ArticlePDF Available

Abstract and Figures

Three experiments tested the hypotheses that while forming stereotypes of social groups, people abstract the central tendency and variability of different attribute dimensions to determine which ones best differentiate the groups and that more differentiating dimensions are more likely to become stereotypical in the sense of becoming strongly associated with the groups in memory. Supporting these hypotheses. Experiment 1 found that, after viewing behaviors performed by members of 2 groups, Ss characterized the groups more in terms of attribute dimensions indicating larger differences between the central tendencies of the groups, and Experiment 2 showed that this effect did not occur when Ss formed impressions of only 1 group. Experiment 3 found that Ss also characterized groups more in terms of attribute dimensions indicating lower within-group variability.
Content may be subject to copyright.
Journal of Personal ily and Social Psychology
1992,
Vol. 63, No. 3, 356-367Copyright 1992 by the American Psychological Association, Inc.
0022-3514/92/$3.00
The Role of Diagnosticity in Stereotype Formation:
Perceiving Group Means and Variances
Thomas E. Ford and Charles Stangor
University of Maryland, College Park
Three experiments tested the hypotheses that while forming stereotypes of social groups, people
abstract the central tendency and variability of different attribute dimensions to determine which
ones best differentiate the groups and that more differentiating dimensions are more likely to
become stereotypical in the sense of becoming strongly associated with the groups in memory.
Supporting these hypotheses, Experiment
1
found that, after viewing behaviors performed by
members of
2
groups,
Ss characterized the groups more in terms of attribute dimensions indicating
larger differences between the central tendencies of the
groups,
and Experiment
2
showed that this
effect did not occur when Ss formed impressions of only
1
group. Experiment 3 found that Ss
also characterized groups more in terms of attribute dimensions indicating lower within-group
variability.
Social stereotypes have traditionally been regarded as fea-
tures that are descriptive of social groups. However, both the
theoretical question of exactly what makes an attribute stereo-
typical of a given group and the methodological question of
how to measure stereotypes continue to be debated (e.g., Ash-
more & Delboca, 1979; Brigham, 1971; Gardner, Kirby, &
Findlay, 1973; Gardner, Lalonde, Nero, & Young, 1988). For
instance, Katz and Braly
(1933)
considered attributes that were
most commonly used to describe a group on a checklist mea-
sure to be those that were stereotypical of that group. Others
(e.g., Brigham, 1971) have assumed that group stereotypes are
represented by the percentage of group members who are
thought to possess a given
attribute.
Stereotypes
have also
been
denned in terms of the "amount" of a given attribute possessed
by individual members of a group and have been measured
using Likert-type scales
(e.g.,
Sigall &
Page,
1971). These tradi-
tional methods of measuring group stereotypes are similar in
that respondents are asked to consider a number of attributes
and judge the value of each for describing a specified group.
Thus,
these methods
assess
the attributes that are considered to
be descriptive of the relevant social group.
More recently, theorists have turned their attention away
from identifying the attributes that are used to describe social
groups and have begun to develop models of
how
stereotypes
are represented in memory. These models assume that stereo-
types are
cognitive representations or knowledge structures that
link descriptive attributes to the group
(e.g.,
Dovidio, Evans, &
Tyler, 1986; Gaertner & McLaughlin,
1983;
Taylor & Crocker,
This research was presented at the third annual meeting of the Amer-
ican Psychological Society as part of the student caucus
research
award
given to Thomas E. Ford.
This research was supported in part by National Science Founda-
tion Research Grant BNS-8818926 to Charles Stangor.
Correspondence concerning this article should be addressed to
Thomas E. Ford or Charles Stangor, Department of Psychology,
University of Maryland, College Park, Maryland 20742.
1981).
The idea that stereotypes are represented
by
associations
between different attributes and a group in memory
is
valuable
because it allows being more specific about the underlying
structure of a group stereotype as well as about expected rela-
tionships among different methods of measuring stereotypes.
Within
this cognitive
representational framework,
the
"tradi-
tional" methods of measuring stereotypes described earlier
as-
sess the attributes that are available within one's representation
of
a
given social
group.
However, the attributes contained in the
representation may also be expected to vary in how strongly
associated they are with the group label and thus how impor-
tant they are in defining the group
stereotype.
The importance
of strength of association for defining a group stereotype was
well articulated by Hamilton and Trolier
(1986)
in a discussion
concerning the content of racial stereotypes. They noted that
"attributes that are strongly associated with a particular racial
concept, then, would compose the content of the stereotype
about that
group.
In contrast, content that is less strongly asso-
ciated with the concept presumably
is less
central to the stereo-
type"
(p. 141).
Conceptualizing stereotypes in terms of the attributes that
are relatively strongly associated with social groups in memory
has a particular advantage in that
it
provides a theoretical frame-
work, on the basis of the principles of construct accessibility,
for explaining the influence of stereotypes on person percep-
tion.
Stereotypes,
like
other knowledge
structures,
mediate per-
son perception in important ways. They influence how we in-
terpret others' behavior (e.g., Duncan, 1976), what we re-
member about others
(e.g.,
Cohen,
1981;
Stangor & McMillan,
1992),
and how
we
behave toward others
(e.g.,
Snyder,
Tanke,
&
Berscheid,
1977;
Word, Zanna, & Cooper, 1974). But which of
the attributes that are descriptive of a group are likely to be
most important in this regard? Attributes that
have
strong
asso-
ciative links to a given group in memory will be more likely to
be activated or "come to mind" when one thinks about the
group or encounters an individual who is a member of that
group. Constructs that most readily come to mind are known
356
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
DETECTING DIAGNOSTIC ATTRIBUTES357
to be particularly important in influencing judgments (e.g.,
Higgins,
Rholes,
&
Jones,
1977)
and behaviors
(Neuberg,
1988).
Thus,
because attributes that are strongly connected to the
groups in memory are likely to be highly accessible when inter-
acting with group
members,
they can
be
expected
to
play
a
large
role in person perception.
In the present research, we considered stereotypes in terms
of the strength of association between attributes and social
groups. Thus, we measured stereotypes by assessing the attrib-
utes
that
come
to mind
when
thinking about
those
groups.
This
method of measuring stereotypes may be expected to reveal a
different pattern of results than traditional methods because
although many attributes may be available in memory to de-
scribe a social
group,
it is the attributes that
have
strong associa-
tive links with the group that will spontaneously come to mind
when one thinks about the group.
The formation of stereotypes is the major focus of this article.
