Personality and Individual Differences 30 (2001) 977±994

www.elsevier.com/locate/paid

Perspectives on ¯uid and crystallized intelligence: facets for

verbal, numerical, and ®gural intelligence
Andre Beauducel a,*, Burkhard Brocke a, Detlev Liepmann b
a
Technische UniversitaÈt Dresden, Institut fuÈr Klinische, Diagnostische und Di€erentielle Psychologie,
Mommsenstr. 13, 01069, Dresden, Germany
b
Freie UniversitaÈt Berlin, Berlin, Germany

Received 19 November 1999; received in revised form 10 March 2000

Abstract
Fluid intelligence is often measured with ®gural tests, whereas crystallized intelligence is often assessed
with verbal tests. It is argued that construct-irrelevant ®gural variance is included in ¯uid intelligence and
construct-irrelevant verbal variance is included in crystallized intelligence. The speci®cation of a content
facet comprising verbal, numerical, and ®gural abilities for ¯uid and crystallized intelligence would reduce
the construct irrelevant variance. This faceted view of ¯uid and crystallized abilities is regarded as more
convincing than a purely hierarchical structure. Although the present approach is partly similar to Gutt-
man's Radex model, no radial partitioning of the tasks is expected. Seven hundred and six German parti-
cipants aged between 14 and 50 years were tested with the I-S-T 2000, a test comprising verbal, numerical,
and ®gural reasoning tasks, as well as verbal, numerical, and ®gural knowledge tests. In smallest space
analysis, a simplex for ¯uid and crystallized intelligence emerged as well as a radial or a polar facet for
verbal, numerical, and ®gural content. The faceted structure for ¯uid and crystallized intelligence was also
shown in con®rmatory factor analysis and ®tted the data more completely than the hierarchical model. The
implications for the conceptualization and the assessment of ¯uid and crystallized intelligence are dis-
cussed. # 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Fluid intelligence; Crystallized intelligence; Facet theory; I-S-T 2000

* Corresponding author. Tel.: +351-463-6996; fax: +351-463-6993.

E-mail address: [email protected] dresden.de (A. Beauducel).

0191-8869/01/$ - see front matter # 2001 Elsevier Science Ltd. All rights reserved.
PII: S0191-8869(00)00087-8
978 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

1. Introduction

The present study examines possibilities for further development of the `classical' theory of
¯uid (gf) and crystallized (gc) general intelligence (Cattell, 1963, 1987; Horn, 1988). The intended
development is based on the introduction of a facet for verbal, numerical, and ®gural intelligence.
Important parts of this approach were already introduced in the manual for a German intelli-
gence test, the `Intelligenz-Struktur-Test 2000' (Amthauer, Brocke, Liepmann & Beauducel,
1999), which is a completely revised form of Amthauer's (1971) intelligence test. The revision of
the test focussed on the theoretical conception and was largely based on the critique in Brocke,
Beauducel and Tasche (1998). Since research concerning the gf±gc theory has mainly been con-
ducted in English-speaking countries, the further theoretical elaboration and the discussion of the
present approach should be presented in English, too.

1.1. The measurement of gf and gc

The gf±gc theory is widely used and discussed in current research (e.g. Carroll, 1993; Flanagan,
Genshaft & Harrison, 1997; HaÈrnqvist, Gustafsson, MutheÂn & Nelson, 1994; Lindenberger &
Baltes, 1997; Staudinger, Maciel, Smith & Baltes, 1998). It is not easy to de®ne simple scales
containing few variables for measuring gf and gc, because the gf and gc factors were de®ned by
large sets of di€erent variables, and even included personality-trait markers in the ®rst studies
(e.g. Cattell, 1963; Horn & Cattell, 1966). According to Horn (1988), gc has prominent relation-
ships with verbal knowledge, following instructions, information about the humanities, social and
physical sciences, culture in general, as well as problem de®nition. Horn and Noll (1997) called gc
`acculturation knowledge', expressing the importance of the knowledge domain for the con-
ceptualization of gc. Fluid intelligence appears to operate whenever the sheer perception of
complex relations is involved (Cattell, 1987). It spreads over many kinds of relationships: classi-
®catory similarities, causal relations, inductive reasoning, abstract relations in numbers, and
inferential relations. According to Horn (1988), good measures of gf are inductive reasoning,
concept formation, visual conceptualization, e€ectiveness in using problem-solving strategies, and
numbers reversed memory.

1.2. Problems and conceptual perspectives in current gf±gc measurement

The high generality and variable operationalization of gf and gc has some implications for the
theoretical conception of the constructs. Both broader and more speci®c interpretations of gf and
gc are sometimes proposed. The more speci®c view of gf focuses on di€erent forms of reasoning
as primaries for gf (e.g. Carroll, 1993; Flanagan & McGrew, 1997). The more speci®c view of gc
focuses on di€erent forms of knowledge and language aptitudes (Carroll, 1993; Flanagan &
McGrew, 1997). The focus on knowledge as a basis of gc is in line with Ackerman's (1996) view
of `intellect-as-knowledge'. However, according to the broader perspective, gf comprises reason-
ing as well as memory and perceptual speed, whereas gc comprises knowledge and ¯uency (e.g.
Lindenberger & Baltes, 1997). These broader conceptions of gf and gc are surely interesting and
well founded (see Hakstian & Cattell, 1978; Horn & Cattell, 1982; Gilardi, Holling & Schmidt,
1983) and have some potential especially for developmental psychology (e.g. Baltes, Lindenberger
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 979

