (the extent to
which a person relies on intuition and past experience).
The central purpose of this experiment was to examine
individual differences in performance on multiple model
syllogismsthatconsistentlyproducethesame initialmodel.
The predictionsare a little complex. In one condition, we
used necessityinstructions,under which only conclusions
(
The uses of objects test was employed. This asked the participants
that necessarilyfollow should logicallybe endorsed. If the to list as many uses as they could think of in 5 min for a brick and a
paper clip (adapted from Hudson, 1966). The test was scored in terms
target conclusionis necessary (i.e., true in all models), one
of the total number of uses produced after the elimination of any ob-
would expect it to be accepted by all participants; there
vious repetitions.
should be no difference dependenton the search for alter-
The alternatives generation task was modeled closely on that of
nativemodelssince such a search is unnecessary. Butif the
Torrens et al. (1999). Four pairs of quantified premises were presented
conclusionis a strong one (i.e., true in the first model con- with neutral content. These were all multiple model syllogisms,
structed but false in other models), one would expect such though the content was different than that in the syllogistic reason-
ing task. The instructions were similar to those in the diagrams con-
differences.The participantswho constructonlyone model
dition of Experiment 1. However, after the first overlap diagram had
should accept strong conclusions as necessarily true, but
been presented, it was pointed out that other alternatives were pos-
those who construct alternative models should be more
likely to find a disconfirming model. Hence, under neces-
sible, and a diagram presenting an overall subset relationship was
given. The instructions indicated that the participants should draw as
sity instructions,onewouldpredicta correlationbetween al- many diagrams as they could until they could think of no more.
ternativesgenerationand accuracy on strong conclusions. Their responses were scored in terms of the number of different but
correct diagrams produced.
In a second condition, we used possibility instructions,
In the main syllogism task, we used eight syllogisms selected on
under which any conclusion that could be true should be
the basis of the results of Experiment 1. We chose eight problems that
accepted. If the conclusion is impossible, it should be re-
reliably elicited the same first model as indicated by the findings in
jectedby all participantsirrespectiveof whether they try to
generate alternativemodels. However, thingsare different
each of the two conditions. The problems used were Syllogisms 4,
, 6, 9, 11, 12, 13, and 15 in Table 1. These were presented in a book-
5
if the conclusion is a weak one (i.e., one that is not possi- let in two groups of four. In one group of four, necessity instructions

1
34
NEWSTEAD, THOMPSON, AND HANDLEY
Table 2
Under necessity instructions, two syllogisms had conclusions that
were necessarily true and two had conclusions that were invalid but
consistent with the first model constructed (i.e., strong). Under pos-
sibility instructions, two problems had impossible conclusions and
two had conclusions that were possible but not consistent with the
first model constructed (i.e., weak). Each syllogism had just one
conclusion to be evaluated, with the pairing of conclusion type with
syllogism being systematically varied.
Conclusions Used in Experiment 2
Syllogism
Number
Necessary
Strong
Weak
Impossible
4
5
6
9
1
2
3
5
Oca
Oca
Oca
Oac
Oac
Oac
Oac
Oca
Eac
Eac
Eac
Eac
Eac
Eac
Eac
Eac
Iac
Iac
Ica
Iac
Iac
Iac
Iac
Iac
Aca
Aca
Aca
Aac
Aac
Aac
Aac
Aca
1
1
1
1
Results
Syllogisms task. Under the necessity instructions, the
participants accepted more necessary (66%) than strong
Notes—A, all; E, no; I, some; O, some. . . not. Direction of conclusion
either A to C or C to A) is indicated in lower case letters.
(
52%) conclusions, a difference that is statisticallysignif-
(
icant [t(152) = 2.88, p < .01]. However, the proportion of
strong conclusionsaccepted was very high and consistent
with the suggestion that many participants did not look
beyond the first model that came to mind. Under possibil-
ity instructions, more weak (40%) than impossible (10%)
conclusions were accepted, and this difference was signif-
were given indicating that a conclusion should be accepted only if it
was necessarily true. The other set of four syllogisms used possibil-
ity instructions, explaining that a conclusion should be accepted if it
was possible and rejected if it was impossible.