That is, which of the many attributes that are descriptive of
a
group become most strongly associated
with
the group
in
mem-
ory? One commonly proposed explanation for the ubiquity of
social stereotypes is that stereotypes simplify the processing of
interpersonal information and facilitate judgments of individ-
uals
(e.g.,
Allport,
1954;
Ashmore
&
Delboca,
1981;
Hamilton,
1979;
Taylor & Crocker, 1981). To the extent that stereotypes
fill such a simplifying function, they should be based on attrib-
utes that differentiate groups from each other, because differ-
entiating attributes would be more useful for making judg-
ments based on group membership than would less differen-
tiating attributes.
For example, a stereotype of Russians based on an attribute
such as
intelligent
would not be very useful for making judg-
ments if many other ethnic groups were also thought to be
intelligent. This is because the intelligence stereotype would
not provide a unique set of expectations about individuals
based on their group membership. However, a stereotype of
Russians based on attributes such as stem or drinks vodka
would be more useful because these attributes are not only
representative of Russians but
also
distinguish them from other
ethnic groups.
The assumption that group differentiation is an important
criterion for stereotypes has been demonstrated in research by
McCauley and Stitt (1978). McCauley and Stitt developed a
"diagnostic ratio" measure of group stereotypes obtained by
dividing, for each subject, the probability that a given attribute
is
possessed
by
a member of a particular group by the probabil-
ity that the attribute is possessed by people in general. They
found that attributes that differentiate members of a social
group from people
as a whole
are more likely than
less
differen-
tiating attributes to be seen as stereotypical of that group. On
the
basis
of these
results,
McCauley,
Stitt,
& Segal (1980)
argued
that stereotypes can be denned as "those generalizations about
a class of people that distinguish them from others" (p. 197).
A fundamental assumption of the present research is that
diagnostic
attributes (i.e, those that differentiate groups) are
more likely to become stereotypical of a social group than are
less diagnostic, but also applicable, attributes. However, this
article goes beyond the issues addressed by McCauley and Stitt
(1978)
by focusing on how the relative diagnosticity of different
attributes is detected in the social world; that is, the role of
diflFerentiation in stereotype formation. One possibility is that,
through contact with members of different groups, the social
perceiver infers the diagnosticity of different attributes from
the distributional qualities (i.e, central tendency and variabil-
ity) of the different attributes.
That the means and variances
along
certain dimensions pro-
vide diagnostic information about groups
is well
known. Many
statistical methods
are
based on the
use
of means and variances
to provide information about group differences. The
analysis of
variance (ANOVA) F ratio can be thought of as an index of
diagnosticity. A large F ratio indicates that two or more groups
are different from one another
on
some measure, and a small F
ratio indicates that the groups are not different from one an-
other on that
measure.
The
size
of the F ratio (and therefore the
diagnosticity of a particular measure) is jointly determined by
the
size
of the difference between group means and the amount
of variability within the groups.
The idea that people are sensitive to group means and vari-
ances underlies many theories of categorization, both of natu-
ral objects
(e.g.,
Rosch, 1978) and of social groups
(e.g.,
Turner,
1987).
Turner
(1987),
for
example,
suggested that social catego-
rization "follows the principle of
meta-contrast"
(p. 46). This
principle suggests that a collection of individuals is more likely
to be considered as a group to the extent that the mean differ-
ence between that set of individuals and other sets of individ-
uals along
some
relevant dimension
is
large relative to the vari-
ability of that dimension within the
sets.
The metacontrast ra-
tio,
then, is like an F ratio in that it is a ratio of the difference
between groups over the variability within groups and can be
interpreted as an index of diagnosticity for categorization.
This research extends this interpretation of
the
principle of
metacontrast to the issue of stereotype formation. The predic-
tions are (a) that attribute dimensions that best differentiate
social groups are likely to become stereotypical of those groups
in
the sense
that they become relatively strongly associated with
the group in memory and
(b)
that the social perceiver considers
both group means and group variances with respect to differ-
ent attribute dimensions in forming diagnostic group stereo-
types.
In this
article,
we present three experiments designed to
examine the use of group means and variances in stereotype
formation. Experiment
1
and Experiment
2
investigate whether
the difference between group means on different attribute di-
mensions influences formation of group stereotypes. Experi-
ment
3
investigates whether the within-group variance on dif-
ferent attribute dimensions influences formation of
group
ste-
reotypes.
Experiment
1
It is well documented that people abstract central tendency
information to cognitively represent social categories. Posner
and Keele (1970) and Rothbart and John (1985), for example,
proposed prototype models of social category representation.
According
to
these models, the perceiver abstracts information
regarding category members to form a unitary representation
or prototype of the category. The prototype represents the cen-
tral tendency of relevant attributes across the known category
members.
In one relevant study, Krueger, Rothbart, and Sriram (1989)
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
358THOMAS E. FORD AND CHARLES STANGOR
reported that subjects used central tendency information
to
dis-
tinguish between two groups and were sensitive to changes in
central
tendency.
Moreover, subjects tended to overestimate the
extent to which new information increased this intergroup dis-
tinction. This research indicates that people differentiate
groups
by
noting the central tendencies of group-relevant
attrib-
utes
and process new information in such a
way as
to maximize
mean differences between groups.
Given that people can infer central tendencies of groups with
respect to relevant attribute dimensions, it is possible that peo-
ple also use differences in central tendencies among groups on
different attribute dimensions to determine which dimensions
are most useful for characterizing the groups. Experiment 1
tested the hypothesis that, because a large mean difference
between groups with respect to an attribute dimension indi-
cates greater diagnosticity than smaller mean differences,
group stereotypes, as assessed by strength of association, are
more likely to be based on attribute dimensions for which the
difference between group means is large than on dimensions
for which the mean difference is small.
Method
Overview. Subjects were presented with behaviors performed by
members of two fictitious groups and asked to form impressions of
each
group.
Each group performed behaviors related to both friendli-
ness and
intelligence.
The behaviors were selected such that the groups
differed in terms of both attribute dimensions. However, for one half
of the subjects, the difference between group means was larger for
intelligence than for friendliness, whereas this pattern was the oppo-
site for the other half of the subjects.
We
predicted that subjects, when
asked to freely characterize the two groups, would use the attribute
dimension that indicated the larger between-groups difference, show-
ing that the dimension had become more strongly associated with the
groups in memory.
Subjects. Subjects were
12
male and
29
female introductory psychol-
ogy students who participated in the experiment in exchange for extra
credit in their psychology
courses.1
Subjects were randomly assigned
to one of
two
experimental conditions (larger mean difference on in-
telligence or larger mean difference on friendliness).
