Full text of DOI:10.3758/BF03195270

Memory & Cognition
2002, 30 (1), 107-118
Effect of theory-based feature correlations
on typicality judgments
WOO-KYOUNG AHN, JESSECAE K. MARSH, and CHRISTIAN C. LUHMANN
Vanderbilt University, Nashville, Tennessee
and
KEVIN LEE
Yale University, New Haven, Connecticut
In the present study, we examine what types of feature correlations are salientin our conceptual rep-
resentations. It was hypothesized that of all possible feature pairs, those that are explicitly recognized
as correlated (i.e., explicit pairs) and affect typicality judgments are the ones that are more likely the-
ory based than are those that are not explicitly recognized (i.e., implicit pairs). Real-world categories
and their properties, taken from Malt and Smith (1984), were examined. We found that explicit pairs
had a greater number of asymmetric dependency relations (i.e., one feature depends on the other fea-
ture, but not vice versa) and stronger dependency relations than did implicit pairs, which were statisti-
callycorrelatedin the environment but were not recognizedas such. In addition, people more oftenpro-
vided specific relation labels for explicit pairs than for implicit pairs; these labels were most often
causalrelations.Finally,typicalityjudgments were more affectedwhen explicit correlationswere broken
than when implicit correlations were broken. It is concluded that in natural categories, feature corre-
lations that are explicitly represented and affect typicalityjudgments are the ones about which people
have theories.
It is generallyagreed thatour conceptsare formed around propertieswhenlearningcategories.Instead, people’sknowl-
clusters of correlated features (Rosch, 1978). For instance, edge of a domain emphasizes only a subset of the statisti-
having wings
being able to fly
is more likely to occur with
cal correlations that occur. We believe that this possibility
has yet to be tested, especially with natural categories, as
having feathers
being able to swim
hav-
and
than with
and
ing gills
bird
. Thus, we give thename
to creaturesthat have will be elaborated upon below. In our studies we will strive
wings and feathers and are able to fly. Within these corre- to (1) demonstrate that explicitly represented feature cor-
lation clusters, there are specific feature correlations that relationsin conceptsare theory based, (2) examinethecon-
are more important and explicit in our conceptual repre- tent of these naive theories, and (3) demonstratethat these
bird
havingwings
beingable
and
sentations.In our
to fly
concept,
would be more closely tied to each other than
feathers chirping
theory-based, explicitly represented feature correlations
are more determinativeof categorymembers’typicalitythan
having
and
. The goal of thepresentstudyis toex- are theory-neutral feature correlations. To achieve these
amine which feature correlations,among all possiblepairs goals, we must first review previous studies on the effect
in a cluster, are explicitly represented in concepts and af- of correlated features, particularly noting studies that in-
fect typicality ratings of exemplars beyond the contribu- volved correlated features in real-life categories.
tion of their individualfeatures.
The work on people’s naive theoriesof naturalcategories Effect of Correlated Features
(Carey, 1985; Keil, 1989; Murphy & Medin, 1985) sheds Found in Previous Studies
some lighton our goals. Accordingtothisapproach,people
Are people sensitive to all correlations among features
do not simply notice and record statistically correlated when learningnovelcategories?In an attempttoanswer this,
Medin, Altom, Edelson, and Freko (1982)had participants
learn artificial categories. When asked to judge category
This project was supported by National Institute of Mental Health
membership for new items, those that preserved the study
Grant RO1 MH57737,awarded to the first author.We thank Kathy John-
phase correlation were more likely to be selected as mem-
son, Barbara Malt, Thomas Palmeri, and Steven Sloman for their ex-
bers, evenwhen theitemwiththepreserved correlationcon-
tremely helpful comments on earlier drafts of this paper. We also thank
tained fewer typical features (see also Wattenmaker, 1991,
1993). However, these results might not generalize to real-
world cases, because the correlation between features was
Barbara Malt for providingthe list of basic properties and the family re-
semblance weightings of properties used in Experiment 4. Correspon-
dence concerning this article should be sent to W.-K. Ahn, Department
of Psychology, Vanderbilt University, Wilson Hall, 111 21st Avenue
South, Nashville, TN 37203(e-mail: woo-kyoung.ahn@vanderbilt.edu).
having wings
perfect, unlike in natural concepts (e.g.,
and
). Furthermore, in these studies, a rule in-
volving the correlated features was more diagnostic and
being able to fly
—Accepted by previous editorial team.
107
Copyright 2002 Psychonomic Society, Inc.

108
AHN, MARSH, LUHMANN, AND LEE
simplistic than any rule involving individual features a category. In Part 2 of their Experiment 2, they used only
alone, making feature correlations unfairly advantaged correlated features in their exemplars that were explicitly
over individual features (see also Wattenmaker, 1993). judgedto be correlated. Participantsnow found exemplars
Therefore, it is difficult to conclude that people would be with intact correlations to be more typical than exemplars
all
sensitive to correlationsamong features in natural con- breaking such correlations.
cepts, as these studies have demonstrated with artificial
stimuli.
The specific question investigatedin our present exper-
iments is why only a subset of the statistically significant
Using natural categories, Malt and Smith (1984) found correlations were recognized as being correlated and why
that peopleare sensitiveto feature correlationsto some de- only this subset affected typicality judgments. Malt and
gree. Althoughthis work has been regularlycited as a clas- Smith (1984) end their article by pointing out the need for
sic study demonstratingsensitivityto correlated features, a this very investigation:
close examination reveals that their results are not so
Exactly what that subset of [correlated] pairs consists of,
straightforward. In their study, about two thirds of the sta-
however, has not been systematically explored. Whether
tistically correlated properties in natural categories were
they are primarily pairs that are functionallyrelated—such
not explicitly noticed as being correlated. Furthermore,
as large birds requiring large wings in order to fly—or are
Malt and Smith did not find that these statistical correla-
empiricallyplausiblecombinations—such as birds far from
tions invariably determined goodness of exemplars in a
category. Sincethe presentstudy is based on the stimuliand
procedure used in their study, we will give a detailed de-
scription.
Malt and Smith (1984)beganby askingwhat features are
statistically correlated within a category. In Part 1 of their
the ocean being unlikely to find fish to eat—or are simply
themost frequentlyoccurringcombinations,is not clear.One
step toward better understandingthe human categorization
processwould seem to be in identifyingwhat types of prop-
erties and property pairs are used. (p. 269)
Experiment1, participantswere asked to listpropertiesthat The Role of Background Knowledge
are generallytrue of members (e.g., robin, sofa) of various
The number of possiblecorrelations in the world is dis-
categories (e.g., birds, furniture). Properties that were turbinglylarge (e.g., Keil, 1981;Murphy& Medin, 1985).
listed by at least a third of the participants and present in For instance, in a category with only seven binary-valued
at least two members in a category were then provided to dimensions,84 correlationsexist.Out of these,whichtypes
another group of participants, who rated the properties as would people notice?
to how much each appliedto each categorymember. An ex-
One way of gettingaround the computationalcomplex-
3
emplar property table was created, containingmean rat- ity is to be guidedby existing backgroundknowledge.Nu-
ings for every category member on all properties. Pearson merous studies have demonstrated how people’s judg-
correlations were conducted on the mean ratings across ments of correlationbetween two events are influencedby
category members for all possible pairs of properties. For their initialhypothesis(e.g., Chapman & Chapman, 1967;
instance,if two propertiesare correlatedin a category, cat- see Alloy & Tabachnik,1984,for a review). These findings
egory members with high mean ratings on one should shouldalso holdin detectingcorrelationswithin concepts.