Each syllogism had four associated conclusions, one which was icant [t(152) = 8.39, p < .001]. Again, the relatively small
necessarily true, one which was strong (i.e., invalid but consistent number of weak conclusions that was accepted is consis-
with the initial model as determined in Experiment 1), one which was
tent with the suggestionthat some of the participantscon-
weak (i.e., possible but not consistent with the initial model), and one
structed a first model in which the conclusion was not
which was impossible. These conclusions are presented in Table 2.
The syllogisms in both conditions appeared as in the following
example:
possible and did not proceed to generate alternative mod-
els. Perhaps the most striking illustrationof this is that rea-
soners were more likely to (incorrectly) accept a strong
conclusion as necessarily following from the premises
than they were to (correctly) accept a weak conclusion as
All the chefs are bankers
None of the chefs are workaholics
Therefore, some of the bankers are workaholics
possiblyfollowingfrom thepremises[
t(152)= 2.48,p < .01].
Yes
N o
Individual differences measures. The correlationsbe-
tween the various measures are shown in Table 3. Of cen-
tral interest to the present study are the correlations be-
The task was to circle the correct answer (which varied, of course,
with the instructions given). The content was chosen to be neutral,
involving occupations, hobbies, and general descriptors. The eight
syllogisms all had different content, and to avoid any materials ef- tweenthe individualdifferencesmeasures and performance
fects, four different random combinations of material with syllo- on the syllogism task. The alternatives generation mea-
gisms were constructed.
Procedure. The participants were run in a single large group. The
order of presentation of the tasks was as follows: uses of objects,
syllogisms, REI, and alternatives generation. The participants were
sure correlated negatively with acceptance of strong con-
clusions and positivelywith weak conclusions.This is ex-
actly what was predicted and suggeststhat the participants
who generated alternativeswere more likely to search be-
presented with syllogisms under both necessity and possibility in-
structions, with the necessity condition being presented first to half yond their initial model. With strong conclusions, they
of the participants and second to the other half. Four different syllo- were likely to find instances in which the conclusion was
gisms were presented in each condition. Under necessity instruc-
invalidated (i.e., was not necessary), whereas with weak
tions, two necessary and two strong conclusions were given, and
conclusions,they found instancesthat showed the conclu-
under possibility instructions, two impossible and two weak conclu-
sion to be valid (i.e., where the conclusion was possible).
sions were presented. The order in which the conclusion types were
In contrast, alternativesgeneration was not expected to
presented was counterbalanced across participants.
correlate with the rejection of impossible conclusions.As
To summarize, each participant received eight syllogisms, four
under necessity instructions and four under possibility instructions. is indicated in Table 2, the correlation was close to zero,
Table 3
Correlations Between Measures Used in Experiment 2
Uses of Objects Rationality
Experientiality Alternatives Generation Strong
Impossible
Necessary
Weak
Uses of objects
Rationality
Experientiality
Alternatives
generation
Strong
Impossible
Necessary
Weak
–
.27*
–
.13
.22*
–
.17*
.21*
2.03
.12
.02
.11
2.08
2.03
.18*
2.01
.02
.14
.03
.10
.07
–
2.17*
2.06
2.02
–
.21*
2.26**
2.18*
–
.20*
2.26**
.04
–
.14*
Note—The measures of syllogism performance are based on the number of conclusions of each type accepted. *p < .05. **p < .01.

GENERATING ALTERNATIVES AND HUMAN REASONING
135
despite the fact that we had sufficient power to find a cor-
Our data are clearly consistent with the claim that peo-
relation (we had 80% power to detect a correlation of .2). ple are capable of constructingalternativerepresentations
Finally, we also observed that alternativesgeneration was of syllogisms beyond the initial model constructed but do
positivelycorrelated with correct acceptanceof necessary not always do so. Necessary conclusions were more com-
conclusions.
monly accepted than were conclusions that were consis-
The other individual differences measures were much tent with the preferred model but not logically necessary.
weaker predictorsof performance on syllogisms. None of However, the latter were accepted by about half the par-
them predicted performance on strong and weak conclu- ticipants, suggesting that they may not have gone beyond
sions: The only significantcorrelation was between expe- the first model. Similarly, impossibleconclusionswere re-
rientiality and impossible conclusions. The two measures jected more often than were conclusions that were possi-
designed to test the ability to producealternatives(uses of ble but not consistent with the initial model, but the latter
objects and the alternatives generation task) correlated were still rejected quitefrequently. It would appear that on
with each other and with rationality. It would appear that a significant number of trials, the participantsconstructed
there is some consistency among our various measures of just a single model and did not attempt to falsify it.