Stimulus
materials.
Booklets containing behavioral descriptions,
similar
to
those used by Hastieand Kumar
(1979),
Srull,
Lichtenstein,
and Rothbart (1985) and Stangor and Duan (1991), were prepared-
Each behavior was typed on a single page along with the name of
a
social group identified
by
color.
For
example,
one page (indicating
low
intelligence) said "A member of the blue group failed his written
driver's test for the fourth
time."
Each target group (the
blue
group and
the red group) was described by 18 different behaviors: 6 related to
friendliness, 6 related to intelligence, and 6 that were not related to
either friendliness or intelligence.
Each group
was
described more positively than the other
in
terms of
one attribute dimension and more negatively in terms of the other
dimension. The blue group always performed behaviors that were
more friendly than those performed by the red group and behaviors
that were less intelligent than those performed by the red group. The
behaviors were selected on the basis of pretest ratings given by 22
subjects from the same population (but who did not participate in the
experiment) such that the mean difference between the Hue group and
the red group along each attribute dimension varied by condition.
These ratings were made on
9-point
tcales (e.g., 1 «= extremely un-
friendly;
9 =
extremely
friendly).
In the
larger-mean-difference-on-intelligence
condition, the behav-
iors were selected to create a large mean difference between the two
groups in terms of intelligence
(Ms =
2.2 for the blue group and 7.6 for
the red group), F(l, 20)
=
425.60, p
<
.001,
and a smaller mean differ-
ence
in
terms of friendliness (Ms=6.8 for the blue group and 2.8 for the
red group), F(l, 20) = 244.10, p <
.001.
The difference between the
groups was significantly greater for the intelligence attribute dimen-
sion, F(l,
20) =
13.01,
p
<
.002.
In the
larger-mean-difference-on-friend-
liness condition, the behaviors were selected to create a large mean
difference between the two groups in terms of friendliness
(Ms =
7.8
for the
blue
group and
1.8
for the red
group),
F(l,
20) =
548.72,
p
<
.001,
and a smaller mean difference in terms of intelligence
(Ms
=3.4 for the
blue group and 6.5 for the red group), F(l,
20) =
149.44, p<
.001.
The
difference between the groups
was
greater for the friendliness attribute
dimension, F(l, 20) = 119.09, p
<
.001.
To ensure that mean differences between groups were not con-
founded with the amount of within-group variability,
we
held the vari-
ability of the behaviors describing the groups along each attribute di-
mension, also determined on the basis of pretesting, constant. In the
larger-mean-difference-on-intelligence condition, the standard devia-
tion of mean ratings for behaviors describing the blue group was 0.33
on the intelligence dimension and
0.35
on the friendliness dimension.
The standard deviation of mean ratings for behaviors describing the
red group was 0.39 on the intelligence dimension and 0.59 on the
friendliness dimension. Bartlett-Box F tests of homogeneity of vari-
ance indicated that the variability of group behaviors along the two
dimensions was not significant for the blue group, F(l, 10) = .02, p =
.89,
or for the red group, F(l, 10) = .76, p = .38.
In the larger-mean-difference-on-friendliness condition, the stan-
dard deviation of mean ratings for behaviors describing the blue group
was 0.37 on the intelligence dimension and 0.20 on the friendliness
dimension. The standard deviation of mean ratings for behaviors de-
scribing the red group
was
0.34 on the intelligence dimension and 0.33
on the friendliness dimension. The difference in the variability of
group behaviors along the two dimensions was not significant for the
blue group, F(l,
10)
= 1.44, p
=
.23,
or for the red
group,
F(l,
10)
= .08,
p =
.93.
The results of the
tests
of homogeneity of variance suggest that
behaviors describing each group along the larger mean difference at-
tribute dimension were not associated with smaller within-group vari-
ances.
The order of the behaviors appearing in the stimulus booklets was
controlled through a blocking procedure. First, we randomly assigned
behaviors to blocks of six behaviors, each consisting of one behavior
from the friendliness dimension, one behavior from the intelligence
dimension, and one irrelevant behavior for each group. Booklets con-
taining six blocks of behaviors were constructed. The behaviors within
each block were randomized, and the order in which the blocks ap-
peared in each booklet was randomly determined for each subject.
This blocking procedure ensured that the serial position of each type
of behavior for each group was approximately equal and that each
behavior appeared approximately equally often in each position.
Procedure. Subjects completed the experiment individually or in
small groups of up to 4. All subjects were told that the experiment
concerned impression formation based on limited information and
that they would be given some information about people from differ-
ent groups and later asked to make some judgments about them. Sub-
jects were told that the groups they would be reading about were real
social groups such as clubs and fraternities from a large university in
the Midwest who had been observed in another study conducted at
that
university.
Subjects
were
told that for the purposes of the study the
groups would be labeled with colors and that each behavior had been
1 There
were no
meaningful sex-of-subject effects in any of the exper-
iments reported in this article, thus all analyses are collapsed across
this variable.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
DETECTING DIAGNOSTIC ATTRIBUTES359
performed by a different member of the group. All subjects were told
to read each of the behaviors carefully and to form an impression of
what each group
was
like.
The experimenter timed the subjects
as
they
read the behaviors at the required rate of 8 s per behavior.
A fter viewing the
behaviors,
subjects completed the dependent mea-
sures.
First came the primary stereotyping measure. Because
we
con-
ceive of stereotypes
as
being comprised of attributes that are relatively
strongly associated with a particular
group,
our group stereotype mea-
sure
was designed to assess the associative strength between the groups
and the descriptive attributes. Subjects were given a sheet of paper
containing the names of the two groups they read about, with four
boxes under each group name. Subjects characterized the two groups
by writing one thought about what the blue group was like in each of
the four
boxes
appearing below the blue group and one thought about
what the red group
was
like in each of the four
boxes
appearing below
the red
group.
The number of thoughts for each group pertaining to a
given attribute dimension provided an index of the strength of associa-
tion between that dimension and the group labels.
Next, subjects were given a free-recall measure. They were given
sheets of paper with the names of each of the groups they read about
listed on the top and were asked
to
recall
as
many behaviors performed
by each group as possible. Then, on a final page, subjects rated each
group on both of the attribute dimensions (friendliness and intelli-
gence) using
9-point
Likert-type scales
(e.g.,
1
=
extremely
unfriendly;
9 =
extremely
friendly).
Subjects were then debriefed and dismissed.