& Staudinger, 1998). But, as mentioned above, the broader conceptions of gf and gc are not
consistently found in the literature. For example, in Carroll (1993) ¯uency forms an additional
factor at the same level of generality as gc. This means that ¯uency cannot be simply integrated in
gc. Also in Carroll (1993), memory forms a second order-factor in addition to gf and cannot be
simply integrated in gf. However, both the more general and the more speci®c view of gf and gc
show some empirical and theoretical evidence. Consequently, it might be premature to exclude
one of the interpretations. But even the more speci®c interpretation of gf and gc is complex,
leading to complex strategies for the assessment of gf and gc (see Flanagan & McGrew, 1997).
Therefore, the more speci®c interpretation of gf Ð including only aspects of reasoning (see also
Carroll, 1993) and the more speci®c interpretation of gc, not including ¯uency (see also Horn,
1988) Ð is adopted here.
The problem of construct underrepresentation is considered in Flanagan and McGrew (1997)
for the case of interpreting the information subtest in the Wechsler Intelligence Scale for
Children Ð third edition (WISC-III, Wechsler, 1991) as a measure of gc. Flanagan and McGrew
(1997, p. 315) were aware of the problem of underrepresentation and conclude: ``Currently, no
intelligence test battery adequately operationalizes the gf±gc theory of intelligence''. They
recommend an improvement of the assessment of gf and gc by means of their `cross-battery
approach'. The cross-battery approach integrates a number of di€erent tests. However, the aim of
the cross-battery approach is mainly pragmatic in that the assessment of gf and gc is improved
using a combination of the instruments available.
The idea of a cross-battery sample of tests for the assessment of gf and gc takes the current
conceptualizations of gf and gc as they are reported in Cattell (1987), Horn (1988) or Carroll
(1993) as a starting point. This is reasonable and correct, if past results and concepts of gf and gc
are to be represented with tests. There can, however, be another way of dealing with the complex
assessment of gf and gc: perhaps the diculties in assessment of gf and gc are due to some pro-
blems in the theoretical conceptualization of gf and gc.

1.3. A new theoretical perspective for gf and gc

The gf±gc theory essentially is a hierarchical model of intelligence. However, as Sternberg and
Powell (1982) pointed out, besides the hierarchical models, models for overlapping abilities (for
instance, faceted models) have been proposed (e.g. Guilford, 1967). According to Sternberg and
Powell (1982), at the highest level of theories of intelligence structure the conceptions of hier-
archical and overlapping abilities are integrated. They mention Guttman's (1965) radex model as
an example for the highest level integrating hierarchy in the complexity facet and overlap in the
content facet. In the German-speaking area, there is another prominent hierarchical and faceted
model: the Berlin Model of Intelligence Structure (BIS; JaÈger, 1982, in English see Wittmann,
1988; Bucik & Neubauer, 1996).
With respect to the gf±gc theory, a lack of integration of overlapping abilities in the sense of
Sternberg and Powell (1982) into the hierarchical perspective might have produced some weak-
nesses. These weaknesses can be shown, when one takes a closer look at the typical shifts in the
conceptualization of gf and gc: a typical shift is the conceptualization of gf as non-verbal or ®g-
ural and gc as verbal (e.g. Pry, 1995; Schoenthaler, Amos, Eysenck, Peritz & Yudkin 1991;
Thorndike, Hagen & Sattler, 1986). Such interpretations are supported by the frequent loadings
980 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

of verbal tasks on a gc factor and of the ®gural/topological tasks on a gf factor (e.g. Cattell,
1987). In this context, the performance in verbal tasks is mostly seen as being strongly dependent
on learning and acculturation whereas the ®gural/topological abilities are assumed to be quite
independent from acculturation. Of course, these interpretations may be correct in many cases,
but they suggest the equation of gc with verbal and gf with ®gural abilities. This view leads away
from the original conceptualization of gf and gc not as (®gural or verbal) content factors, but as
dependent on the degree of acculturation (e.g. Horn, 1988).
It is conceptual reductionism to assume that verbal tasks simply measure gc. Cattell (1987)
acknowledges the possibility that verbal tasks can load on ¯uid intelligence when they include
vocabulary which is available for all participants (see also Horn, 1988). For comparable reasons,
Sternberg and Gastel (1989) interpreted their verbal reasoning tasks as measures for gf. More-
over, Sternberg and Gardner (1983) showed that information processing on verbal and ®gural
reasoning tasks can be very similar. According to Flanagan and McGrew (1997), the interpreta-
tion of the Wechsler Verbal IQ as gc is an example for the inclusion of construct-irrelevant var-
iance. The same holds for numerical calculations: Horn (1988) states that numerical calculations
will measure gf when they mainly require reasoning rather than knowledge of mathematics as
such. Interestingly, even for the ®gural abilities, a corresponding argument is given in Horn
(1988): if a visual conceptualization task mainly requires reasoning, it will mainly indicate gf. But
if the task mainly requires the eye to ¯ow over a pattern and see how ®gures can change
appearance as they are moved through space, the task will mainly indicate broad visual intelli-
gence. The example shows that it is not the ®gural or visual context itself which ensures that a
task loads mainly on gf. According to Horn (1988) it is empirically evident that the degree of
`pure' reasoning demands of a task is essential for the degree of gf measured.
Obviously, the degree of gf can be manipulated across three types of content (i.e. verbal,
numerical, and ®gural) and the degree of gc can be manipulated in two di€erent types of content
(i.e. in verbal and numerical tasks). But if this is the case, gf and gc should not be identi®ed with
any speci®c type of content, because they are abilities which are relevant in di€erent types of
content. Of course, due to speci®c distributions of knowledge in di€erent groups of participants,
sometimes a high correlation between a content factor and gf or gc may occur. However, the
close relations of gf or gc to speci®c content factors can be explained within the context of
acculturation, and may therefore not be interpreted as characteristic for gf and gc themselves. It
may, consequently, be theoretically useful to protect the conceptualization and interpretation of
gf and gc against the contamination with content factors, because the relation of gf and gc to
content factors may change from one sample to another.
The problems with the conception and measurement of gf and gc can be seen as the result of a
lack of the inclusion of overlapping abilities in the sense of Sternberg and Powell (1982) into the
model. The above-mentioned problems could be reduced when overlapping abilities are inte-
grated in the gf±gc theory by means of a content facet for verbal, numerical and ®gural abilities.
If the gf and gc factors or scales are aggregates across three di€erent types of content, the in¯u-
ence of speci®c types of content should be reduced (for the principle of suppression of unwanted
variance see Humphreys, 1962; JaÈger, 1982). If a content facet is introduced for gf, the above-
mentioned group-speci®c and task-speci®c relations of gf and gc to verbal, numerical, and ®gural
or visualization abilities should be balanced out. Furthermore, when overlapping abilities are
integrated in the hierarchical gf±gc model by means of a content facet all Ð until now Ð available
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 981

possibilities of structuring intelligence variance, would be included in the model. Thereby the gf±
gc model would become a model of the highest level of theories of intelligence structure according
to Sternberg and Powell (1982).