have also received high ratings on the other. On average, If people understand the reasons behind correlations,
33.5% of the property pairs were significantly correlated these correlations will be noticed more easily. For in-
within a category at the .05 level. Henceforth, we will call stance, consider some of the statistical correlations found
statisticallycorrelated feature pairs
Beinggray/white
these pairs
The next question Malt and Smith (1984) raised was in birds was positivelycorrelatedwith
ing buttons having long sleeves
.
inExperiment1 ofMaltandSmith(1984):
eating fish
hav-
, and
were positively corre-
whetherthesestatisticallysignificantcorrelationsaffect typ-
and
icality judgments of exemplars. In Part 2 of their Experi- lated in clothing. Yet, the causal links underlying these
ment 1, participants received exemplars in which these correlationsare notimmediatelyapparent.In contrast,most
correlations were either preserved or broken, while fam- laypeople would know causal connections underlying the
ily resemblance scores1 were equated to examine the cor- correlation between
relations’ effects on typicality judgments separately. Sur- for clothing, or
and
eating fish
for
being made of wool
being warm
living near the ocean
and
prisingly, they failed to find an effect. That is, unlikein the birds. Indeed, the latter pairs, and not the former, are the
previousstudiesusing artificial stimuli(e.g., Medin et al., ones that were shown to have been explicitly noticed in
1982), people were not sensitive to the correlated struc- Malt and Smith. Thus, we propose that people tend to no-
tures presented in natural categories.
tice correlationsin conceptsif they are meaningfulto them
One explanationthat Malt and Smith (1984) offered for in light of their existing domain theories.2
this failure was that property correlations were statistical
co-occurrences and were not necessarily present in peo- influence typicality judgments. According to the
We further suggestthatit is explainablecorrelationsthat
theory-
based
ple’s representations. In their Experiment 2, participants
approach to categorization,conceptsare embedded
rated the extent to which they thought each pair of proper- in theories and are represented by causal links and ex-
ties was related within the category. Of the original pairs, planatory relations. Murphy and Medin (1985) state that
only 33% were explicitly rated as correlated. Therefore, “the connectionbetween those [correlated] features is not
peoplewere not aware of all feature correlationspresent in a simple link, but a whole causal explanation for how the

FEATURE CORRELATION
109
two are related”(p. 300).From thisperspective,conceptual Overview of Experiments
coherence is based on explanations or domain theories,3
We carried out four experiments to test the hypothesis
rather than on clusters of undifferentiated correlations. If that domain theories determinewhich feature correlations
peopleview conceptualcoherencethisway, exemplars that would be salientin our representationof naturalcategories.
preserve theory-basedcorrelationswill be judgedto be bet- First, in Experiment 1, we determined which feature corre-
tercategorymembers thanare exemplarsthatdonotpreserve lationswere salient in people’s conceptualrepresentations
such correlations, all else being equal. On the other hand, by replicating Malt and Smith’s (1984) study. In Experi-
mere statistical correlations that are not linked by ex- ments2 and3, we examinedwhetherthefeature correlations
planatory relationswould not influencemembership judg- that were salientwere more theory based than theonesthat
ment as much, because these relations are not critical in were notsalient.Finally, inExperiment4, we tested whether
conceptual representations. Therefore, our second hypoth- these two types of feature correlationswould differentially
esis is that the types of correlations that should matter in affect typicality judgments of category exemplars.
typicality judgments are the ones that are linked within
one’s domain theories.
EXPERIMENT 1
Unfortunately, few previous studies have directly ex-
amined the relationship between background knowledge
In Experiment1, we examinedwhichfeature correlations
and sensitivity to correlations in natural categories. Some in concepts people explicitlynotice. This is a crucial step,
studiesused artificialstimulito examinethisissue, butthey because the explicitnessof relationsserves as the main in-
were limited in various ways. For instance, Barrett, Abdi, dependent variable for the remaining experiments. Al-
Murphy, and Gallagher (1993) found that children were though Malt and Smith (1984) already had provided such
havinga
more sensitive to theory-based correlations(e.g.,
data, we replicated this study with the same participant
) than to theory-neutralcorre- population as in the rest of the experiments.
big brain a good memory
and
lations. As the authors acknowledged,however, this result
could have been due to a simple heuristic (e.g., big things Method
Participants. Twenty-one undergraduates from Vanderbilt Uni-
are good, so big brains go with good memory), rather than
to children’s specific domain theories. In their Experi-
ment 2, designedto avoidthis problem, children had more
difficulty;more than 50% of theparticipantswere dropped
for failing to show consistent responses. Thus, it is diffi-
cult to tell how prevalent the effects of theory-based cor-
relations really were.
Murphy and Wisniewski (1989) studied the effect of
theory-basedcorrelationsin adults’categorizationof natural
concepts. In this study, when a feature in a theory-related
versity participated in partial fulfillment of introductory psychology
course requirements. One of these participants was excluded from
the study because more than 50% of his/her responses were made
within 500 msec, which was deemed not enough time to read the
stimuli under normal circumstances.
Materials and Procedure. The procedure was modeled after
Part 1 of Malt and Smith’s (1984) Experiment 2, with modifications
being made to support a computer presentation of the stimuli. We
used property pairs found to be statistically correlated in Malt and
Smith’s Experiment 1 for six different categories. These categories
were Bird, Clothing, Flower, Fruit, Furniture, and Tree. The partic-
ipants were shown one property pair on the computer screen at a
time, with the appropriate category name appearing at the top of the
screen in capital letters. The participants were told that for each pair,
they were to “make a judgment as to how well the features go to-
gether in that category.” The participants were instructed to use a 7-
eats fish
lives under water
and ) was substituted
pair (e.g.,
with a feature in a contrast category (e.g.,
lives under water
eats wheat
and
), the broken relation significantly low-
ered categorymembership judgments.As in this example,
broken correlations in this study resulted in bizarre ex-
amples, violating layperson’s ontological assumptions
(e.g., Keil, 1981). Murphy and Wisniewski proposed that
people use theirdomain theories to rule out implausibleor
-
point scale ranging from 3 (a strong negative relation) to +3 (a
strong positive relation), with a midpoint of zero standing for no re-
lationship. The participants were given examples of feature pairs
within the category Deer that would fit both of the scale’s extremes
bizarreco-occurrencesof featureswithina concept.Although (is male and has horns for the positive extreme and is female and has
horns for the negative extreme) and the midpoint (has horns and
lives near a lake) during the initial instructions. During pair presen-
tation, a reminder scale that depicted the rating range was shown at
thiscertainlywould be onemechanismthroughwhichprior
knowledge can play a role in concept learning, it does not
answer the initial question that we posed in relation to
the bottom of the computer screen. The pairs were blocked by cate-
Malt and Smith’s (1984)results. In Malt and Smith’s study,
gory, and pair presentation was randomized within each category.
the effect of correlated features was found even when the
items with broken correlations were not bizarre.
Each subject received all 178 correlated property pairs. The experi-
ment was programmed using SuperLab, Version 1.75.
