Sie sind auf Seite 1von 4

Psychological Bulletin Copyright 2005 by the American Psychological Association

2005, Vol. 131, No. 1, 7275 0033-2909/05/$12.00 DOI: 10.1037/0033-2909.131.1.72

Working Memory and Intelligence: Different Constructs.


Reply to Oberauer et al. (2005) and Kane et al. (2005)

Margaret E. Beier Phillip L. Ackerman


Rice University Georgia Institute of Technology

The authors address agreements and disagreements with the M. J. Kane, D. Z. Hambrick, and A. R. A.
Conway (2005) and K. Oberauer, R. Schulze, O. Wilhelm, and H.-M. Su (2005) commentaries on P. L.
Ackerman, M. E. Beier, and M. O. Boyle (2005). They discuss the following issues: (a) the relationship
between working memory (WM) and general intelligence (g), (b) the reanalyses included in the
comments, (c) the use of a fixed-effects model versus a random-effects model for the meta-analysis, (d)
the use of structural equation modeling analyses and structural coefficients as equivocal evidence for the
relationship between WM and intelligence, and (e) the problem of confirmation bias in research on WM.
Although the authors disagree with their commentators about the magnitude of the relationship between
WM and g, in the final analysis it appears that all concerned parties agree that WM and intelligence are
different constructs.

The Kane, Hambrick, and Conway (2005) and Oberauer, tators emphasized the similarity between them. After reading their
Schulze, Wilhelm, and Su (2005) commentaries on Ackerman, comments, we remain unconvinced of the similarity of these
Beier, and Boyle (2005) raise issues about the meta-analysis of constructs.
working memory (WM) and general intelligence (or g) relations Kane et al. (2005) chose a subset of studies and tasks for
and some more general issues. There are several points of dis- reanalysistasks that they believed would allow for a better
agreement between their commentaries and the conclusions of the evaluation of the relationship between WM and general fluid
target article, but there are several points of agreement as well. We intelligence (Gf)/reasoning than found in the Ackerman et al.
begin by discussing issues related to the relationship between WM (2005) meta-analysis. However, the tests designated by Kane et al.
and g and then address some of the methodological questions. We (2005) as markers for Gf/reasoning ability frequently contradict
conclude by discussing ongoing concerns about research on WM the identification of the underlying abilities for these tests. For
and intelligence. example, space relations, cards, and flags are generally considered
tests of speeded rotation; cube comparison and paper folding are
The Relationship Between WM and g spatial visualization tests; and problem solving and necessary facts
tests have complex loadings on fluid, crystallized, and numerical
The main conclusion of Ackerman et al. (2005) was the refuta- abilities (see, e.g., Ackerman, Bowen, Beier, & Kanfer, 2001;
tion of the claim that WM and g are isomorphic constructs (e.g., Carroll, 1993; Lohman, 1979). Across the various sets of tasks
Engle, 2002). We estimated the shared variance between WM and used by Kane et al. (2005), it is doubtful that they well represent
g was about 25%. In their comments, Kane et al. (2005) claimed Gf/reasoning ability. Nonetheless, Kane et al. (2005) reported that
that it is 50%, and Oberauer et al. (2005) claimed that it is 72.3%. the relationship between WM and Gf/reasoning in their subset of
On the fundamental question raised by Ackerman et al. (2005; i.e., studies was larger than that reported in Ackerman et al. (2005). We
same or different constructs?), it seems clear that there is no have no disagreement that a selective sampling might result in a
substantial disagreement with the assertion that WM and g are larger estimated relationship than we reported (though the lack of
indeed different constructs. However, we assert that the differ- fit statistic information makes this representation difficult to eval-
ences between the constructs are substantial, whereas the commen- uate in that context). However, we disagree that this type of
selective inclusion of studies for meta-analysis provides the best
evidence for the size of the relationship. Consequently, we remain
unconvinced that the relationship between WM and g is larger than
We thank Werner Wittmann and Roberto Colom for their helpful that reported in Ackerman et al. (2005).
discussions related to the issues of working memory and intelligence and Also, we were confused by Kane et al.s (2005) statement that
this reply.
a Gf latent variable based on syllogistic and analogical reasoning
Correspondence concerning this article should be addressed to Margaret
E. Beier, Department of Psychology, Rice University, MS-25, 6100 Main
tests is free of deduction- and induction-specific ability; a [work-
Street, Houston, TX 77005, or Phillip L. Ackerman, School of Psychol- ing memory capacity] latent variable based on tests for words and
ogy, Georgia Institute of Technology, MC 0170, 654 Cherry Street, At- for arrows is free of content-specific storage (p. 67). The confu-
lanta, GA 30332-0170. E-mail: beier@rice.edu or phillip.ackerman@ sion seems to stem from their unusual usage of terms for factor
psych.gatech.edu analysis. According to factor analytic theory (see, e.g., Harman,