the capacityto producealternatives,suggestingthat they re-
A possible alternative explanation for our results is in
flect some common dimension,perhaps theabilityto think terms of intellectualability.Ability has been found to cor-
flexibly.The fact that only the alternativesgenerationtask relate with logicalperformance on syllogisms(e.g., Stano-
predicted performance on syllogisms suggests that it may vich & West, 1998; Torrens et al., 1999) and it also corre-
also measure a task-specificabilityto generate alternative lates with alternatives generation (Newstead et al., 2002;
models in deductive reasoning tasks.
Torrens et al., 1999). Perhaps, then, our results are little
Correlationsbetween measures onthe syllogismtask are more than a demonstration of the importance of ability in
also of interest. One would expect a negative correlation syllogisticreasoning.There are, however, a number of rea-
between acceptanceof strong conclusionsand weak ones, sons for believing that alternatives generation measures a
since the former suggests a failure to search for alterna- process over and above general ability. Most importantly,
tives, whereas the latter indicates a tendency to conduct we chose problems for which a correct answer could only
such a search. The predicted negative correlation can be have been reached by searching for alternatives. That is,
seen in Table 2. Acceptanceof necessary conclusionscor- the participants could only reject the strong conclusions
related with all of the other conclusions. The correlation and accept the weak conclusions if they searched beyond
was positive with the acceptance of weak conclusionsand their initial representation of the premises. Thus, it seems
was negative with the acceptance of both strong and im- unlikely that correlations with these measures can be at-
possible conclusions.
tributedsolelyto a general logicalability insteadof a more
specific ability to construct alternative representations.
In addition,in ourown data, we didnot find a correlation
Discussion
The presentfindingssupportthemain predictions.There between alternatives generation and performance on im-
is evidence that individual differences in the production possible conclusionsas might have been expected if alter-
of alternativerepresentations, as measured by the alterna- natives generation is closely related to intellectualability.
tives generation task, correlate in predictable ways with Furthermore, other studies have demonstrated that alter-
performance on weak and strong conclusions. In general, natives generation is in many ways a better predictor of
the more likely a person is to generate alternatives, the some aspects of reasoning performance than are standard
more likely they are to find the relevant model to discon- measures of intellectual ability (Newstead et al., 2002;
firm the conclusion suggested by the initial model. This Torrens et al., 1999). Finally, alternatives generation cor-
clearly supports the claims made by the mental models related with uses of objects (a measure of creativity) and
theory that the search for falsifying conclusionsunderlies rationality (a measure of cognitive motivation),neither of
the reasoning process.
which are strong correlates of intellectualability, and nei-
As predicted, alternatives generation did not correlate ther of which independentlypredicted syllogistic reason-
with performance on impossible conclusions.It did, how- ing performance.This suggeststhat alternativesgeneration
ever, correlate with necessary conclusions. There may be is more a measure of processing style than of intelligence.
a very goodreason for this. Necessary conclusionsalways These arguments are not conclusive, but taken together,
used the quantifiersome not, but we know from the results they do suggest that alternatives generation is not simply
of Experiment1 that the exclusionmodel is likely to be the a measure of intellectual ability.
first one constructed. Although logically some not is con-
sistent with the exclusion relationship, we know that
participants do not like to give this conclusion (Roberts,
Newstead, & Griggs, 2001). As a consequence, reasoners
GENERAL DISCUSSION
The present data providestraightforward answers to the
may reject the conclusionunless they are able to construct questions raised in the introduction. There is clear evi-
an alternativerepresentationthatis consistentwitha layper- dence for individualdifferences in the tendency to gener-
son ’s interpretationof some not. This means, of course, that ate alternative representations of deductive reasoning
peoplewho can construct such alternativerepresentations problems. At one extreme, some people seem to produce
are more likely to accept necessary conclusions.
just a single representation and to provide a conclusion

1
36
NEWSTEAD, THOMPSON, AND HANDLEY
consistent with this. They thus accept as necessary conclu- is, of course, entirely consistent with this theory and in-
sionsthat are in fact compatibleonlywith the initialmodel; deed is derived from it. Furthermore, the similarities be-
and they also reject as not possibleconclusionsthat are in- tween diagram generationand initialconclusionsobserved
compatible with the initial model but that are compatible in Experiment 1 are exactly what would have been pre-
with alternative models. At the other extreme, some peo- dicted by this theory (though the discrepancies on some
ple seem to attempt to falsify initialconclusionsand do so syllogisms may also be cause for concern).
by generating alternative models.