Results
Attribute
ratings.
It was expected that the ratings of each
group in
terms
of intelligence and friendliness would
show
that
subjects perceived the groups
to
differ on both attribute dimen-
sions and that the central tendencies of the two groups would
be perceived
to
differ
more
on the larger mean difference
attrib-
ute dimension.
Planned comparisons indicated that, in the
larger-mean-dif-
ference-on-intelligence condition, the red group was rated as
more intelligent (M= 8.2) than the blue group (M
=
2.7), F(l,
18) =
127.01,
p < .001, and the red group was rated as less
friendly (M = 3.3) than the blue group (M = 7.2), F(l, 18) =
63.36,
p < .001. These tests indicated that subjects distin-
guished the two groups along both the intelligence and the
friendliness attribute dimensions. Also, as expected, subjects
perceived a greater difference between the two groups in terms
of intelligence rather than friendliness,
F(l
,18)=
9.79,
p
=
.006.
In the larger-mean-difference-on-friendliness condition, the
red group was rated as less friendly (M
=
3.2) than the blue
group (M= 8.0), F(l, 21) = 141.16, p
<
.001,
and the red group
was
rated
as
more intelligent (M
=
7.2)
than the
blue
group (M=
4.1),
F(l, 21) = 39.69, p < .001. Also, as expected, subjects
perceived the groups as more different in terms of friendliness
than intelligence, F(l, 21) = 9.06, p < .007. Taken together,
these tests indicate that, as expected, subjects distinguished
between the two groups along both the intelligence and the
friendliness attribute dimensions.
Associative
strength.
The associative strength measure con-
sisted of the thoughts generated for each
group.
The major pre-
diction of this experiment was that attribute dimensions that
best differentiate between the two groups would become more
strongly associated with the groups than less differentiating
attribute dimensions. Therefore, it was expected that when
asked to list thoughts about the two groups, subjects would list
more thoughts pertaining to the attribute dimension that indi-
cated a greater mean difference between the two groups.
The experimenter and an independent judge
(an
undergradu-
ate
assistant),
each blind
to
condition, coded
the
thoughts listed
for each group by each subject. Thoughts were coded as being
either associated with friendliness, associated with intelligence,
or irrelevant to both attribute dimensions.2 Agreement be-
tween the judges
was
96%
(Cohen's kappa
=
.83),
and disagree-
ments were decided by the experimenter. Because attribute-
irrelevant thoughts are not theoretically interesting, they were
not included in subsequent analyses. Because the hypothesis
suggests that one attribute dimension will be used more to
describe both groups, depending on which dimension indi-
cated a larger between-groups difference, thought listings for
the blue group were combined with thought listings for the red
group.
The mean number of thoughts pertaining to intelligence and
friendliness generated in each condition are shown in Figure
1.
In the larger-mean-difference-on-intelligence condition, there
were a greater number of thoughts generated pertaining to in-
telligence (M=
3.63)
than pertaining to friendliness (M = 2.05).
In the larger-mean-difference-on-friendliness condition, there
were a greater number of thoughts generated pertaining
to friendliness (M = 4.13) than pertaining to intelligence
(M=1.27).
A 2
(larger mean difference dimension)
X 2
(generated attrib-
ute dimension) ANOVft was conducted. The only significant
effect was the interaction, F(l,
39)
= 42.29, p
<
.001,
indicating
that, as predicted, subjects generated more thoughts about the
groups in terms of the attribute dimension that represented
larger differences in central tendency between the two groups.
Planned comparisons revealed that in the larger-mean-differ-
ence-on-intelligence condition, the number of thoughts gener-
ated to describe each group in terms of intelligence
was
signifi-
cantly greater than the number generated to describe them in
terms
of friendliness, F(l,
39) =
10.84,
p
<
.004.
Likewise, in the
larger-mean-difference-on-friendliness condition, there were
significantly more thoughts generated to describe the groups in
terms of friendliness than in terms of intelligence, F(l, 39) =
47.01,
p< .001.
Free
recall.
One explanation of the above results is that sub-
jects abstracted the utility of
the
two attribute dimensions for
differentiating the groups and found the larger-mean-differ-
ence attribute dimension more useful for characterizing them.
However, it is possible that because the behaviors depicted in
the larger-mean-difference attribute dimension were more ex-
treme, they might have been more memorable and thus more
likely to be used in the characterizations (Rothbart, Fulero,
Jensen, Howard, & Birrell,
1978;
Thompson, Reyes, & Bower,
1978).
To test this explanation, we examined recall for behaviors
describing each group. Subjects in the larger-mean-difference-
on-intelligence condition recalled more behaviors related to
intelligence (M
=
2.00) than behaviors related to friendliness
2
The valence
of
the
generated
thoughts
(i.e.,
friendly
vs.
unfriendly)
was not of
interest.
Both thoughts are equally related to the friendli-
ness
attribute dimension and
were
coded
as
such.
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
360THOMAS E. FORD AND CHARLES STANGOR
Mean Number of Generated Thoughts
Intelligence Friendliness
Larger Mean Difference
3 Intelligence CD Friendliness
Figure
1.
Mean number of generated thoughts pertaining to intelli-
gence and
friendliness in each condition of Experiment
1.
(M
=
1.61).
In the larger-mean-difference-on-friendliness con-
dition, subjects recalled more behaviors related to friendliness
(M =
2.22)
than behaviors related to intelligence (M
=
1.64).
A 2
(larger-mean-difference attribute dimension)
X
2 (attribute di-
mension recalled) ANOVA was conducted on the mean num-
ber of
behaviors
recalled. The Larger-Mean-Difference Attrib-
ute Dimension
X
Generated Attribute Dimension interaction
was significant, F(l, 39) = 14.80, p < .001.
These results suggest that the observed effects
may
merely be
due to superior memorability of the behaviors from the larger-
mean-difference attribute dimension. However, when an analy-
sis of covariance (ANCOV\) was conducted on the associative
strength measure with the number of intelligence- and friendli-
ness-related behaviors recalled serving as a covariate, the
Larger-Mean-Difference Attribute Dimension
X
Generated At-
tribute Dimension interaction remained significant, F(l, 38) =
34.56,
p
<
.001.
In addition, two path analyses were conducted
to evaluate the hypothesis that memory played a mediating role
in the thought listings. The first path model posits that the
manipulation of the larger mean difference attribute dimen-
sion has a direct effect on thought listings and that this effect is
not mediated by memorability. Thus, the direct path from the
attribute dimension manipulation to the thought listings was
estimated, whereas the path from the attribute dimension ma-
nipulation to memory for group behaviors and the path from
memory for group behaviors to the thought listings were each
constrained to be 0. As expected, the path from the attribute
dimension manipulation to the thought listings
was
significant
(y = .71,p<.05).