To summarize, previous studies have shown that back-
ground knowledge affects people’s sensitivity to correla- Results and Discussion
tions and suggestedpossibleunderlyingmechanisms (e.g.,
The goal of this experiment was to obtain updated data
background knowledge rules out implausible correla- on what feature pairs were explicitly noted as correlated.
tions). However, none of the studies has provided strong For each feature pair, a mean rating was computed across
evidencethatthepresenceofnaivedomaintheoriesaccounts all participants. Following the criterion used in Malt and
for why certain correlationsmattered more in typicalityjudg- Smith (1984), pairs rated as greater than or equal to +1.5
-
ments of real-life concepts, as was found by Malt and for positive correlations and less than or equal to 1.5 for
Smith (1984).
negative correlations were defined as saliently related

110
AHN, MARSH, LUHMANN, AND LEE
ex-
havingcushions
having cush-
. We predicted that this
pairs. These pairs will from now on be referred to as
doesnot dependon
ions having four legs
, nor does
plicitpairs
. Feature pairs that fell within the cutoffs will be
depend on
implicitpairs 4
referred to as
. Overall, 51 feature pairs were would be a dominantpattern for the implicitpairs. To sum-
explicitlyrated as correlated, representing29% of the total marize, we predicted that in the explicit pairs, one of the
pairs. The percentageof pairs explicitlyrated as correlated two dependency judgments within a pair would be high,
varied between categories, ranging from 13% for the whereas in the implicit pairs, both would be low. Thus, a
clothingcategoryto 44% for furniture. A list of the explicit maximum dependencyrating within a pair would be higher
pairs and their mean ratings can be found in Appendix A. in the explicit pairs than in the implicit pairs. In addition,
Our results were fairly consistentwith Malt and Smith’s the difference between the two dependency judgments
(1984) results, in that 83.1% of the pairs were consistent within a pair would be higher in the explicitpairs (because
in their classification across the two studies. Of the 30 B depends on A but A does not depend on B) than in the
pairs that showed discrepancy, 19 were within a 0.5 range implicitpairs (because both judgmentswould be low in the
from a cutoff point, indicating that most of the discrep- implicit pairs).5
ancy was due to marginal cases. Unless noted otherwise,
we use the results from Experiment 1 in the subsequent Method
Participants. Eighteen undergraduates from Yale University and
experiments when we make a distinctionbetween explicit
10 undergraduates from Vanderbilt University participated in this
and implicit feature pairs.
experiment in partial fulfillment of an introductory psychology
course credit.
EXPERIMENT 2
Materials. The same property pairs as those from Malt and Smith
(1984) were used, as in our Experiment 1. For each property pair,
Experiments2 and 3 testthehypothesisthat inreal-world
concepts,thekindsof correlationsthat peopleexplicitlyno-
ticeare theonesfor which peoplehaveexplanations.In Ex-
periment 2, we treattheory-relatedcorrelationsina relatively
content-free manner by assuming that they come in only
one kind (Thagard, 1989). Following Sloman, Love, and
two statements were generated, one to reflect each direction of pos-
sible dependency. For example, using the properties eats fish and
lives near the ocean, two test items were generated: “For a bird,
whether or not it eats fish depends on whether or not it lives near the
ocean” and “For a bird, whether or not it lives near the ocean depends
on whether or not it eats fish.” Since there were 178 feature pairs in
our study, this led to a total of 356 statements (106 explicit state-
ments and 250 implicit statements). Because responding to all 356
statements is too demanding, each participant received roughly half
of these statements. Data from our Experiment 1 was not available
at the time of this study, so we used Malt and Smith’s distinction of
explicit and implicit feature pairs in sampling statements for each
participant. Using Malt and Smith’s distinction, all the participants
received all of the explicit statements, but they received approximately
half of the implicit statements, sampled across six categories.6 In this
dependen-
Ahn (1998), we call these unlabeled relations
cies
. The term is general enough to cover causal, categor-
ical, temporal, or indeed any other kind of directional re-
lation between features. For instance, a piece of furniture
is comfortable because it has cushions, and therefore,
beingcomfortable
dependson
people might not believethat
havingcushions
being gray/white
. In contrast,
in birds de-
living near the ocean
ticipantsonlyaboutdependencyrelations(e.g., “Does
pends on
or vice versa. Asking par- way, the number of implicit and explicit statements for each partic-
ipant was equated as much as possible. Note that Malt and Smith’s
explicit/implicit distinctions were used only when assigning the items
to the participants. All the analyses reported in this study are based on
the explicit/implicit distinctions obtained from our Experiment 1.
Procedure. The participants were instructed to judge the degree
to which they agreed with each statement on a 9-point scale an-
chored with strongly disagree for 1 and strongly agree for 9. The par-
ticipants were alerted to be sensitive to asymmetric dependency re-
lations, rather than simple co-occurrences, as follows:
being
?”) is a less confusing
able to fly
having wings
depend on
and more succinct procedure than listing all possible
theory-based relations in a question (e.g., “Does
having
wings
cause/allow/follow/is depended upon by/trigger/
being able to fly
increase/decrease/etc.
?”).
Experiment 2 used pairs of features that were found to
be statisticallycorrelatedin Malt and Smith’s (1984)Exper-
iment 1. As was found in our Experiment 1, some of these
pairs were more explicitly noted as being correlated (ex-
plicitpairs) than were others (implicitpairs). For each pair
of features (say A and B), we asked participantsto rate the
strengths in two directions (i.e., the degree to which A de-
pends on B, and the degree to which B depends on A). If
a person believes that A causes/allows/is followed by/
triggers/increases/decreases/etc.B, theratingon “B depends
on A” would be high, whereas the rating on “A dependson
Also remember that some relationships might be plausible in one direc-
tion, such as: “For a chair, whether or not you can sit on it depends on
whether or not it is sturdy,” but not in the other direction: “For a chair,
whether or not it is sturdy depends on whether or not you can sit on it.”
Be sure to pay attention to which property is depended upon (the second
property in the statement) and which is dependent (the first property).
After reading the instructions, the participants proceeded to rate
randomly presented statements. The experiment was programmed
using Psyscope 1.1 (Cohen, MacWhinney, Flatt, & Provost, 1993).
B” wouldbe low. Wepredictedthatthiswouldbea dominant Results
pattern in explicitpairs, because we hypothesizedthat peo-
Do explicitpairs havestrongerdependencyrelationsthan
ple have specific explanations for the correlations that do implicit pairs? To answer this question, we examined
maximum dependencyscores
they explicitly notice. On the other hand, if a person rep- what we termed
within each
resents the pair as being merely correlated, without any pair. Recall that the participantsrated dependencystrengths
explanationsfor why, bothratingswill be low. For instance, in two different directions within each pair (A to B and B
having four legs
having cushions
and
for furniture were to A). The larger of the two ratings within each pair (i.e.,
having four legs
found to be statistically correlated, but
the maximum dependency score) is the dependency

FEATURE CORRELATION
111
P
strength of that pair. For each participant, we obtained an would be able to tell that in tires (made by Goodyear
|
P
average maximum score for the explicit pairs and an aver- black) is lower than (black made by Goodyear). Thus,
age maximum score for the implicit pairs. A paired test the relationship between the two is asymmetric, but it is
|
t
showed that maximum scores were significantly higher very unlikely that people know any reason for this rela-
for the explicit pairs (6.18) than for the implicit pairs tionship. We highly suspect that in a case like this, people
t
p
[3.62; (27) = 15.48, < .0001]. As is shown in Table 1, would not have given a high dependency rating. That is, it
the direction of difference was consistentacross all six cat- is quite unlikely that people would say that “whether or
egories. Therefore, the results strongly support the con- not a tire is made by Goodyear depends on whether or not
clusion that explicit pairs have stronger dependency rela- a tire is black” or “whether or not a tire is black depends
tions than do implicit pairs.
on whether or not a tire is made by Goodyear.”The phrase
dependson
We also examined the difference score within each pair.
impliesmuch more than just a conditionalprob-
As was explainedearlier, if, for instance,A causes B, then ability. Therefore, both the maximum and the difference
the difference between the rating on “A dependson B” and dependency rating would have been lower in this kind of
“B dependson A” wouldbe large. However, ifA merelycor- nontheoretical,asymmetric case. Yet, it is still possiblethat
relates with B, the difference would be small.7 For a sta- dependency relations might have implied nontheoretical,
tistical analysis, we first calculated a difference between asymmetric relations, and the difference between the ex-
a rating on “A depends on B” and a rating on “B depends plicit and the implicit pairs might have been due to these
on A” in each pair for each participant.The average differ- nontheoreticalones. Experiment 3 will address this issue.
ence scores for the implicit pairs and those for the explicit
Alternatively, using dependency as an umbrella for all
pairs within each participantwere used as inputfor a paired theory-based relations might have been a gross simplifi-
t
test. The results showed that difference scores were sig- cation; there may have been many other theory-based rela-
depends on
nificantly higher for explicit pairs (2.09) than for implicit tions that are not easily captured by the term
.
t
p
pairs [1.47; (27)= 5.04, < .0001]. As is shown in Table 1, This, however, does not necessarily pose a problem for
the results were consistent across all six categories.