Explanationsof ourfindingsinterms ofothertheoriesare
We have also shown that it is possible to measure inde- possible but seem less natural. For example, mental logi-
pendentlythis tendencyto constructmore than one model. cians such as Rips (1994) could no doubt argue that alter-
The alternatives generation task is one that seems to pro- natives generation is related to some more general logical
vide a reasonablemeasure of this and that we now know to ability and that the correlations between this measure and
be a predictorofbothsyllogisticreasoningandvarious kinds reasoning performance are mediated by this more general
of conditional reasoning (Newstead et al., 2002; Torrens ability. However, althoughalternativesgenerationcorrelates
et al., 1999). What is more, this measure seems to corre- with intellectual ability, we have argued that it is more a
late with other, more generic measures of flexible think- measure of processing style than of reasoning ability.
ing: the uses of objects test and rationality scores on the There is nothingin the mental logicapproach that captures
REI. This suggests that it is related to measures of flexi- why alternativesgenerationshould be a key componentin
ble thinking.
human reasoning.In contrast, generatingalternativesis at
In some ways, it may seem a little surprising that the the heart of mental models theory, and this view provides
alternatives generation measure proved to be a good pre- a natural and compelling account of the findings pre-
dictor of syllogistic reasoning performance. The measure sented here.
involvesproductionof Euler circlesto representsyllogisms,
and this may lead to the adoption of different representa-
REFERENCES
tionalstrategiesto those used in syllogisticreasoningitself.
Bucciarelli, M., & Johnson-Laird, P. N. (1999). Strategies in syllo-
Nevertheless, it is clearly a fairly demanding task, and the
gistic reasoning. Cognitive Science, 23, 247-303.
Byrne, R. M. J. (1989). Suppressing valid inferences with conditionals.
skillsrequiredtoproducethesediagramsseem tohavesome
generality. The ability to generate alternatives may be a
Cognition, 31, 61-83.
skill that more able thinkers develop, in the same way that Cummins, D. D. (1995). Naive theories and causal deduction. Memory
&
Cognition, 23, 646-658.
experts develop skills that enable them to use long-term
memory efficiently (Ericsson & Kintsch, 1995).
Erickson, J. R. (1974). A set analysis theory of behavior in formal syl-
logisticreasoning tasks. In R. L. Solso (Ed.), Theoriesin cognitivepsy-
chology: Loyola symposium (pp. 305-330). Potomac, MD: Erlbaum.
It is important to note that the success of the alterna-
tives generation task in predicting performance does not Ericsson, K. A., & Kintsch, W. (1995). Long-term working memory.
Psychological Review, 102, 211-245.
Evans, J. St. B. T., Handley, S. J., Harper, C. N. J., & Johnson-
Laird, P. N. (1999). Reasoning about necessity and possibility:A test
of the mental model theory of deduction. Journal of Experimental
commit us to a specific form of representationand, in par-
ticular, does not commit us to assuming that Euler circles
are implicated in syllogistic reasoning. A number of the-
orists have claimed that reasoners use Euler circles as a
Psychology: Learning, Memory, & Cognition, 25, 1495-1513.
means of representing syllogisms (Erickson, 1974; Ford, Evans, J. St. B. T., Newstead, S. E., Allen, J. L., & Pollard, P.
1994). Debiasing by instruction: The case of belief bias. European
(
1
995; Stenning & Oberlander, 1995). We are using alter-
Journal of Cognitive Psychology, 6, 263-285.
Ford, M. (1995). Two modes of mental representation and problem so-
lution in syllogistic reasoning. Cognition, 54, 1-71.
natives generation as a general measure of how well peo-
ple can produce conceptuallydistinct representations; we
do not commit ourselves to claims as to the form of those Handley, S. J., Dennis, I., Evans, J. St. B. T., & Capon, A. (2000). In-
dividual differences and the search for counter-examples in syllogis-
tic reasoning. In W. Schaeken, A. Vandierendonck, & G. de Vooght
representations.