Compare this model with one that requires the effect of the
attribute dimension manipulation to be mediated by better
memory for that dimension. For this model, the path from the
attribute dimension manipulation to memory for group behav-
iors and the path from memory for group behaviors to the
thought listings were each estimated, whereas the direct path
from the attribute dimension manipulation to the thought list-
ings was constrained to be 0. Neither the path from the attrib-
ute dimension manipulation to memory for group behaviors
(y
=
-.012) nor the path from memory for group behaviors to
the thought listings (/? = .016) was significant (p
<
.50 for each
path coefficient).
Goodness-of-fit indices indicated that assigning memory for
group behaviors as a mediator of the effects of the attribute
dimension manipulation on the thought listings resulted in a
significant reduction in model fit. The first model, which pos-
tulated only a direct effect, provided an accurate and parsi-
monious explanation of the results, x2(2, N
=
41) = .05, p =
.975.
The second model, however, provided a much poorer ex-
planation, x2(l, N = 41) =
315.51,
p < .001. The difference
between these two models
is
significant, x2G,
JV=
41)
=
315.46,
p<.00\.
Discussion
Results of the attribute ratings indicate that subjects per-
ceived the groups to differ along both attribute dimensions.
The blue group was perceived as significantly less intelligent
and significantly more friendly than the red group in both con-
ditions. Thus, both dimensions
were
available in
the
group rep-
resentations. Nevertheless, when asked to give spontaneous
characterizations of the groups, subjects characterized the two
groups more in terms of the larger-mean-difference attribute
dimension, as predicted. Overall, these results suggest that at-
tributes are likely to become relatively strongly associated with
social groups in memory to the extent that they differentiate
between groups.
Our proposed explanation for these findings is that subjects
used the size of
the
differences between group means with re-
spect
to
each attribute dimension
as
an indication of their
diag-
nosticity or utility for differentiating between the two groups.
To the extent that a certain attribute dimension differentiates
between
groups,
a
stereotype based on that attribute dimension
defines a set of unique expectations about group members and
therefore facilitates inferences about individuals on the basis of
group membership.
The results of the
ANCOVA
on the group stereotype measure
with recall as a covariate cast doubt on an alternative explana-
tion in terms of memorability. If the greater use of the larger
mean difference attribute dimension was due to better memory
for those behaviors, then it would be expected that the Larger-
Mean-Difference Attribute Dimension
X
Generated Attribute
Dimension interaction would be diminished when controlling
for memory. The results indicate that this interaction
is
present
even with memory held constant. Consistent with these results,
the path analyses indicate that memorability did not mediate
the effect of the attribute dimension manipulation on group
stereotypes.
Experiment 2
Another way to test the hypothesis
that group
differentiation
is the
crucial determinant of stereotype formation
is
to provide
conditions in which there would or would not be a motivation
to form diagnostic stereotypes
to
distinguish groups from each
other. One way to accomplish this would be to provide condi-
tions in
which
subjects
learned about only
one
group.
In
Experi-
ment 2, conditions were included in which subjects received
information regarding the friendliness and intelligence of
only
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
DETECTING DIAGNOSTIC ATTRIBUTES361
one
group,
as well as
conditions in which subjects learned about
two
groups.
In both
cases,
the central tendency of the group
(or
groups) was more extreme on friendliness than on intelligence.
If differentiation is critical, then subjects should characterize
the groups more in terms of friendliness only in the two-group
condition, because there is no need to differentiate groups in
the one-group condition.
A potential interpretive problem of Experiment
1 is
that
sub-
jects were asked simultaneously to describe both groups. It is
possible that asking subjects to characterize each group on the
same measure at the same time might have artificially encour-
aged them to compare the two groups. Subjects might have
relied on the larger mean difference attribute dimension to
characterize
the
groups because they felt compelled
to
compare
the groups. To help rule out this potential artifact, in the two-
group condition, we asked subjects to express their thoughts
about each group separately, thus reducing the possibility of
experimentally induced comparison at the time of judgment.
Method
Overview.
The procedure for Experiment
2 was
similar to that used
in Experiment 1. Subjects read behaviors denoting intelligence and
friendliness performed
by
members of either
(a)
the
blue
group and the
red group, (b) the blue group alone, or (c) the red group alone. The
behaviors were the same as those used in the larger-mean-difference-
on-friendliness condition of Experiment 1. After reading the behav-
iors,
all subjects were asked to express their thoughts about what each
group they had read about was like.
Subjects. Subjects were 19 male and 25 female students in lower
level psychology courses
who
participated in exchange for extra credit
in their
classes.
Subjects were randomly assigned to either
(a)
the blue-
and-red-group condition, (b) the blue-group-only condition, or
(c)
the
red-group-only condition.
Procedure. The procedure
was
the same
as
in Experiment
1,
except
that subjects completed the stereotyping measure separately for each
group. After reading the behaviors describing the groups, subjects in
the two-group condition were asked to characterize one of the two
groups by writing their thoughts
in
the four boxes that appeared on the
page.
Only after that page
was
completed
were
subjects given a second
form asking them to characterize the other group. The order of the
thought listings was counterbalanced such that half of the subjects
listed their thoughts about the
blue
group then the red group, whereas
the other half listed their thoughts about the red group then the blue
group.
Subjects in the one-group conditions listed thoughts only about
the group they had read about.
Results
Attribute
ratings.
Subjects' ratings of each group in terms of
intelligence and friendliness served as a measure of perceived
central tendency. Planned comparisons revealed that subjects
in the two-group condition perceived the blue group as more
friendly (M = 7.2) than the red group {M = 3.4), F(l, 14) =
24.31,
p <
.01.
Likewise, the blue group was perceived as less
intelligent (M
=
4.6) than the red group (M
=
6.9), F(l, 14) =
9.49, p < .01. Finally, subjects perceived a greater difference
between the two groups in terms of friendliness (M = 3.8) than
in terms of intelligence (M
=
2.3), F(\,
14) =
32.08,
p
<
.001.
In
addition, subjects in the one-group conditions formed
the
same
impressions of
the
groups as subjects in the two-group condi-
tion. The blue group was perceived as friendly (M = 7.7) and
unintelligent (M= 4.9), whereas the red group
was
perceived
as
unfriendly (M= 3.9) and intelligent (M= 7.1). None of these
means differed from the corresponding means in the two-
group condition.