1.4. Hierarchical and faceted structures of gf and gc

However, the question arises whether the combination of gf and gc with content factors for
verbal, numerical, and ®gural abilities could not be described in terms of a purely hierarchical
model as well. Therefore, a theoretical alternative can be speci®ed with six factors at the bottom
of the hierarchy, three factors for verbal, numerical, and ®gural reasoning and three factors for
verbal, numerical, and ®gural knowledge. On the next level of hierarchy the three content factors
for reasoning would combine to a second order reasoning factor, which would correspond to gf
and the three content factors for knowledge would combine to a second order knowledge factor
which would correspond to gc. These second order factors are expected to correlate which can be
explained by the e€ect of a factor for general intelligence.
This model would be a reasonable alternative to the above-mentioned model and has simila-
rities to Gustafsson's (1984) hierarchical model. However, it is hypothesized in the present con-
text that this type of a purely hierarchical model would not ®t the data as well as the faceted one.
The reason is, that the hierarchical model would not allow the variances of each task to split up
directly between gf and gc on the one hand and verbal, numerical, and ®gural abilities on the
other. Moreover, the relations between gf and gc would only be accounted for by the correlation
between both factors, i.e. the variances of the reasoning tasks cannot split up directly between gf
and gc. However, the partitioning of the variances at the level of the reasoning tasks and not at
the level of factors is regarded as an important feature of the above-mentioned model. According
to Horn (1988) it is assumed that typical reasoning tasks involve both aspects of pure reasoning
and knowledge. Moreover, it is assumed that pure reasoning and knowledge are involved in the
reasoning tasks to very di€erent degrees. In order to represent these di€erent degrees of pure
reasoning and knowledge involved in the tasks, it seems important to split up the variances of the
reasoning tasks directly. Thus, there are several reasons to prefer the faceted model in the present
context. However, the purely hierarchical model will be tested as an alternative.

1.5. Relations of the faceted concept of gf and gc to the radex-model

The introduction of a content facet for verbal, numerical, and ®gural abilities in gf and gc can
be regarded as a theoretical integration of aspects of the radex-model (Guttman, 1965; Guttman
& Levy, 1991) in gf±gc theory. Guttman and Levy (1991) propose a facet `rule task' for rule
inference, rule application, and learning. The relations of gf±gc to rule inference and rule appli-
cation are further clari®ed, when the tasks, which are supposed to measure rule inference and rule
application, are considered: rule inference is measured with inductive reasoning tasks, which are
typically used to measure gf (Cattell, 1987; Horn, 1988). Rule application is measured with
deductive reasoning tasks, as well as vocabulary and knowledge tests (Adler & Guttman, 1982;
Guttman & Levy, 1991; Shye, 1988), which typically measure gc (Horn, 1988). The parallels
between gf and gc on the one hand, and rule inference and rule application on the other, is also shown
in Marshalek, Lohman and Snow (1983): gf can be located in the center of the radex, where rule
982 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

inference is usually located in the radex-model, and gc is located in the radial segment of rule
application. In the radex model, rule inference and rule application are di€erentiated by means of
a polar content facet for verbal, numerical, and ®gural abilities (e.g. Guttman & Levy, 1991). The
parallels between gf±gc and rule inference±rule application suggests that a content facet for verbal,
numerical, and ®gural abilities can be introduced in the gf±gc theory. The di€erentiation of gf and
gc on the basis of a content facet can be regarded as an integration of the radex model and gf±gc
theory. The advantages of the two models should be saved and integrated in this new perspective.
For gf±gc theory it is the dynamic interaction of both factors, i.e. the investment theory with its
developmental implications, for the radex model it is the faceted conceptualization of intelligence.
However, the question arises whether the Cattell±Horn concept is just replaced by the Guttman
concept in the present modi®cation of gf and gc. In this case, rule inference would just be inter-
preted as gf, and rule application as gc. If this is the best view, the radex structure would be the
optimal representation of the relations between gf and gc, and the concepts of rule inference and
rule application would be replaced by the concepts of gf and gc. However, empirically both, gf
and gc, regularly emerge as second-order or second-stratum factors (e.g. Cattell, 1963; Hakstian
& Cattell, 1978; Horn & Cattell, 1966), i.e. they are both homogeneous enough to constitute dif-
ferent factors. If the radex model (e.g. Guttman & Levy, 1991) would have been the best repre-
sentation in smalest space analysis (SSA), the gf tasks would have been located in the center of
the radex with the gc tasks around them. In this case, it would have been dicult or even
impossible to yield a gc factor in factor analysis, because the tasks, which constitute gc, would
have been too heterogeneous. However, Marshalek et al. (1983) were able to demonstrate a gf, gc,
and a general visualization factor in factor analysis and to locate the marker variables of these
factors in SSA. In their SSA solution, the gc markers were not arranged in a circumplex around
the gf markers. Only when the gc markers were taken together with parts of the markers for
general visualization, some radial partitioning was possible. Moreover, tasks which involve
knowledge and the application of rules are not necessarily less complex. Therefore, in the present
context, the di€erence between gf and gc is still interpreted according to the Cattell±Horn model,
and no radial partitioning for the gc tasks around the gf tasks is expected, since both factors were
suciently homogeneous to emerge regularly in factor analyses.

1.6. Purpose of the present study

To sum up, the present investigation has three aims: (1) to avoid construct-irrelevant variance,
i.e. the contamination of verbal and ®gural intelligence with gc and gf; (2) to demonstrate that in
the present context a faceted structure for gf, gc, on the one hand and verbal, numerical, and
®gural abilities on the other will be superior to a hierarchical model for these abilities; and (3)
that the faceted structure ®tting the data will not be a radex structure, even when gf and gc are
measured with verbal, ®gural, and numerical tests. These points suggest the following hypotheses.

1. For gf and gc, a content facet for verbal (V), numerical (N), and ®gural (F) intelligence can
be demonstrated in multivariate analysis, when matrices and further reasoning tasks with
verbal, numerical, and ®gural content are included as markers for gf, and a test for verbal,
numerical, and ®gural knowledge is included as a marker for gc (see Horn, 1988; Amthauer
et al., 1999).
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 983

2. A hierarchical model with gf and gc at the top and factors for verbal, numerical, and ®gural
reasoning and verbal, numerical, and ®gural knowledge at the bottom will not ®t the data as
well as a faceted model with gf and gc arranged in one facet and verbal, numerical, and ®g-
ural abilities in the other.
3. A radial partitioning for gf and gc cannot be demonstrated, since the gc tasks are as homo-
geneous as the gf tasks. The gc tasks are as homogeneous as the gf tasks, even when tests
based on di€erent knowledge domains are included (also the second hypothesis implies that
a radex structure should not emerge).