These findingshave clear implicationsfor mental mod-
els theory. On the one hand, the data support the work of
(
Eds.), Deductive reasoning and strategies (pp. 241-265). Mahwah,
NJ: Erlbaum.
those (e.g., Newstead et al., 1999; Polk & Newell, 1995) Handley, S. J., Feeney, A., & Harper, C. (in press). Alternative an-
who have suggested that the search for alternative models
is not a compulsory part of the reasoningprocess. Instead,
the default procedure appears to be one in which reason-
ers proceed on the basis of a single model and are willing
to accept a conclusion if it is consistent with that model.
On the other hand, there is also clear evidence that rea-
soners are capable of constructing alternative models and
do so when either individual inclinations or task-specific
characteristicspromptthemto. Moreover, theabilitytocon-
struct alternatives is linked to logically correct reasoning
performance.
tecedents, probabilities, and the suppression of inferences in Wason’s
selection task. Quarterly Journal of Experimental Psychology.
Hudson, L. (1966).Contrary imaginations:A psychologicalstudy of the
English schoolboy. Harmondsworth, U.K.: Penguin.
Johnson-Laird, P. N. (1983). Mental models: Towards a cognitive sci-
ence of language, inference, and consciousness. Cambridge: Cam-
bridge University Press.
Johnson-Laird, P. N., & Byrne, R. M. J. (1991). Deduction. Hove,
U.K.: Erlbaum.
Klaczynski, P. A., Fauth, J. M., & Swanger, A. (1998). Adolescent
identity: Rational vs. experiential processing, formal operations, and
critical thinkingbeliefs. Journalof Youth & Adolescence, 27, 185-207.
Klauer, K. C., Musch, J., & Naumer, B. (2000). On belief bias in syl-
logistic reasoning. Psychological Review, 107, 852-884.
Markovits, H. (1984). Awareness of the “possible” as a mediator of for-
mal thinking in conditional reasoning problems. British Journal of
Psychology, 75, 367-376.
Proponentsof mental models theory will thus take con-
siderablecomfort from the present findings.The proposed
centrality of alternatives generation to human reasoning

GENERATING ALTERNATIVES AND HUMAN REASONING
137
Newstead, S. E., Handley, S. J., & Buck, E. (1999). Falsifying men-
tal models: Testing the predictionsof theories of syllogisticreasoning.
Memory & Cognition, 27, 344-354.
Newstead, S. E., Handley, S. J., Harley, C., Wright, H., & Far-
relly,D. (2002).Individualdifferences in deductive reasoning. Man-
uscript submitted for publication.
interpretation and syllogistic reasoning. Thinking & Reasoning, 7,
173-204.
Stanovich, K. E., & West, R. F. (1998).Individualdifferences in rational
thought. Journal of Experimental Psychology: General, 127, 161-188.
Stenning, K., & Oberlander,J. (1995). A cognitive theory of graph-
ical and linguistic reasoning: Logic and implementation. Cognitive
Science, 19, 97-140.
Newstead, S. E., Pollard, P.,Evans,J. St. B. T., & Allen,J. L. (1992).
The source of belief bias in syllogistic reasoning. Cognition, 45, 257- Thompson, V. A. (1995).Conditionalreasoning:The necessary and suf-
2
84.
ficient conditions.CanadianJournalof Experimental Psychology, 49,
1-58.
tial processing styles to personality, basic beliefs, and the ratio-bias Thompson, V. A. (2000). The task-specific nature of domain-general
Pacini, R., & Epstein, S. (1999). The relation of rational and experien-
phenomenon.JournalofPersonality& SocialPsychology, 76, 972-987.
Polk, T. A., & Newell, A. (1995). Deduction as verbal reasoning. Psy-
chological Review, 102, 533-566.
Rips, L. J. (1994). The psychology of proof. London: MIT Press.
Roberts, M. J. (2000). Strategies in relational inference. Thinking &
Reasoning, 6, 1-26.
reasoning. Cognition, 76, 209-268.
Torrens, D., Thompson, V. A., & Cramer, K. M. (1999). Individual
differences and the belief bias effect: Mental models, logical necessity,
and abstract reasoning. Thinking & Reasoning, 5, 1-28.
(Manuscript received October 23, 2000;
Roberts, M. J., Newstead, S. E., & Griggs, R. A. (2001). Quantifier
revision accepted for publication August 24, 2001.)