Associative
strength.
The experimenter and an independent
judge (an undergraduate assistant), each blind to condition,
coded the thought listings for each group for each subject.
Thoughts were coded as being associated with friendliness or
intelligence, or irrelevant to either attribute dimension. The
agreement between the judges was
84%
(Cohen's kappa = .84).
Disagreements
were
decided by the experimenter.
As
in Exper-
iment
1,
attribute-irrelevant thoughts were not included in sub-
sequent analyses.
One purpose of this experiment
was
to replicate the findings
of Experiment
1
in the two-group condition, but when thought
listings for each group were collected separately. A 2 (order of
collection: blue group first or red group first) X 2 (generated
attribute dimension: friendliness or intelligence) X 2 (group
rated) ANOY\ was conducted on the thought listings in the
two-group condition. Replicating the
findings
of Experiment
1,
there was a significant main effect of attribute dimension, F(l,
14) = 33.35, p < .001, indicating that subjects listed more
thoughts pertaining to friendliness (M=
4.93)
than intelligence
(M =
1.33).
There were no other significant main effects or
interactions. Most notably, there was no effect of order.
Thoughts generated to describe the blue and the red groups did
not differ depending on whether the blue or the red group was
considered first.
To test the prediction that subjects would list more thoughts
pertaining to friendliness only in the two-group condition, the
thought
listings
for each group
in
the two-group condition were
compared with the corresponding thought listings from the
one-group conditions. Mean thought listings pertaining to
friendliness and intelligence were analyzed separately for each
group using 2 (number of groups condition: one group or two
group)
X 2
(generated attribute
dimension:
friendliness or
intel-
ligence)
ANGVAs.
Means are presented in Table 1.
For the
blue
group,
the Number of Groups Condition
X
Gen-
erated Attribute Dimension interaction was significant, F(L,
28) =
22.96,
p
<
.001.
Planned comparisons revealed that in the
two-group condition, subjects generated more thoughts per-
taining to friendliness (M= 2.53) than intelligence (M
=
0.60),
F(l,
28) =
49.18,
p
<
.001.
In
the
one-group condition, however,
Table
1
Mean Number of Generated Thoughts Pertaining to Intelligence
and Friendliness for Each Group in Experiment 2
Attribute dimension
Condition
Blue group
Two group
One group
Red group
Two group
One group
n
15
15
15
14
Intelligence
0.60
1.13
0.73
1.21
Friendliness
2.53
1.20
2.40
1.50
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
362THOMAS E. FORD AND CHARLES STANGOR
there was not a significant difference between the number of
thoughts pertaining to friendliness (M
=
1.20) and intelligence
(M=
1.13),
F(l, 28) < 1.00.
Turning to the thought listings for
the
red
group,
the Number
of Groups Condition
X
Generated Attribute Dimension inter-
action was also significant, F(l, 27) = 5.16, p- .03. Planned
comparisons revealed that in the two-group condition, subjects
generated more thoughts pertaining to friendliness (Af
=
2.40)
than intelligence (M = 0.73), F(l, 27) =
15.63,
p
<
.001.
In the one-group condition, however, there was not a
significant difference between the number of thoughts pertain-
ing to friendliness (M
=
1.50) and intelligence (M
=
1.21),
F(l,
27) < 1.00.
Free
recall.
To
test the hypothesis that memory for behaviors
influenced the results, recall was examined. A 2 (number of
groups condition: two group or one group) X 2 (attribute di-
mension
recalled:
friendliness or intelligence)
ANOVA
was con-
ducted on the mean number of behaviors recalled for both the
blue group and the red group. The Number of Groups Condi-
tion
X
Attribute Dimension Recalled interaction failed to reach
significance for either group (both Fs <
1.00).
In the two-group
condition the difference between the number of behaviors re-
called associated with each group pertaining to friendliness
(M= 2.32) and intelligence (M
=
1.56) also fell short of signifi-
cance, F(l, 14) = 3.35, p
=
.08.
Discussion
Consistent with the contention that subjects attempted to
differentiate the two groups, results from the thought listings
indicate that subjects in the two-group condition characterized
the groups more in terms of friendliness than intelligence,
whereas subjects in both one-group conditions did not rely
more heavily on the friendliness
dimension.
This was
predicted
because only in the two-group condition was it necessary to use
the most informative attribute dimension to differentiate the
groups.
In addition, the present results suggest that the greater
use
of
the more informative attribute dimension in Experiment
1
was
not due to an artifact of the procedure. Even when group
stereo-
types were assessed separately, subjects characterized the
groups more in
terms
of the attribute dimension with the larger
mean differences. Thus, demand to compare the groups does
not seem to have produced the effects.
Overall, the results of Experiment
2,
together with results of
Experiment
1,
suggest that individuals
use
the
size
of the differ-
ence between the central tendencies of groups with respect to
different attribute dimensions when forming group stereo-
types.
Experiment
3
addresses
the
question of whether variabil-
ity information is abstracted and used in a similar manner.
Experiment 3
Recent research has demonstrated that variability of group
members with respect to stereotypic attributes is important in
understanding perception of groups. In one relevant study,
Park and Hastie (1987) found that the degree of variability
among group members within a social group on an attribute
dimension influenced the extent to which that dimension was
used to make judgments about the group and its members.
Subjects who perceived a group as highly variable were less
likely to generalize from the attributes of one group member to
the rest of the group, suggesting that group membership was
less
informative about
the
personal characteristics of individual
members when the group was highly variable.
These findings indicate that perceptions of groups are in-
fluenced by variability among group members and that the
more variable
a group is
regarding
a
particular attribute dimen-
sion, the less useful a group stereotype based on that attribute
dimension would be for making judgments of individuals. It
follows that, as stereotypes are being formed, they are more
likely to be based on attribute dimensions for which the vari-
ability within a group is perceived to be low than on ones for
which within-group variability is higher. Experiment 3 tested
this hypothesis.
Method
Overview. Experiment
3
used a procedure similar to that of Experi-
ment
1.
Subjects were asked to form impressions of two groups while
reading behaviors performed
by
members of each
group.
In this exper-
iment, the behaviors were chosen to create differences in the variabil-
ity of the groups rather than differences in central tendency. We pre-
dicted that subjects would be more likely to use the attribute dimen-
sion for which there
was
small within-group variability
to
characterize
the two groups when asked to express their thoughts about what each
group was like.