There has been some debate concerning the question as to whether a large amount of age var-
iance should be included in factor analyses of intelligence tasks (Guilford, 1980). Horn and Cat-
tell (1966, 1982) defended the analysis of age-heterogeneous groups with the developmental
perspective of the investment theory (see also Cattell, 1987, Horn, 1988). Since the developmental
assumptions are important issues for the gf±gc theory (e.g. Cattell, 1987; Kaufman, Kaufman-
Packer, McLean & Reynolds, 1991), the above-mentioned hypotheses are tested in a group which
is heterogeneous with respect to age. Additional analyses are conducted with more age-homo-
geneous subgroups, in order to reduce age-speci®c variance as recommended in Guilford (1980)
and to test whether the expected results are stable across age. Of course, a more comprehensive
developmental framework requires far more sophisticated analyses (e.g. Cattell, 1987; Horn,
1988; Lindenberger & Baltes, 1997), but the present study focuses on the structural relations
between gf and gc within a facet approach.

2. Method

2.1. Participants

In a project of test-construction, 468 participants aged between 15 and 60 years were tested
with the I-S-T 2000 (Amthauer et al., 1999). Meanwhile, 238 additional participants were tested.
Altogether, data from 706 participants were included in the analyses (see Table 1).
The mean age of the total group is 24.2 years. The standard deviation is 8.9 years, thus, the
group is heterogeneous with respect to age, but the standard deviation is still smaller than in
Horn and Cattell (1966). However, since the analysis of age-heterogeneous groups has been

Table 1
Age and gender distributiona

Age in years Female Male Total

14±18 121 131 252 (36)

19±24 121 74 195 (28)
25±34 82 70 152 (22)
35±50 68 39 107 (15)
392 (54) 314 (46) 706 (100)
a
Percent values are given in brackets.
984 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

criticized (e.g. Guilford, 1980), more age-homogeneous subgroups were formed for further ana-
lyses. One group of participants were aged between 14 and 24 years (N=447; M=18.6; SD=3.1)
and a second group of participants were aged between 24 and 50 years (N=259; M=33.8;
SD=7.4). The second group is more heterogeneous than the ®rst. But due to the lower number of
older participants in the total sample, it was not possible to form a more age-homogeneous group
of older participants, which is suciently large for multivariate analyses. Concerning the educa-
tional level, only high school graduates (Gymnasium pupils) can be di€erentiated from partici-
pants without high school grade. Participants, who were about to reach their high school grade,
but were not completely ®nished with high school, were classi®ed as graduates. Participants who
were younger than 16 years could not be di€erentiated as graduates or not. These participants
were classi®ed as schoolchildren (see Table 2).

2.2. Material and procedure

First, the nine reasoning tasks of the I-S-T 2000 were applied (three verbal, three numerical,
three ®gural; time: 92 min). After that, 611 of the 706 participants worked on a verbal and a ®g-
ural memory task (time: 12 min). Because the present study focuses on the relation between gf
and gc, based on the distinction between reasoning and knowledge, the memory tasks were not
included in the analyses. Finally, a test for verbal (26 questions), numerical (25 questions), and
®gural knowledge (22 questions) was applied (time: 37 min). The knowledge test comprises
knowledge on art-literature, geography-history, mathematics, science, and daily life. Table 3 gives
some examples for verbal, numerical, and ®gural reasoning tasks and verbal, numerical, and ®g-
ural knowledge questions in the I-S-T 2000. The original items are mostly in multiple choice
form, but they have been shortened in order to save space.
There are two ways of forming knowledge subscales from the material. First, six subscales
based on the content of knowledge can be formed: two for verbal, two for numerical, and two for
®gural knowledge (see Amthauer et al., 1999). Second, eight smaller aggregates, each for verbal
and numerical knowledge on geography-history, verbal and numerical knowledge on science,
numerical, and ®gural knowledge on mathematics, verbal knowledge on art and knowledge of
®gures from daily life can be formed. Thus, the second way of aggregation di€erentiates the
domain of knowledge as well as the content (i.e. verbal, numerical, and ®gural). The complete
testing session took about 2 h 30 min.

Table 2
Educational level and agea

Age in years Schoolchildren No grade With grade Total

14±18 125 22 105 252 (36)

19±24 27 168 195 (28)
25±34 51 101 152 (22)
35±50 62 45 107 (15)
125 (18) 162 (23) 419 (59) 706 (100)
a
Percent values are given in brackets.
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 985

2.3. Data analysis

The analyses were performed with spss 8 and lisrel 8 (JoÈreskog & SoÈrbom, 1993). A faceted
structure cannot be tested directly by means of exploratory factor analysis (but see JaÈger, 1982 for
a method based on controlled aggregation), for this purpose SSA (Borg & Lingoes, 1987) is more
appropriate. Faceted structures can also be demonstrated by means of con®rmatory factor ana-
lysis (CFA), as was shown in Amthauer et al. (1999) as well as Bucik and Neubauer (1996).
Therefore, CFA and SSA will be used in the present context. Both ways of aggregation of
knowledge can be used in separate analyses.