Experiment2, because according to this argument, the use
of dependency relations would have underestimated the
differencesbetween explicitand implicitpairs. Yet, thiscould
Discussion
To summarize, we found that explicit pairs contained still be a problem if most of the implicit pairs happenedto
stronger dependency links than did implicit pairs, as was consist of theoretical relations that could not have been
shown by thedifferencebetween the two kindsof pairs with described as dependency. In this case, we cannot conclude
respect to maximum scores. Furthermore, explicit pairs that the explicit pairs are more theory based than are the
had more asymmetric dependencyrelations.These results implicit pairs, despite the fact that the explicit pairs had
support the hypothesis that explicitly noticed correlated stronger dependency relations than did the implicit pairs.
pairs are more theory based than are those that are not ex- Although this possibility would be a large coincidence,
plicitly noticed.
Experiment 3 was carried out to also address this concern.
Relying onlyon dependencyrelationshad an advantage
in that the participants were not concerned about how to
classify or label relations, making the task easier. How-
ever, this method might have been limited in determining
EXPERIMENT 3
Experiment3 uses a more open-endedtaskby lettingpar-
which feature pairs are theory based; dependencyrelations ticipantsdraw relationsby using any labels that they want.
might have been either too broad or too narrow as opera- The participantswere presented with pairs of features and
tional definitionsfor theory-based relations. Each will be were asked to draw an arrow with a label on top to indicate
discussed.
any specific relation that they believe existed. They were
First, dependency relations might have been too broad specifically told not to draw mere correlations. Thus, it
a measure in that they might have also captured nontheo- was possible to examine whether feature correlations that
retical, asymmetric relations.For instance,most laypeople are explicitlynoticedwere indeedbased on specifictheory-
based relations,rather than on merely salient correlations.
From thesedata, we also examinedthecontentof theory-
based relations.Proponentsof theory-based categorization
(e.g., Carey, 1985;Murphy& Medin,1985;Wellman,1990)
have emphasized causality as a central componentin con-
ceptualortheoryrepresentations.We examinedwhetherthis
indeed was the case.
Table 1
Mean Maximum and Difference Scores for Explicit
and Implicit Pairs
Maximum Scores
Explicit Implicit
6.30 3.21
Difference Scores
Category
Explicit
Implicit
Bird
1.95
2.59
1.90
2.48
2.06
2.47
1.16
1.48
0.90
1.52
1.77
1.46
Clothing
Flower
Fruit
Furniture
Tree
5.63
7.13
6.30
5.76
6.28
3.38
2.15
3.78
4.30
3.71
Method
Participants. Twenty-one Vanderbilt University undergraduates
participated in partial fulfillment of the requirements of an intro-
ductory psychology course. One participant did not finish the task
and was therefore excluded from the study.
Average
SD
6.18
1.26
3.62
1.09
2.09
0.54
1.47
0.50
Materials. Again, the property pairs found to be statistically cor-
related by Malt and Smith (1984) were used. Pairs were arranged on

112
AHN, MARSH, LUHMANN, AND LEE
Figure 1. Sample item from Experiment 3. The label in italics and the arrow represent a
hypothetical addition by a participant.
paper in a table format. Each row of a table was divided into two strengthof 0), andit was foundthatexplicitpairs (2.50)had
columns: a large column on the left and a small column on the right.
significantlystrongerrelationsthan did implicitpairs [1.33;
Within the left column, a feature pair was presented with a large gap
t
p
(19) = 12.16, < .0001]. As is shown in Table 2, the direc-
between the features. The right column was left empty. Figure 1 rep-
resents what each item looked like after a participant added a labeled
arrow and rated the feature pair.
tion of difference was consistent across all six categories.
Yet, one might be concerned that nearly half the time
(43.4%), the implicit pairs were labeled to have specific
relations. However, the implicit pairs that fell close to the
designated cutoff of 1.5 made up the majority of the im-
plicit pairs found to have a specific relation. Indeed, the
correlationbetween the mean number of specific relations
and the mean explicitness ratings from Experiment 1 was
Property pairs were grouped by category, with the category name
appearing at the top of each page. Within each category, the pair order
was randomized so that no property was presented in more than two
consecutive pairs. A second version of the packet was created by re-
versing the order of the properties within pairs (i.e., a feature that ap-
peared on the left in a pair in Version 1 appeared on the right in Ver-
sion 2, and vice versa) and rerandomizing the property pairs within
categories. Ten participants received each version. For each partici-
pant, the order of the six categories within the packet was random-
ized. Each subject received all 178 property pairs.
Procedure. The participants were asked to decide whether a re-
lationship existed between the members of a feature pair. If they be-
lieved that there was a relationship between the members of a feature
pair, they were asked to (1) draw an arrow indicating the directional-
ity of the relationship, (2) label the arrow as to the type of relation-
ship it represented, and (3) rate the strength of the relationship on a
scale of 1 to 5, where 1 represented a weak relationship and 5 rep-
resented a strong relationship. The participants were given the fol-
lowing list of suggested labels: depends on, precedes, is a precondi-
tion for, allows, enables, affects, causes, is a subset of, is an example
of, increases, and decreases. This list was created by the first three
authors of this article and was intended to be exhaustive. To include
any plausible relations that were not thought of, the participants were
encouraged to use labels of their own invention where they felt it to
be appropriate. The participants were instructed that if they believed
a bidirectional relationship existed between members of a pair, to
draw arrows in both directions and label both accordingly. The par-
ticipants did all arrow drawing and labeling in the space between
features in the table and recorded the strength of the relationship in
the empty right column. If the participants did not believe a rela-
tionship existed between members of a feature pair, they were in-
structed to leave the area between a feature pair blank and to write a
zero in the corresponding strength box.
r
significantly positive even within implicit pairs only ( =
.57, < .001).
p
Taken together, the results strongly support the claim
that the more explicita pair is, the more likelypeopleknow
specific relations. In other words, when people explicitly
notice correlations of features in concepts, they do not
simply code them as mere correlations; they have beliefs
about what these relations are.
In addition,a correlational analysis was carried out be-
tween the strengths of the specific relations found in Ex-
periment 3 and the strengthsof the general relationsfound
inExperiment2, and a significantlypositivecorrelationwas
found ( = .71, < .01). This result indicatesthat the mea-
sure of dependency strengths generally converge with the
measure of specific relations used in Experiment 3.
Then what is the contentof these relations? Some of the
labels that participants produced could be easily grouped
as the same kind. For instance,as is shown as notes of Ap-
r
p
decreases
minimizes
are groupedas the same
pendixA,
and
label. Throughdiscussion,thefirst three authorsof the pre-
sent article grouped the labels into seven categories. Ap-
pendix A explains the groupingand lists the relationsthat
were mentioned by at least 4 participantsfor each explicit
pair, with the number of participants who mentioned that
label in parentheses.As can be easily seen, causal relations
were most frequentlylisted(52.4% of all labeledrelations).