Subjects. Subjects
were 17
male and
23
female introductory psychol-
ogy students
who
participated in the experiment in exchange for extra
credit in their psychology
courses.
Subjects
were
randomly assigned to
one of two experimental conditions
(lower variability on intelligence
or
lower variability on friendliness).
Stimulus materials. All subjects received booklets describing behav-
iors performed by members of
two
target groups (the blue group and
the red group). Each group
was
described by
18
different behaviors: 6
related to friendliness, 6 related to intelligence, and 6 that were not
related
to
either friendliness or
intelligence.
Behaviors
were
selected on
the
basis
of a pretest rating
(see
Experiment 1)
so
that
in
each condition
the variability of behaviors describing the groups was greater along
one dimension than along the other.
In the lower-variability-on-intelligence condition, the standard de-
viation of mean ratings for behaviors describing the blue group was
0.12 on the intelligence dimension and
2.11
on the friendliness dimen-
sion. The standard deviation of mean ratings for behaviors describing
the red group
was
0.23 on the intelligence dimension and
2.41
on the
friendliness dimension. Bartlett-Box F tests of homogeneity of vari-
ance indicated that there
was a
significant difference between the vari-
ability of behaviors along the two dimensions for the blue group, F(l,
10) = 21.42, p
<
.001,
and for the red group, F(\,
10) =
15.72, p
<
.001.
In the lower-variability-on-friendliness condition, the standard de-
viation of mean ratings for behaviors describing the blue group was
2.08 on the intelligence dimension and 0.52 on the friendliness dimen-
sion. The standard deviation of mean ratings for behaviors describing
the red group was 1.99 on the intelligence dimension and 0.73 on the
friendliness dimension. Bartlett-Box F tests of homogeneity of
vari-
ance indicated that there
was a
significant difference between the vari-
ability of behaviors along the two dimensions for the blue group, F(l.
10)
= 6.99, p <
.01,
and for the red group, F(l, 10) = 3.99, p
<
.05.
Although the variabilities of the attribute dimensions differed, the
difference between group means did not. Separate
2
(attribute valence:
positive or negative) X 2 (attribute dimension) ANOVAs were con-
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
DETECTING DIAGNOSTIC ATTRIBUTES363
ducted for each experimental condition. In the lower-variability-on-in-
telligence condition, the interaction effect was not significant, F(l,
20) < 1.00, indicating that the group means were approximately
equally different on the intelligence dimension (Aft = 6.2 for the posi-
tively described red group and 3.9 for the negatively described blue
group) as on the friendliness dimension (Aft = 5.9 for the positively
described blue group and 3.7 for the negatively described red group).
Similarly, in the lower-variabilhy-on-friendliness condition, the two-
way interaction was not significant, F(l, 20) < 1.00. The difference
between the group means on intelligence (Aft = 6.0 for the positively
described red group and 3.9 for the negatively described blue group)
was the same as the mean difference on friendliness (Aft = 6.1 for the
positively described blue group and 3.5 for the negatively described red
group).
Procedure. The procedure was exactly the same as in Experiment 1,
except that subjects completed an additional questionnaire. In addi-
tion to making their estimates of the central tendency of each group,
subjects also completed a measure designed to assess the perceived
variability of each group in terms of intelligence and friendliness (e.g.,
Linville, Fischer, & Salovey, 1989). Subjects were presented with seven
levels of each attribute dimension (e.g, extremely
unfriendly, unfriendly,
somewhat unfriendly,
neutral,
somewhat friendly, friendly, and ex-
tremely friendly) and were instructed to estimate the percentage of
members of each group who would fall into each of the seven levels.
Results
Perceived
variability.
We used the maximum percentage as-
signed to any one level on the variability scale as an index of
perceived variability. It was assumed that the maximum per-
centage
of group members assigned to
a single level
on the scale
would be larger if the group
was
perceived to
have
low variabil-
ity than if it was perceived to have higher
variability.
Therefore,
it was expected that the maximum percentage of group
members assigned to one
level
on the variability
scale
would be
greater for the low-variability attribute dimension than for the
high-variability attribute dimension.
In the lower-variability-on-intelligence condition, the maxi-
mum was larger for the intelligence attribute dimension (M
=
47.4) than for the friendliness dimension (M= 43.5). Likewise,
in the lower-variability-on-friendliness condition, the maxi-
mum was larger for the friendliness attribute dimension (M =
52.3) than for
the
intelligence attribute dimension (M= 49.75).
A 2 (variability condition) X 2 (attribute dimension) ANOVA
was conducted
on these
means.
Although in the predicted
direc-
tion, the Variability Condition
X
Attribute Dimension interac-
tion fell short of significance, F(l, 38) = 2.14, p = .15. It is
possible that our measure was not sensitive enough to detect
significant differences in perceived variability with respect to
the high- and low-variability attribute dimensions. Neverthe-
less,
data from the pretest showed that the stimulus materials
did have the desired characteristics.
Attribute ratings. Because the
analyses
of the stimulus mate-
rials showed that the central tendency of the groups with re-
spect to each attribute dimension did not vary between condi-
tions,
it was expected that the perceived difference between
mean ratings of each group would be the same for both the
high- and the low-variability attribute dimensions. This as-
sumption was supported for the intelligence dimension.
Separate
2
(group)
X 2
(variability
condition:
intelligence low
variability or friendliness low variability) ANOVAs were con-
ducted on the intelligence and friendliness ratings. On the di-
mension of intelligence, the groups were seen as equally ex-
treme in both variability conditions. There was no main effect
or interaction involving the variability condition. Planned com-
parisons revealed that,
as
expected,
in the lower-variability-on-
inteliigence condition the red group was rated as more intelli-
gent (M= 7.3) than the blue group (A/= 4.4), F(l, 38) = 52.74,
p <
.001.
Likewise, in the low-variability-in-friendliness condi-
tion, the red group
was
rated
as
more intelligent (Af = 7.6) than
the blue group
(A/
=
4.3), F\\, 38) = 62.17, p<
.001.
On the friendliness dimension, however, the low-variability
groups were perceived
as
more extreme than the high-variabil-
ity groups. The Variability Condition x Group Rated interac-
tion was significant, F(l, 38) = 8.70, p <
.005.
A planned com-
parison revealed that in the lower-variability-on-friendliness
condition, the red group was rated as less friendly (M = 4.4)
than the blue group (A/- 7.4), F(l,
38) =
16.40,
p<
.001.