3. Results

Reliabilities and further psychometric properties of the tasks are presented in Amthauer et al.
(1999). Cronbach's Alpha for the nine reasoning tasks ranged between 0.60 and 0.90 (Amthauer
et al., 1999). Only Cronbach's Alpha for the knowledge subscales is presented here in order to
compare the content- and domain-speci®c aggregates with the content-speci®c aggregates (see
Table 4).
Of course, Cronbach's Alpha was considerably lower for the shorter domain- and content-
speci®c aggregates. These lower reliabilities may act in favor of lower intercorrelations of the
scales, and thereby may produce more heterogeneous patterns in SSA. Since homogeneity is
assumed for the gc markers, the lower reliabilities act against our second hypothesis. If the

Table 3
Examples for the tasks of the I-S-T 2000
986 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

patterns of the gc tasks in SSA were homogeneous in spite of partly low reliabilities, the evidence
in favor of the second hypothesis would be quite compelling.
First, a metric SSA was performed with all 706 participants. According to the hypotheses, a
content facet for gf and gc should emerge, and the gc tasks should be as homogeneous as the gf
tasks. A solution with three dimensions with the nine reasoning tasks and the six content sub-
scales of the knowledge test had an acceptable stress of 0.099 (according to Kruskal's stress-for-
mula 1; R2=0.94). The intended structure emerged in dimensions one and two (see Fig. 1). The

Table 4
Cronbach's alpha for content and combined domain and content scales

Content scales Cronbach's  Domain and content scales Cronbach's 

Verbal 1 0.62 Verbal geography/history 0.56

Verbal 2 0.64 Numerical geography/history 0.46
Number 1 0.72 Verbal science 0.54
Number 2 0.67 Numerical science 0.61
Figural 1 0.61 Numerical mathematics 0.80
Figural 2 0.64 Figural mathematics 0.69
Verbal arts 0.62
Figural arts 0.43
Figural daily life 0.66

Fig. 1. SSA for nine reasoning tasks and six knowledge subscales: dimension one and two, V=verbal abilities;
N=numerical abilities; F=®gural abilities.
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 987

content facet emerged in a radial partition, the tasks formed a simplex along the ®rst dimension.
The di€erence between knowledge tests as markers for gc and reasoning tests as markers for gf
was very clear. According to the third hypothesis, the knowledge subscales were at least as
homogeneous as the reasoning tasks.
The present structure di€ers from the radex model because the knowledge tasks involving rule
application according to Guttman and Levy (1991) were not located in a circumplex around the
tasks involving primarily rule inference. An analogous con®guration was found in dimensions
one and three of the present solution. In order to check the stability of the results, the SSA was
calculated separately for the younger and the older group of participants. For the younger group
(N=447), the stress was 0.12 (R2=0.91) and for the older group, the stress was 0.10 (R2=0.93).
In both groups, the partition was the same as in the total group, i.e. there was a circumplex for
content and a simplex for gf±gc. An additional weighted SSA (see Carroll & Chang, 1970) for
individual di€erences between the two age groups with three dimensions had a moderate stress
(stress=0.130; R2=0.89). The ®rst dimension had a larger weight in the younger age-group (0.84)
than in the older group (0.67), indicating that the di€erentiation between gf and gc tasks was
greater in the younger group. The content facet was represented with one mislocated variable
(numerical knowledge) by a polar partitioning in dimensions two and three. As assumed in the
third hypothesis, the knowledge scales were at least as homogeneous as the reasoning tasks in this
solution.
However, the di€erent domains of knowledge were eliminated in the content subscales of the
knowledge test. This procedure might have enhanced their homogeneity in SSA. To test whether
the homogeneity of the knowledge subscales was due to the way of aggregation, a SSA was per-
formed with the reasoning tasks and the nine domain- and content-speci®c aggregates. The stress
of a three-dimensional metric solution was acceptable (stress=0.118; R2=0.90). The di€erentia-
tion between gf and gc was demonstrated in the ®rst two dimensions. The knowledge aggregates
were grouped according to the domain of knowledge and the reasoning tasks were grouped
according to verbal, numerical, and ®gural content (see Fig. 2).
With the exception of geography-history and daily life Ð which formed only a single aggregate
for the domain Ð all variables with the same domain of knowledge were grouped together. But
even with these aggregates, which were not all extremely reliable, the knowledge scales were
nearly as homogeneous as the reasoning scales. This is in line with the third hypothesis.
The assumed content facet, the di€erenciation of gf and gc, and the comparable homogeneity
of gc tasks and gf tasks were demonstrated with SSA. When the reasoning tasks were analyzed
together in a metric SSA with the content-aggregates and domain-aggregates in the group aged
between 14 and 24 years (N=447), the stress for a three-dimensional solution was moderate
(stress=0.142; R2=0.87). In dimensions one and two, the same con®guration was found as for
the total group, with the di€erentiation of gf and gc on the one hand and the domains of
knowledge on the other. In dimensions two and three a polar content facet was found without
any mislocated variable. For the group of participants aged between 25 and 50 years (N=259)
also a moderate stress was found for a solution with three dimensions (stress=0.139; R2=0.84).
The content facet was also represented in dimensions two and three, however with two mislocated
variables. The two mislocated variables were the verbal knowledge on science and the ®gural
knowledge on mathematics. Instead of being located in their respective area of content (i.e. verbal
or ®gural) they were both located in the numerical area. Probably numerical abilities were
988 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

strongly related to verbal knowledge of science and ®gural knowledge of mathematics in older
people, who work for long periods of time in technical jobs, or in the ®eld of science or mathe-
matics, where they probably had the opportunity to develop their numerical abilities. This could
explain why the domain of science and mathematics correlates strongly with the numerical abil-
ities in the group of older participants. However, 16 variables were correctly located in the polar
content facet, so that the speci®c e€ect did not alter the whole structure.
In order to test the faceted model further, CFA was performed, since generally facets cannot be
shown without speci®c aggregation techniques (e.g. JaÈger, 1982) in exploratory factor analysis.
First, the faceted structure of gf and gc was analyzed with CFA based on the nine reasoning
variables and the six content subscales for knowledge. For all variables, two loadings were set
free: one on a content factor for verbal (V), numerical (N), or ®gural (F) intelligence and one on
the gf- or gc-factor. In order to split up the variance of the reasoning tasks directly, a third
loading on the gc factor was set free for all reasoning tasks with exception of the matrices. Since
one gc-loading of the reasoning tasks had to be ®xed in order to restrict the number of free
parameters, the gc-loading of matrices was ®xed at zero, because the matrices have been used as
marker for gf several times. Interval restrictions were set between 0.30 and 0.80 for all free load-
ings except the gc-loadings of the reasoning tasks, and between 0.30 and 0.60 for all free corre-
lations between factors. The gc-loadings of the reasoning tasks should vary according to their
portion of gc variance. This portion of gc-variance can be zero for some tasks, but is assumed to
be signi®cant in several tasks, even when the gf and gc factor are allowed to correlate. According