In order to providemore objective,quantifiableindices,
The participants were told that any feature pair that they were
shown co-occurred within the given category and, for this reason, to
refrain from labeling relationships as co-occurs or associated. They
were encouraged to use more precise labels that reflected a more
specific relationship. If they could not think of a more specif ic label,
they were asked to rate the pair as no relationship.
causal
we measured how
each relation was from an inde-
pendent group of 8 volunteers(4 undergraduatesand 4 re-
Results and Discussion
First, we looked at what percentagesof explicitand im-
plicitpairs were given a specific relationshiplabel by sub-
jects.The resultsshowed thatexplicitpairswere much more
likely to have been given a specific relation (68.9%) than
Table 2
Mean Strengths of Relations for Explicit and
Implicit Feature Pairs in Experiment 3
Type of Feature Pairs
t
were implicitpairs (43.4%). A paired test was performed
Category
Explicit
Implicit
on proportions of labeled relations in explicit and implicit
pairs for each participant,and a significantdifference was
Bird
2.51
3.69
2.54
3.23
2.33
3.20
1.03
1.66
1.25
1.78
1.83
1.29
Clothing
Flower
Fruit
Furniture
Tree
t
found between the two types of property pairs [ (19) =
p
t
8.45, < .001]. Another paired test was carried out on
mean strength ratings for explicit and implicit pairs for
each participant (with no relations treated as having a

FEATURE CORRELATION
113
search assistants with bachelor’s degrees). They received the final question of the impact of preserving these corre-
a list of all the labels produced by the participants of Ex- lationsontypicalityjudgments.MaltandSmith(1984) were
whether
periment 3. Then they rated each label on “the extent to interested in
feature correlations could influence
which the term implies that there is a (positive or nega- typicality judgments. In the end, they could demonstrate
tive)causal link underlyingthetwo items”ona 9-pointscale the effect when they utilized explicit pairs only. We have
term definitelydoes not imply a causal link
anchored with
taken their findings a step further to see what type of sta-
term definitelyimplies a causallink
for 1 and
rating for each label was obtained and used as a
index
for 9. A mean tisticalcorrelationsmatter most in goodnessof exemplars.
causality
indicating how causal that label is. These indices in our experiments) correlations are much more likely to
SD is an example of
We hypothesizethat explicit(or theory-based,as was found
ranged from 2.38 ( = 1.99) for
to 8.88 influence typicalityjudgmentsthan are implicit(or theory-
. Then, for each feature pair used in neutral) correlations. In Experiment 4, participants were
SD causes
(
= 0.35) for
Experiment3, we calculatedan average causalityindexof presented with pairs of items that either broke or pre-
all relationsmentionedfor that pair. For instance, suppose served a feature correlation and then judged which item
causes
10 out of 20 participants listed
as a relation for a was more typical of a target category.
feature pair, and the other 10 did not list any relation. The
average causality indexfor that feature pair would be 8.88.
Thus, for each pair, we came up with a measure of how
causallythe two features are related, given that it is judged
as having a specific relationship. Across all feature pairs,
the mean causalityindexwas 5.67, which was significantly
higher than 5 [i.e., a midpoint on the 9-point scale used to
Method
Participants. Thirty-three Vanderbilt University undergraduates
participated in partial fulfillment of the requirements of an intro-
ductory psychology course. Three of the participants were excluded
from the study because they did not correctly indicate their re-
sponses in the provided packets.
Materials. From the results obtained in Experiment 1, we selected
all of the property pairs that had been rated as positive explicit rela-
tions.8 These pairs were used to construct artificial exemplars of the
six categories, using the same method as in Malt and Smith (1984).
Specifically, the exemplars for a given category had three com-
mon basic properties. These properties were identical to the ones
used in Malt and Smith (1984) and are listed in Appendix B under
“Constant Features.” In addition, each exemplar had two critical
properties that either formed an explicit pair or did not. In the cases
in which the explicit pair relation was broken, one of the two prop-
erties was replaced by another property of that category that had an
equal or greater family resemblance weight as used by Malt and
Smith.9 As in Malt and Smith’s study, it was ensured that each re-
placement property was not rated as explicitly related or negatively
related to the remaining property in the results of our Experiment 1.
Of the original 35 explicit pairs that were taken under consideration,
25 pairs met the criteria for having matching family resemblance
weights and were therefore used to create exemplars. These pairs are
noted in Appendix A. The pairs varied in number across categories
in the following way: 9 for Furniture, 6 for Tree, 4 each for Bird and
Clothing, and 1 each for Flower and Fruit.
t
p
measure causality indices; (176) = 10.37, < .001]. This
suggeststhat, overall, the specific relationspeople used to
explainfeature correlationsare more causalthannoncausal.
Interestingly, explicitpairs were judged to be more causal
M
M t
= 6.18) than were implicit pairs [ = 5.46; (175) =
(
p
5.49, < .001].
To summarize, we started out with a question of why
some feature correlations are more noticeablethan others
when all are statisticallysignificantlycorrelated. In Exper-
iment 3, we found that feature correlationsthat peopletend
to be aware of differ from feature correlationsthat theytend
to be unaware of, in terms of whether people can provide
explanations for how they are related. People were more
likely to provide specific labels for the explicit pairs than
for the implicit pairs. In addition, we found that causality
is an importantcomponentin theseexplanations,especially
for the ones provided for the explicit pairs.
One final point to be discussed is the finding that a ma-
jority of the participants came up with theory-based rela-
tions between pairs that had a category name as a feature
(i.e., being a bird). As is shown in Appendix A, three such
After the selection of the 25 explicit pairs, 25 implicit pairs were
chosen, using the same selection procedure as that described for the
explicit pairs, and with an additional constraint that there be an equal
number of implicit pairs and explicit pairs within each category. The
implicit pairs were then used to create implicit exemplars of the
is a bird
has
cases were
feathers
lays
causes/allows, is a preconditionfor
has a beak is a bird
also causes/allows same format as the explicit exemplars.
, and
. One might argue that these relations are more defin-
itional than explanatory: Animals that have feathers and
birds
and
In order to verify that the 25 explicit pairs selected for Experiment 4
are more theory based than the 25 implicit pairs, these two sets of
pairs were compared with respect to the two measures used in Ex-
periments 2 and 3. The maximum dependency ratings measured in
Experiment 2 were significantly higher for the selected explicit pairs
(M = 6.38) than for the selected implicit pairs [M = 3.45; t(48) =
8.21, p < .0001]. Similarly, the strengths of the explanatory relations
measured in Experiment 3 were significantly higher for the selected
explicit pairs (M = 3.22) than for the selected implicit pairs [M =
1.11; t(48) = 10.90, p < .0001].
Exemplars were created so that each possessed three basic prop-
erties and two critical properties, except for one bird and one tree ex-
emplar. These two exemplars had four features, because one of the
critical pair features was the same as a constant feature used by Malt
and Smith (1984; i.e., has feathers and has leaves). For each stimulus,
beaks and that lay eggs are called
. Subsequentanaly-
ses showed that the pattern of all results reported in this
study still holds withoutthese pairs. In addition,we argue,
onthebasisofpsychologicalessentialism(Medin &Ortony,
1989), that these relationsare truly believedto be explana-
bird essence
tory. That is, peoplewould havemeant that the
causes birds to have feathers and a beak and to lay eggs.
EXPERIMENT 4
In Experiments 2 and 3, we found that correlations that
are explicitly noticed are theory based. Now, we return to the constant properties were listed first, followed by the critical pair.