In the
lower-variability-on-intelligence condition, however, there was
no perceived difference between the groups (M= 5.8 for both
groups). It
is
possible that, on the friendliness dimension, high
variability was perceived as implying small mean differences.
Associative
strength.
The major prediction of this experi-
ment was that subjects would form stereotypes about groups
along attribute dimensions for which within-group variability
was low and that when asked to list thoughts about the two
groups would list more thoughts pertaining to the attribute
dimension that denoted lower within-group variability.
The
experimenter and an independent judge
(an
undergradu-
ate assistant), each blind to condition, coded the thought list-
ings for each group by each subject. Thoughts were coded as
being related to friendliness or intelligence or irrelevant to ei-
ther attribute dimension. The agreement between the judges
was 89% (Cohen's kappa = .79). Disagreement was decided
by the experimenter. As in Experiment
1
and Experiment 2,
attribute-irrelevant thoughts were not included in subsequent
analyses.
The mean number of thoughts pertaining
to
intelligence and
friendliness generated in each variability condition is pre-
sented in Figure
2.
In the lower-variability-on-intelligence con-
Mcan Number of Qtn«rat«d Thought*
3.6
3
2.6
2
1.6
1
0.6
InMligane* Friandlinau
Lowar Variability
B Intelligano* CD FriandlliM**
Figure 2. Mean number of generated thoughts pertaining to intelli-
gence and friendliness in each condition of Experiment 3.
I1
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
364THOMAS E. FORD AND CHARLES STANGOR
dition, there were a greater number of thoughts pertaining to
intelligence generated about each group (M = 2.90) than
thoughts pertaining to friendliness (M= 2.09). Likewise, in the
lower-variability-on-friendliness condition, there were a
greater number of thoughts pertaining to friendliness gener-
ated about each group
(M =
3.32) than thoughts pertaining to
intelligence
(M =
2.42).
A 2 (variability condition) X 2 (generated attribute dimen-
sion) ANOVA was conducted on the means of the thought list-
ings.
The Variability Condition
X
Attribute Dimension interac-
tion
was
significant, F(l,
38) =
5.27, p
<
.05.
This indicates that
subjects generated more thoughts about the groups in terms of
the attribute dimension that denoted lower within-group vari-
ability. Planned comparisons, however, revealed that in the
lower-variability-on-intelligence condition, the number of
thoughts generated for each group pertaining to intelligence
was not significantly greater than the number of thoughts per-
taining to friendliness, F(l,
38) =
2.51,
p
<
.25.
Similarly, in the
lower-variability-on-friendliness condition, the difference be-
tween the number of thoughts generated for each group in
terms of friendliness and intelligence fell short of significance,
F(l,39)=2.82, p<.09.
Free
recall.
It is assumed that memorability is related to
behavioral extremity (Rothbart et al., 1978), and because we
held differences between group means along each attribute di-
mension constant in this experiment, we predicted no recall
differences as a function of variability condition. In the lower-
variability-on-intelligence condition, the mean number of
be-
haviors recalled pertaining to intelligence was 1.21 and the
mean number of behaviors recalled pertaining to friendliness
was
1.48. In the lower-variability-on-friendliness condition, the
mean number of behaviors recalled pertaining to intelligence
was
1.29 and the mean number of behaviors recalled pertaining
to friendliness was
2.11.
A 2 (variability condition)
X
2 (attrib-
ute dimension recalled)
ANOVA
revealed that,
as
expected,
the
Variability Condition
X
Attribute Dimension Recalled interac-
tion was not significant, F(l, 38) = 2.81, p = .10. The only
significant effect was a main effect of attribute dimension, F(l,
38) =
10.63,
p < .002, indicating that subjects recalled more
behaviors related to friendliness (M = 1.79) than intelligence
(M=1.25).
Discussion
As predicted, when asked to describe two groups, subjects
characterized them more in terms of the attribute dimension
that denoted lower within-group variability. These results sup-
port the hypothesis that group stereotypes are more likely to be
based on attribute dimensions for which within-group variabil-
ity is low than on ones for which within-group variability is
higher. The proposed explanation for these results is that sub-
jects used the size of the within-group variability with respect
to the two attribute dimensions
as
an indication of their utility
for differentiating the groups. To the extent that a social group
is perceived to have low variability on a particular attribute
dimension, a stereotype based on that dimension would pro-
vide diagnostic information about the group and would there-
fore facilitate inferences and judgments of individuals based on
group membership.
As
with Experiments
1
and
2,
the
results
of Experiment
3
also
cannot be explained in terms of memorability differences. An
interesting feature of this design is that the higher variability
attribute dimension contains behaviors that are actually some-
what
more
extreme than the lower variability dimension. Thus,
if stereotypes are based only on attribute dimensions denoting
behavioral extremity, then subjects should have characterized
the groups more in terms of the larger variability attribute di-
mension. However, subjects generated more thoughts related to
the low-variability attribute dimension, which, because it had a
smaller range, actually consisted of behaviors that were less
extreme than the behaviors on the large-variability attribute
dimension.
Although the results of Experiment 3 suggest that subjects
generated more thoughts about the two groups in terms of the
less variable attribute dimension, the effect of the within-group
variability manipulation on the associative strength measure
was weaker than that of the between-groups mean difference
manipulation of Experiments
1
and
2.
One
possible
explanation
for
this
relative weakness
is
that it
is
more difficult to keep track
of within-group variability regarding different attribute di-
mensions. The use of within-group variability would presum-
ably involve estimating central tendencies of the groups regard-
ing different attribute dimensions as well as estimating the ex-
tent to which behavioral instances deviated from those central
tendencies. Thus, abstracting variability information may re-
quire more extensive cognitive processing than abstracting cen-
tral tendency information.
A
second possibility
is
that the vari-
ability manipulation in Experiment
3
was simply not as strong
as the central tendency manipulation of Experiments
1
and 2.
General Discussion
The results of the present research make several contribu-
tions to an understanding of the nature of social
stereotypes,
as
well as having implications for how to measure them. In the
following sections we consider the implications of
the
present
studies for issues of stereotype formation and stereotype mea-
surement.
Importance of Intergroup Differentiation
The present results suggest that, when forming impressions
of social groups, individuals form group stereotypes by using
the central tendency and variability of different attribute di-
mensions as criteria for determining which dimensions are
most diagnostic
(i£.,
provide the most valuable information for
differentiating the groups). The idea that highly diagnostic at-
tributes are most informative about social groups is consistent
with Tajfel's emphasis on the importance of differentiation:
"The characteristics of one's group
a