Fig. 2. SSA for nine reasoning tasks and nine domain and content speci®c knowledge aggregates: dimensions one and
two; V=verbal abilities; N=numerical abilities; F=®gural abilities.
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 989

to the sample size (N=706), the restriction forces the intended loadings to be at least signi®cant.
The correlations between the content factors on the one hand and gf and gc on the other were
®xed at zero, because it is not possible to estimate free loadings and correlations between factors
simultaneously (see also Bucik & Neubauer, 1996). The present model had an acceptable ®t (see
Table 5). This supports the ®rst hypothesis of a content facet for gf and gc.
The above-mentioned model was also speci®ed for the two age-homogeneous groups. The
model ®ts the data for both groups (for the younger group: 2(DF=63; N=447)=156.69;
p0.001; GFI=0.95; CFI=0.97; for the older group: 2(DF=63; N=259)=104.75; p0.001;
GFI=0.95; CFI=0.98). There was some notable di€erence in the correlation between gf and gc
for the younger and the older groups: the correlation was 0.30 in the younger group, which was
the lowest value allowed with interval restrictions, and it was 0.47 in the older group. However,
even when in multiple group analysis the loadings, factor-intercorrelations, and error variance for
the younger group were ®xed to be equal to values in the older group, the model ®t was accep-
table (2(DF=183; N1=447, N2=259)=353.49; p0.001; GFI=0.92; CFI=0.97).
It is more dicult to prove the content facet of gc with knowledge aggregates, which are not
only aggregated with respect to content, but also with respect to the domain of knowledge.
Nevertheless, the model for the domain- and content-speci®c knowledge scales based on the same

Table 5
Faceted model, loadings of the reasoning tasks and content-speci®c knowledge aggregates on latent variables in CFA
(completely standardized solution)a

V N F gf gc Spec.

Sentence completion 0.42 0.35 0.19 0.63

Analogies 0.53 0.50 0.18 0.39
Similarities 0.44 0.51 0.12 0.50
Calculations 0.58 0.50 0.20 0.30
Number series 0.42 0.62 0.08 0.40
Signs 0.50 0.57 0.06 0.39
Figures 0.78 0.29 ÿ0.10 0.32
Cubes 0.76 0.28 ÿ0.13 0.35
Matrices 0.34 0.53 0.61
Verbal 1 0.30 0.73 0.38
Verbal 2 0.29 0.74 0.37
Numerical 1 0.61 0.62 0.24
Numerical 2 0.50 0.66 0.32
Figural 1 0.30 0.69 0.43
Figural 2 0.30 0.73 0.38
Phi
V 1.00
N 0.47 1.00
F 0.60 0.60 1.00
gf 1.00
gc 0.30 1.00
a
Goodness-of-®t: 2(DF=63; N=706)=182.22 (p0.001); root mean square residual (RMR)=0.048; goodness-of-
®t index (GFI)=0.97; adjusted goodness-of-®t index (AGFI)=0.94; comparative ®t index (CFI)=0.98. V=verbal
intelligence, N=numerical intelligence, F=®gural intelligence, Spec.=speci®c variance.
990 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

parameter setting as the solution in Table 5 had an acceptable ®t (2(63, N=706)=182.22,

p0.001; GFI=0.97; CFI=0.98). When analyzed with separate age groups, the model ®tted for
the younger group (2(105, N=447)=254.70, p0.001; GFI=0.94; CFI=0.95), but it did not ®t
well for the older group (2(105, N=259)=306.71, p0.001; GFI=0.88; CFI=0.91). The
aggregates for verbal knowledge on science and ®gural knowledge on mathematics were located
in the partition for numerical intelligence in SSA. When these two aggregates were allowed to
load also on the factor for numerical intelligence, the model ®t improved signi®cantly
(2di€=49.61, DF=2, p<0.001) and the model ®t became moderate (2(103, N=259)=257.10,
p0.001; GFI=0.90; CFI=0.93). Probably, the domain speci®c knowledge interacts with the
content type of the knowledge in older participants.
In spite of these speci®c results for the content- and domain-speci®c aggregates, the content
factors and gf and gc were demonstrated simultaneously in CFA. This corresponds to the ®rst

Table 6
Hierarchical model, loadings of the reasoning tasks and content-speci®c knowledge aggregates on latent variables in
CFA (completely standardized solution)a

Variables RV RN RF KV KN KF Spec.

First order factors

Sentence completion 0.58 0.66
Analogies 0.81 0.35
Similarities 0.68 0.53
Calculations 0.83 0.31
Number series 0.75 0.43
Signs 0.78 0.40
Figures 0.81 0.35
Cubes 0.75 0.44
Matrices 0.52 0.73
Verbal 1 0.81 0.35
Verbal 2 0.80 0.35
Numerical 1 0.83 0.31
Numerical 2 0.85 0.27
Figural 1 0.77 0.41
Figural 2 0.78 0.39

Second order factors gf gc

RV 0.86
RN 0.86
RF 0.65
KV 0.91
KN 0.91
KF 0.95
Phi 0.65
a
Goodness-of-®t: 2(DF=83; N=706)=429.57 (p0.001); root mean square residual (RMR)=0.067; goodness-of-
®t index (GFI)=0.92; adjusted goodness-of-®t index (AGFI)=0.89; comparative ®t index (CFI)=0.93. RV=verbal
reasoning, RN=numerical reasoning, RF=®gural reasoning, KV, verbal knowledge, KN=numerical knowledge,
KF=®gural knowledge, Spec.=speci®c variance.
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 991

hypothesis and also indirectly to the third hypothesis, since the gc variables were homogeneous
enough to comprise a factor.
Finally, a purely hierarchical model was tested. Therefore, the three content factors for verbal,
numerical, and ®gural reasoning as well as the three content factors for knowledge were con-
ceptualized as latent endogenous variables. The two second order factors were conceptualized as
latent exogenous variables. The intended loadings on the respective factors were set free or they
were ®xed to zero. The model ®t was acceptable (see Table 6), but it was not as excellent as for the
faceted model (see Table 5). Especially, the AGFI is at the lower bound in the hierarchical model
(see Table 6).