114
AHN, MARSH, LUHMANN, AND LEE
An exemplar preserving the original explicit or implicit relation clearly supported the hypothesisthat explicitcorrelations,
and an exemplar breaking this relation appeared on each page of the
butnot implicitcorrelations,matter intypicalityjudgments.
experimental packet. In half of the explicit and implicit pair items,
Anotherway of lookingat this interactionis to compare
the preserved exemplars were placed on the left side of a page, and
each of the preserved and broken correlations across the
in the other half they were placed on the right side of a page. In all
explicit and the implicit pairs. For items with broken cor-
of the items, the position of the feature that differed between the two
relations, there was no difference between the explicit
sides was randomly placed as either last or second to last in the fea-
ture list. Exemplars were blocked by category, and category presen-
tation was randomized for each participant. Within each category,
explicit and implicit exemplars were randomly mixed by page. A
second set of materials was created from the first version by revers-
ing the side on which the preserved correlation could be found. The
second set of materials was identical to the first in every other re-
spect. The instructions for the booklet were typed on the first page
of the experimental packet.
Procedure. The procedure was identical to Part 2 of Malt and
Smith’s (1984) Experiment 2. To reiterate, the participants were pro-
vided with pairs of exemplars and were asked to choose which of the
pair, the left or the right, seemed more typical of the category listed
at the top of the page. They were then asked to rate both exemplars
on a 1 to 7 scale as to exactly how typical each was of its category,
where 1 was very atypical and 7 was very typical. They were told to
t
p
pairs andtheimplicitpairs [ (29) = 0.60, = .55]. However,
the items with preserved explicitcorrelationswere judged
to be significantly more typical than the items with pre-
t
p
served implicitcorrelations[ (29) = 4.69, < .0001]. That
is, the interaction was not obtained because breaking cor-
relations supported by domain theories greatly lowered
typicalityratings. Instead, it was obtainedbecause preserv-
ing correlations supported by domain theories increased
thetypicalityofexemplars.TheseresultscontrastwithMur-
phy and Wisniewski (1989), who found that the effect of
theory-based correlations was obtained because items
with broken correlations were bizarre in light of domain
eats wheat
lives under water
and ), as was
theories (e.g.,
avoid making their decisions by simply comparing the two proper- discussed earlier in the introduction.That is, according to
ties that differed and, instead, to consider each exemplar as a whole.
their proposal, background theories matter because do-
maintheoriesrule outimplausiblecases. Our findingsshow
that preserving theory-based correlations could even en-
hance the typicality of exemplars.
Results and Discussion
As was predicted,the participantswere much more likely
to chooseexemplars preservinga correlationwhen thecor-
relationwas an explicittype(or theory based,as was found
in previousexperiments)than when it was an implicittype
(or theory neutral). For each participant,the mean number
of selections of an exemplar with a preserved correlation
over a broken correlation was calculated for explicit and
GENERAL DISCUSSION
It has often been describedthat, in concepts,feature cor-
relations affect categorization beyond the contribution of
their individualfeatures. Medin et al. (1982) provides one
of the strongest sources of evidence for this. Yet, we ar-
gued that their results from artificial stimuli are difficult
to generalizeto naturalconcepts.Using featurecorrelations
that were intended to be more similar to those in natural
concepts, Murphy and Wisniewski (1989) failed to find
any evidencethat peoplebasetheircategorizationdecisions
on the statisticalrelationof individualfeatures. When nat-
t
implicitpairexemplars.A paired testbased onthesescores
showed a significant difference between explicit pair ex-
M
M
=
emplars ( = 70.8%) and implicit pair exemplars [
t
p
49.0%; (29) = 9.362, < .0001]. In fact, choices on im-
plicit pair exemplars did not differ from the chance level
t
p
t
of 50% [ (29) = 0.64, = .53]. Another test was performed
by treating items as random variables, and a significant
t
p
difference was found [ (48) = 3.31, < .01].
Finally, the participants’typicalityratingswere analyzed.
For each participant,a mean rating was obtainedfor eachof
the four types of feature pairs (i.e., correlation-preserved
and correlation-brokenitems for explicitand implicitfea-
ture pairs). The overall means for these four types of fea-
ture pairs are shown in Figure 2. A repeated measures
analysis of variance found a significant main effect of
F
whetherornotcorrelationswere preserved[ (1,29) = 30.26,
MSe
p
= 0.026, < .001] and a significant main effect of
F
whethercorrelationsare explicitorimplicit[ (1,29) = 7.32,
MS
p
e = 0.056, < .05]. But these main effects should be in-
terpreted in light of the significant interaction between
MS
F
p
e = 0.042, < .001].
these two factors [ (1,29) = 16.34,
The interaction was significantbecause the typicality rat-
M
ings on items with preserved explicit correlations (
=
5.11) were significantly higher than the typicality ratings
M
on items with broken explicit correlations [ = 4.79;
t
p
(29) = 6.24, < .001], whereas the typicality ratings on
items with preserved implicit correlations (4.84) did not
differ from those on items with broken implicit correla-
Figure 2. Mean typicality ratings for items that preserved or
t
p
tions [4.83; (29) = 0.23, = .82]. Therefore, the results broke explicit or implicit correlations.

FEATURE CORRELATION
115
ural categories are examined, Malt and Smith (1984) fines of artificially created stimuli. Thus, we found that
found that only about one third of the feature correlations theoretical knowledge has a significant influence on the
that were statistically significant were noticed as being perception of coherence in real-life categories. As was
correlated. Furthermore, Malt and Smith failed to find any mentionedabove, it has been shown that one’s domainthe-
effect of feature correlations when all statistically signif- ory can help rule out incoherent exemplars (Murphy &
icant feature correlations were examined. Instead, feature Wisniewski,1989).In thepresent study, we haveshown that
correlations that were explicitly noticed as being corre- explanationsunderlying feature correlations can increase
lated influencedtypicalityratings over and above each in- the typicality of an item above and beyond the impact of
dividual feature’s family resemblance score. Our main individual feature frequencies, whereas mere statistical
goal was to examine why only a subset of feature correla- correlations that could not be explained failed to do so.
tions are explicitly noticed and influence categorization. This findingprovidesevidenceof a robust relationshipbe-
The present study providesevidencesuggestingthat the tween conceptual coherence and explanatory coherence.
correlationsthat are present in our representationsof real- Finally, we have provided the first empirical evidence for
world conceptsare the ones with explanatoryrelations.In the general assumption that causality constitutesa major-
Experiment 1, we found that people did not explicitlyrec- ity of relationsin laypeople’s representationsof naturalcon-
ognize the relationships between all the features that sta- cepts (Carey, 1985; Wellman, 1990).
tistically co-occurred within a category, replicating Malt
Our results focused on the effect of theory-based cor-
and Smith’s (1984) results. In Experiment2, we found that relations only on typicality judgments. Will theory-based
the pairs people explicitlyrecognized had stronger depen- correlationsinfluence other kinds of behaviors associated
dency relations. In Experiment 3, we found that explicitly with concepts,such as inductiveinferences? We speculate
recognized feature correlations were much more likely to that the answer is yes. Lassaline (1996) found that induc-
be associated with specifically labeled relations and that tive strengthsof features are greater when the features are
those relations tended to be associated with causality. In bound by causal relationsthan by noncausalrelations.For
Experiment 4, we demonstrated that preserving these ex- instance, participants were given the descriptions of two
plicit,theory-basedcorrelationsmattered in typicalityrat- animals(e.g., AnimalA and Animal B). Theywere told that
ings. Thus, an item with a theoretically related pair (e.g., each animal had a property (e.g., Property X) and that, ad-
flies
sits in trees
for birds) was judged to be a better ditionally, Animal B possessed another property (e.g.,
and
member of a category than was an item that did not pre- Property Y). Furthermore, the participantswere told that,
has a beak sits in trees
serve such a correlation (e.g.,
and
), for Animal B, there was a relationship between X and Y.