4. Discussion

Some new perspectives for the gf±gc theory are presented here, based on three hypotheses,
tested in the present study. The ®rst hypothesis applying to a content facet for verbal, numerical,
and ®gural intelligence for gf and gc can be con®rmed in SSA and CFA. When content- and
domain-speci®c knowledge scales were used, the results were less convincing for the older group
of participants. Probably some speci®c interaction between the domain of knowledge and the
numerical abilities occurred in older participants. Numerical abilities and verbal knowledge about
science as well as ®gural knowledge on mathematics correlated highly in older participants. This
may be due to e€ects of acculturation during work in professions of the natural sciences and of
technology. But for the total group, the content facet could be demonstrated, even with domain-
speci®c and content-speci®c knowledge scales.
The possibility of introducing a content facet in gf and gc may have some important implica-
tions for the assessment of these constructs. The measurement of construct irrelevant variance as
pointed out by Flanagan and McGrew (1997, i.e. the measurement of gc with a scale for verbal
intelligence) can be reduced, when the content is balanced out in scales for the measurement of gf
and gc. In the German-speaking area, a test for gf and gc, in which the content is balanced out, is
already available (see Amthauer et al., 1999).
An alternative model which was strictly hierarchical and which comprised verbal, numerical,
and ®gural reasoning and verbal, numerical, and ®gural knowledge factors as well as second
order factors corresponding to gf and gc was tested. The hierarchical model had an acceptable ®t,
but the ®t was not as excellent as for the faceted model. The hierarchical model can be regarded
as an alternative to the faceted model, but it is empirically less compelling. Moreover, the faceted
model has the theoretical advantage of integrating aspects of Guttman's radex model. However,
the similarity to the radex model is not complete, since, according to the third hypothesis, no
radial partitioning is expected.
The third hypothesis refers to the homogeneity of the knowledge aggregates. The knowledge
scales are supposed to be as homogeneous as the reasoning tasks. Therefore, they would not form
a radex structure in SSA. This was shown in the respective analyses. Furthermore, the fact that a
gc factor could be demonstrated within CFA, even with domain- and content-speci®c aggregates,
underlines the homogeneity of these aggregates. It should be noted that three content types (ver-
bal, numerical, and ®gural) and ®ve domains of knowledge are included in the tests (geography-
history, science, mathematics, art-literature, daily life) so that the material is perhaps more het-
992 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

erogeneous than the verbal and vocabulary scales often used to measure gc. Therefore the
homogeneity cannot be attributed to speci®city of the tests.
Analyses for di€erent age groups revealed, that the proposed faceted structure for gf and gc has
some stability across age. However, age seems to have a moderating e€ect on the di€erentiation
of gf and gc. The weight for the ®rst dimension was far more important in the weighted SSA for
individual di€erences in the younger group than in the older group. Thus, the di€erentiation
between gf and gc, which goes along with the ®rst dimension in SSA, seems to be more pro-
nounced in the younger participants. In addition, the correlation between gf and gc was higher in
the older group in separate CFA. Because this is a cross-sectional result, it should be interpreted
with caution, because they may represent mainly cohort di€erences (e.g. Schaie & Baltes, 1977).
However, the present result is not common, since mostly the correlation between gf and gc is
expected to decrease with age (e.g. Horn & Cattell, 1982). This would be in line with the early
hypothesis of age-di€erentiation of cognitive abilities (e.g. Garrett, 1946). But according to Car-
roll (1993), the age-di€erentiation of cognitive abilities is hard to demonstrate. Moreover, already
Balinsky (1941) reported an increase of the general factor variance in the older age group, which
contradicts the hypothesis of age-di€erentiation as well as the assumptions in Horn and Cattell
(1982). For the old and very old age, Lindenberger and Baltes (1997) were able to support the
dedi€erentiation hypothesis, i.e. higher intercorrelations of cognitive tests in old and very old
participants. The present results would be in line with the results reported in Lindenberger and
Baltes (1997), because they may be interpreted as a small but early beginning of cognitive ded-
i€erentiation. However, the present results should be replicated and cannot fully be interpreted in
a developmental perspective until longitudinal data become available. Since in the present con-
ceptualization, gf and gc are measured in tasks with verbal, numerical, and ®gural content, the
developmental results may di€er from those reviewed in Cattell (1987), as well as Horn and Noll
(1997). However, the important result in the present context is the stability of the assumed
structure across age groups.
In further research, it would be interesting to examine whether the present faceted approach to
gf and gc can be related to existing facet models of intelligence, as for example the BIS (JaÈger, 1982).
The relations to Guttman's radex model should also be examined further, because the present
approach focuses only on tasks classi®ed as rule inference and rule application, and tasks classi®ed
as learning were not considered here. In addition, the extension of the faceted approach presented
here on the broader view of gf and gc (e.g. Lindenberger & Baltes, 1997) including reasoning,
memory, and perceptual speed in gf and knowledge and ¯uency in gc should be examined.

References

Ackerman, P. L. (1996). A theory of adult intellectual development: Process, personality, interests, and knowledge.
Intelligence, 22, 227±257.
Adler, N., & Guttman, R. (1982). The radex structure of intelligence: A replication. Educational and Psychological
Measurement, 42, 739±748.
Amthauer, R. (1971). I-S-T 70 (3rd ed.). GoÈttingen: Hogrefe.
Amthauer, R., Brocke, B., Liepmann, D., & Beauducel, A. (1999). Intelligenz-Struktur-Test 2000. GoÈttingen: Hogrefe.
Balinsky, B. (1941). An analysis of the mental factors of various age groups from 9-60. Genetic Psychology Mono-
graphs, 23, 191±234.
 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994 993