even when individual features in both items were equally The participantsthen had to judge the likelihoodthat An-
rated as belongingto that category.In contrast, no such ef- imal A also possessed Y. The relationship between X and
fect of feature correlationswas observed with implicit fea- Y was described as causal in her Experiment2 and as non-
ture correlations that the participants failed to notice in causal (a temporal relation was used instead) in her Ex-
Experiment 1 and were less likely to label with specific re- periment 3. Only when the causal descriptionwas supplied
lations as found in Experiment 3.
did the inductive strength increase beyond the strength
The present study has important implications for both generated by the other information alone. On the basis of
the theory-based approach and the similarity-based ap- this, it is reasonable to expect that correlations not sup-
proach to categorization. With respect to the similarity- ported by any causal relations would not increase induc-
based approach,the present study offers an important con- tivestrengthsas much. Other works havealso demonstrated
straint. For instance, Gluck and Bower (1988; see also that people’s causal background knowledge determines
Gluck, 1991) describe a configural-cue network model of how important a feature is in categorization (Ahn, Kim,
human learning that represents stimulus patterns in terms Lassaline, & Dennis, 2000), as well as in similarity judg-
of their individual features and pairwise conjunctions of ments(Ahn & Dennis,2001).Takentogether, thesefindings
features. However, as was notedearlier, a problemwith this suggest that explanatoryrelations, especiallycausal ones,
approach is computationalcomplexity;there are too many play a crucial role in various aspects of our concepts.
conjunctionsof features to keep track of as the number of
The results of the above-mentioned studies lead to a
features increases (Medin, 1983). The present results sug- pressing question for future study: Do people explicitly
gest that, as a way to avoid this complexity, people may notice correlations because they can explain them, or do
tend to explicitly encode only those feature conjunctions people impose explanations after they explicitly notice
that they can explain. That is, they do not encode all con- correlations? For example, people might explicitlynotice
beinga tire
black color
because
junctions of features in a content-free manner, as in the a correlation between
configural-cue network model.
the correlation is unusually strong or because somebody
and
These findings also contribute to the theory-based ap- pointedout the correlation to them, but initially,they have
proach by allowingus to examinetheeffect of theory-based no structure bindingthose two propertiesother than a per-
correlations in real-world concepts. Although it is often ceived correlation. Later, they might develop some theo-
difficultto analyzethe complexnetwork of propertiesand ries as to why tires have to be black (e.g., it is less likely
interrelations found in naturally occurring concepts, we to show dirt). On the other hand, it may be that feature cor-
havedemonstratedthatthe influenceof backgroundknowl- relations are explicitly noticed (even when there is no ac-
edge (e.g., theories) does indeed extend beyond the con- tual correlation;Chapman& Chapman,1967)becausethey

116
AHN, MARSH, LUHMANN, AND LEE
are consistent with people’s existing background knowl- Murphy, G. L., & Medin, D. L. (1985). Therole of theories in conceptual
coherence. Psychological Review, 92, 289-316.
Murphy, G. L., & Wisniewski, E. J. (1989).Feature correlations in con-
ceptual representations. In G. Tiberghien (Ed.), Advances in cognitive
science: Vol. 2. Theory and applications(pp. 23-45). Chichester, U.K.:
edge. The latter possibilityappears more likely because of
the computationalcomplexityproblem noted earlier. With
the exception of extremely high or perceptually salient
correlations (e.g., all tires being black), it would be diffi-
cult to notice correlations without any guidance of one’s
background knowledge. The present experiments are not
informative in relation to this issue, because we examined
natural conceptsthat peoplealready possessed and did not
Ellis Horwood.
Rosch, E. (1978).Principles of categorization. In E. Rosch & B. B. Lloyd
(Eds.), Cognition and categorization (pp. 27-48). Hillsdale, NJ: Erl-
baum.
Sloman, S., Love, B., & Ahn, W. (1998). Feature centrality and concep-
tual coherence. Cognitive Science, 22, 189-228.
examinethe processesby which they had come to be aware Thagard,P. (1989).Explanatorycoherence. Behavioral& BrainSciences,
12, 435-502.
of certain feature pairs. A future studyusingartificialstim-
uli with structures similar to natural categories will fur-
ther shed light on this issue.
Wattenmaker, W. D. (1991).Learning modes, feature correlations, and
memory-based categorization. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 17, 908-923.
Wattenmaker, W. D. (1993). Incidental concept learning, feature fre-
REFERENCES
quency,and correlated properties. JournalofExperimentalPsychology:
Learning, Memory, & Cognition, 19, 203-222.
Wellman, H. M. (1990). The child’s theory of mind. Cambridge, MA:
MIT Press.
Ahn, W., & Dennis, M. (2001). Dissociation between categorization
and similarity judgment. In M. Ramscar & U. Hahn (Eds.), Similarity
andcategorization(pp.87-107).New York:Cambridge University Press.
Ahn, W., Kim, N. S., Lassaline,M. E., & Dennis, M. J. (2000).Causal
status as a determinant of feature centrality. CognitivePsychology, 41,
1-55.
Alloy, L. B., & Tabachnik, N. (1984). Assessment of covariation by
humans and animals: The joint influence of prior expectations and
current situational information. Psychological Review, 91, 112-149.
NOTES
1. Each property was weighted by thenumber of category members for
which it appeared, and thefamily resemblance score for each item consisted
of the sum of the weights of its properties.
2. It is also possible that sometimes people might first notice salient
Barrett, S. E., Abdi, H., Murphy, G. L., & Gallagher,J. M. (1993). correlations and then subsequentlydevelop theories about them. See the
Theory-based correlations and their role in children’s concepts. Child
Development, 64, 1595-1616.
Carey, S. (1985). Conceptual change in childhood. Cambridge, MA:
General Discussion section for further discussion.
3. We treat explanationsand domain theories as the same in naive con-
ceptual representations (Murphy, 2000) and use the terms interchange-
ably.
Plenum.
Chapman, L. J., & Chapman, J. P. (1967). Genesis of popular but erro-
neous psychodiagnosticobservations. Journal of Abnormal Psychol-
ogy, 72, 193-204.
4. The term implicit has traditionally been used to designate abilities
or knowledge that cannot be articulated yet could be demonstrated in an
appropriatetask. We used the same term here because participants failed
Cohen, J., MacWhinney, B., Flatt, M., & Provost, J. (1993). to explicitly notice correlations inherent in their own representations
PsyScope: An interactive graphic system for designing and control- when measured in a task employed in Part 1 of Experiment 1 in Malt and
ling experiments in the psychology laboratory using Macintosh com-
puters. BehavioralResearch Methods, Instruments, & Computers, 25,
257-271.
Gluck, M. A. (1991). Stimulus generalization and representation in
adaptive network models of category learning. Psychological Science,
2, 50-55.
Smith (1984).
5. The exceptionsto these will be discussed at the end of Experiment 2.
6. The participants were given differing numbers of explicit and im-
plicit pairs, as defined by our Experiment 1, depending on which exper-
imental form they were given. Fifteen participants received 90 explicit
statements and 142 implicit statements, and 13 participants received 86
explicit statements and 144 implicit statements. These numbers might
Gluck, M. A., & Bower, G. H. (1988). Evaluating an adaptive network
modelof human learning. JournalofMemory& Language, 27, 166-195. notappear to sum up to the total number of statements, butthis is because
Keil, F. C. (1981). Constraints on knowledge and cognitive develop- some of the statements were presented to both groups of participants.
ment. Psychological Review, 88, 197-227.
Keil, F. C. (1989). Concepts, kinds, and cognitive development. Cam-
bridge, MA: MIT Press.
The followingbreakdowngives the number of pairs in each category, with
the parentheses representing the second form of the experiment: Bird—
18 (17) explicit, 19 (19) implicit; Clothing,— 4 (3) explicit, 17 (18) im-
Lassaline,M. E. (1996). Structuralalignment in inductionand similarity. plicit; Flower—3 (4) explicit, 2 (5) implicit; Fruit—2 (2) explicit, 4 (3)
Journalof Experimental Psychology:Learning,Memory, & Cognition,
22, 754-770.
Malt, B. C., & Smith, E. E. (1984). Correlated properties in natural cat-
implicit; Furniture—13 (13)explicit, 12 (12) implicit; Tree—5 (6) explicit,
15 (15) implicit.
7. This analysis does not capture cyclic dependency relations (e.g., A
egories. Journal of Verbal Learning & Verbal Behavior, 23, 250-269. causes B and B also causes A) as theory-based relations because they
Medin, D. L. (1983).Structural principlesin categorization. In T. J. Tighe will result in small difference scores. Yet, a maximum score should have
& B. E. Shepp (Eds.), Perception, cognition,and development: Interac-
tional analyses (pp. 203-230). Hillsdale, NJ: Erlbaum.
Medin, D. L., Altom, M. W., Edelson, S. M., & Freko,D. (1982).Cor-
related symptoms and simulated medical classification. Journalof Ex-
captured this as a theoretical relation because, in this case, A depends on
B, as well as B depending on A.
8. Negative relations were not used because, by definition, they will
detrimentally lower typicality judgments. Malt and Smith (1984) do not
perimental Psychology: Learning, Memory, & Cognition, 8, 37-50. describe exactly how negative correlations were treated in their Experi-
Medin, D. L., & Ortony, A. (1989).Psychologicalessentialism. In S. Vos-
ment 2. The most sensible thing to do is to negate one of the features to
niadou & A. Ortony (Eds.), Similarity and analogical reasoning make the correlation positive, but that could make a pair identical to one
(pp. 179-196). New York: Cambridge University Press.
Murphy, G. L. (2000).Explanatoryconcepts. In F. C. Keil & R. A. Wilson
(Eds.), Explanation and cognition (pp. 361-392). Cambridge, MA:
MIT Press.
of the other existing pairs (i.e., can’t fly, sits in trees becomes flies, sits
in trees, an already existing pair).
9. These weights were previously not published and were provided to
the present authors by Barbara Malt.

FEATURE CORRELATION
117
APPENDIX A
The following table provides a list of all feature pairs rated as explicitly related in Experiment 1. The labels most often used for
each feature pair in Experiment 3 are provided, along with the number of participantswho used that label in parentheses.
Mean Ratings
on Relatedness
Commonly Used
Labels
Category
Bird
Feature 1
Feature 2
is a bird
is a bird
has feathers
has a beak
is a bird
2.90
2.75
2.65
precondition(10), causes/allows (4)
causes/allows (9), precondition(8)
causes/allows (12)
lays eggs
is large
is large
flies
has a large beak
has large wings*
has a large beak
sits in trees*
eats fish*
2.35
2.20
2.10
1.95
causes/allows (10)
causes/allows (8)
causes/allows (15)
causes/allows (17)
eats fish
has feathers
has legs
has large wings
has a beak
has feathers
is small
is small
is small
is small
can’t fly
lives near the ocean*
has a beak
sits in trees
has a large beak
lays eggs
1.90
1.90
1.75
1.70
1.70
1.65
1.60
21.60
21.75
21.80
22.05
22.40
causes/allows (15)
causes/allows (6)
causes/allows (8), precondition(4)
increases (5)
causes/allows (5)
no common label
causes/allows (6)
decreases/minimizes (7)
lays eggs
eats insects
has a large beak
is large
has large wings
flies
no common label
decreases/minimizes (9)
disables (6)
can’t fly
sits in trees
decreases/minimizes (8), disables (4)
Clothing
is wool
is warm*
is material*
is worn in bad weather
is comfortable*
is material*
2.50
2.20
2.15
2.00
1.65
causes/allows (7), increases (6), subset/example (5)
subset/example (7), precondition(4)
causes/allows (10)
causes/allows (10), increases(4)
subset/example (16)
is clothing
is warm
is cotton
is cotton
Flower
Fruit
has a long stem
is small
is tall
is tall*
has a long stem
is small
2.40
21.80
21.80
causes/allows (17)
decreases/minimizes (6)
no common label
is sweet
is juicy
tastes good
tastes good*
2.35
1.90
causes/allows (15)
causes/allows (11), increases(6)
Furniture is soft
has cushions
is comfortable*
is for sitting on*
is soft*
is comfortable*
is for sitting on*
is for sitting on*
has cushions
is electric*
2.60
2.40
2.40
2.15
2.15
2.00
1.90
1.70
causes/allows (7)
causes/allows (11), subset/example (4)
causes/allows (11), increases(4)
causes/allows (14)
causes/allows (9), increases(4)
causes/allows (6), increases (4), precondition (4)
causes/allows (8), precondition(4)
causes/allows (11)
has cushions
has cushions
is comfortable
is soft
has springs
has a cord
has springs
has springs
has cushions
is comfortable
is for sitting on
is on the wall
is metal
is comfortable*
is for sitting on*
is on the wall
is on the wall
is on the wall
is with a chair
is soft
1.60
1.55
21.50
21.60
21.70
21.75
22.70
causes/allows (9), increases(4)
causes/allows (8)
no common label
no common label
no common label
no common label
decreases/minimizes (7)
Tree
is big
is shady*
is shady*
has bark*
is big*
is tall*
is big*
2.30
2.20
2.15
2.15
1.90
1.85
1.65
causes/allows (11), increases(5)
causes/allows (15)
causes/allows (8), precondition(5)
causes/allows (8), subset/example (7)
causes/allows (10)
has branches
has a trunk
is tall
is old
is old
causes/allows (11), increases(4)
causes/allows (9)
has needles
is green
Note—Labels were condensed into similar groups for ease in interpretation.These groups were collapsed as follows: precondi-
precondition example/subset decreases/mini-
tion, necessary,needs, requires as
; example, subset, type as ; decreases,minimizes as
; causes, enables, allows, because, is due to, depends on, in order to, result of, deter-
mizes
disables
; disables, inhibits, prevents as
causes/allows
increases
mines, affects as
. The label
did not include multiple labels. *These pairs were used for the construction
of exemplars in Experiment 4.

118
AHN, MARSH, LUHMANN, AND LEE
APPENDIX B
Constant Features of Six Categories
Used in Experiment 4
Category
Constant Features
Bird
Clothing
has feathers, has claws, lays eggs
can be bought, can be made, is used
by humans
Tree
has leaves, manufactures chlorophyll,
grows outdoors
Furniture
Fruit
is found in houses, is man-made, can
be bought in department stores
is nutritious, is grown in warm climates,
is eaten
Flower
has petals, grows in soil, is alive
APPENDIX C
Implicit Pairs Used in Experiment 4
Bird
has a beak–is small
has feathers–is black
is gray/white–eats fish
has a large beak–has legs
Clothing
is worn in bad weather–is comfortable
has a zipper–is colored
has buttons–is colored
has long sleeves–is colored
Fruit
tastes good–is a summer fruit
Flower
is yellow–has many petals
Tree
grows in a warm climate–has bark
grows in a warm climate–has a trunk
is shady–has a trunk
has bark–is tall
has bark–is big
has a trunk–is old
Furniture
is metal–has a dial
has a flat top–is with a chair
has four legs–is comfortable
is metal–is electric
has a dial–is electric
is wood–has a flat top
has a dial–has a cord
has four legs–is for writing/working
has four legs–is with a chair
(Manuscript received October 23, 2000;
revision accepted for publication October 9, 2001.)