Baltes, P. B., Lindenberger, U., & Staudinger, U. M. (1998). Life-span theory in developmental psychology. In R. M.
Lerner, Theoretical models of human development: Handbook of child psychology5th ed., Vol. 1. New York: Wiley.
Borg, I., & Lingoes, J. C. (1987). Multidimensional similarity structure analysis. New York: Springer.
Brocke, B., Beauducel, B., & Tasche, K. G. (1998). Der Intelligenz-Struktur-Test: Analysen zur theoretischen Grund-
lage und technischen GuÈte. Diagnostica, 44, 84±99.
Bucik, V., & Neubauer, A. C. (1996). Bimodality in the Berlin Model of Intelligence Structure (BIS): A replication
study. Personality and Individual Di€erences, 21, 987±1005.
Carroll, J. B. (1993). Human cognitive abilities. A survey of factor-analytic studies. Cambridge: Cambridge University
Press.
Carroll, J. D., & Chang, J. J. (1970). Analysis of individual di€erences in multidimensional scaling via an N-way gen-
eralization of ``Eckard-Young'' decomposition. Psychometrika, 35, 283±319.
Cattell, R. B. (1963). Theory of ¯uid and crystallized intelligence: A critical experiment. Journal of Educational Psy-
chology, 54, 1±22.
Cattell, R. B. (1987). Intelligence: Its structure, growth, and action. Amsterdam: Elsevier Science Publishers BV.
Flanagan, D. P., Genshaft, J. L., & Harrison, P. L. (1997). Contemporary intellectual assessment. Theories, tests, and
issues. New York: The Guilford Press.
Flanagan, D. P., & McGrew, K. S. (1997). A cross-battery approach to assessing and interpreting cognitive abilities:
Narrowing the gap between practice and cognitive science. In D. P. Flanagan, J. L. Genshaft, & P. L. Harrison,
Contemporary intellectual assessment. Theories, tests, and issues. New York: The Guilford Press.
Garrett, H. E. (1946). A developmental theory of intelligence. American Psychologist, 1, 372±378.
Gilardi, R., Holling, H., & Schmidt, J. U. (1983). Replikationsanalyse zum CATTELLschen Intelligenzkonzept. Psy-
chologische BeitraÈge, 25, 369±384.
Guilford, J. P. (1967). The nature of human intelligence. New York: McGraw-Hill.
Guilford, J. P. (1980). Fluid and crystallized intelligences: Two fanciful concepts. Psychological Bulletin, 88, 406±412.
Gustafsson, J. E. (1984). A unifying model for the structure of intellectual abilities. Intelligence, 8, 179±203.
Guttman, L. (1965). A faceted de®nition of intelligence. In R. Eiferman, Studies in Psychology, scripta hier-
osolymitanaVol. 14. Jerusalem: The Hebrew University.
Guttman, L., & Levy, S. (1991). Two structural laws for intelligence tests. Intelligence, 15, 79±103.
Hakstian, A. R., & Cattell, R. B. (1978). Higher-stratum ability structures on a basis of twenty primary abilities.
Journal of Educational Psychology, 70, 657±669.
HaÈrnqvist, K., Gustafsson, J. E., MutheÂn, B. O., & Nelson, G. (1994). Hierarchical models of ability at individual and
class levels. Intelligence, 18, 165±187.
Horn, J. (1988). Thinking about human abilities. In J. R. Nesselroade, & R. B. Cattell, Handbook of multivariate
experimental psychology (pp. 645±685). New York: Plenum Press.
Horn, J. L., & Cattell, R. B. (1966). Re®nement and test of the theory of ¯uid and crystallized general intelligences.
Journal of Educational Psychology, 57, 253±270.
Horn, J. L., & Cattell, R. B. (1982). Whimsy and misunderstandings of gf±gc theory: A comment on Guilford. Psy-
chological Bulletin, 91, 623±633.
Horn, J. L., & Noll, J. (1997). Human cognitive capabilities: Gf±Gc theory. In D. P. Flanagan, J. L. Genshaft, & P. L.
Harrison, Contemporary intellectual assessment. Theories, tests, and issues. New York: The Guilford Press.
Humphreys, L. G. (1962). The organisation of human abilities. American Psychologist, 17, 475±483.
JaÈger, A. O. (1982). Mehrmodale Klassi®kation von Intelligenzleistungen: Experimentell kontrollierte Weiter-
entwicklung eines deskriptiven Intelligenzstrukturmodells. Diagnostica, 23, 195±225.
JoÈreskog, K. G., & SoÈrbom, D. (1993). lisrel 8-User's reference guide. Chicago, IL: Scienti®c Software International.
Kaufman, A. S., Kaufman-Packer, J. L., McLean, J. E., & Reynolds, C. R. (1991). Is the pattern of intellectual growth
and decline across the adult life span di€erent for men and women?. Journal of Clinical Psychology, 47, 801±812.
Lindenberger, U., & Baltes, P. B. (1997). Intellectual functioning in old and very old age: Cross-sectional results from
the Berlin Aging Study. Psychology and Aging, 12, 410±432.
Marshalek, B., Lohman, D. F., & Snow, R. E. (1983). The complexity continuum in the radex and hierarchical models
of intelligence. Intelligence, 7, 107±127.
Pry, R. (1995). Intelligence CristalliseÂe et WISC-R. Revue europeÂenne de Psychologie AppliqueÂe, 54, 219±225.
994 A. Beauducel et al. / Personality and Individual Di€erences 30 (2001) 977±994

Schaie, K. W., & Baltes, P. B. (1977). Some faith helps see the forest: A ®nal comment on the Horn and Donaldson
myth of the Baltes-Schaie position on adult intelligence. American Psychologist, 32, 1118±1120.
Schoenthaler, S. J., Amos, S. P., Eysenck, H. J., Peritz, E., & Yudkin, J. (1991). Controlled trial of vitamin-mineral
supplementation: E€ects on intelligence and performance. Personality and Individual Di€erences, 12, 351±362.
Shye, S. (1988). Inductive and deductive reasoning: A structural reanalysis of ability tests. Journal of Applied Psychol-
ogy, 73, 308±311.
Staudinger, U. M., Maciel, A. G., Smith, J., & Baltes, P. (1998). What predicts wisdom-related performance? A ®rst
look at personality, intelligence, and facilitative experimental contexts. European Journal of Personality, 12, 1±17.
Sternberg, R. J., & Gardner, M. K. (1983). Unities in inductive reasoning. Journal of Experimental Psychology: Gen-
eral, 112, 80±116.
Sternberg, R. J., & Gastel, J. (1989). If dancers ate their shoes: Inductive reasoning with factual and counterfactual
premises. Memory & Cognition, 17, 1±10.
Sternberg, R. J., & Powell, J. S. (1982). Theories of intelligence. In R. J. Sternberg, Handbook of human intelligence
Cambridge: University Press.
Thorndike, R. L., Hagen, E. P., & Sattler, J. M. (1986). Technical manual: Stanford-Binet-Intelligence Scale (4th ed.).
Chicago: Riverside.
Wechsler, D. (1991). Wechsler Intelligence Scale for Children (3rd ed.). San Antonio, TX: Psychological Corporation.
Wittmann, W. W. (1988). Multivariate reliability theory: Principles of symmetry and successful validation strate-
giesNesselroade, J. R., & Cattell, R. B.Handbook of multivariate experimental psychology (2nd ed.). New York:
Plenum (S. 505±560).