72
WORKING MEMORY AND INTELLIGENCE: REPLY 73

1976), an individual test variance can be decomposed into com- of these relationships (Hunter & Schmidt, 2000)which was our
mon factor variance, specific factor variance, and error variance. intent. That said, we agree with Oberauer et al. (2005) that a
Given that all specific test variance is orthogonal to common random-effects model would have increased the size of the confi-
factor variance, then it is, by definition, true that a common factor dence intervals we reported. However, what Oberauer et al. did not
is independent of test-specific variance. However, contrary to the acknowledge is that adopting their framework would also have the
implication of Kane et al. (2005), two tests may share variance on effect of reducing the power to show that the relationship between
a number of common factors (including cognitive styles or strat- short-term memory (STM) and ability is statistically different from
egies), so in the example given by Kane et al. (2005), a Gf latent the relationship between WM and ability (something that neither
variable based on syllogistic and analogical reasoning tests might set of commentators would probably view as a positive finding for
actually include deduction-common variance if these processes are their respective approaches). Indeed, recent research has shown
correlated in the sample. Also, the Gf latent variable might include that WM and STM are perhaps more highly related constructs than
variance associated with strategies common to both tests or any originally thought (R. Colom, personal communication, August 5,
number of other influences (see, e.g., Kyllonen, Lohman, & Snow, 2004; Kane et al., 2004) and that STM is more highly related to
1984). As noted by Humphreys (1962), intelligence than researchers have previously suggested (see, e.g.,
Beier & Ackerman, 2004).
from the facet point of view there are no pure tests. One obtains In Ackerman et al. (2005) we acknowledged the structural
better and better measures of an element by adding together as many
equation modeling (SEM) analysis violated many characteristics
elements on other facets as possible. Common variance associated
of a good analysis (i.e., missing data, unequal sample sizes per
with the element is increased while variance associated with other
elements is decreased but cannot be reduced to zero by control of cell), and as Oberauer et al. (2005) pointed out, we were limited to
heterogeneity. (p. 481) using correlational input for the analysis. Indeed, the reason that
the SEM analysis was reported was that its inclusion was required
As pointed out by Oberauer et al. (2005), we did not eliminate by the manuscript reviewers, even though it was highly specula-
WM tasks (WMTs) from our analysis on the basis of recent work tive. We continue to believe that the overreliance on such SEM
identifying certain tasks as poor markers for the construct. Again, analyses in this field may be associated with the fact that they can
we agree that selecting a subset of WMTs would have likely be used to support any number of different approaches; such as the
changed the outcome of the meta-analysis. However, the definition early models with high correlations between WM and greasoning
of a good WMT and, more generally, the definition of the con- that appeared compatible with the conclusion that WM and gGf
struct of WM remain unclear. For example, Oberauer et al. (2005) were highly correlated factors.
indicated that WM is not executive function, meaning that WM We disagree with the commentators that SEM methods provide
is not the supervision of basic processes (p. 61). In contrast, unequivocal evidence for a particular value to the relationship
Kane et al. (2005) stated that WM represents attention-control between WM and g. Rather, we think that many researchers may
capabilities (p. 67). It is unclear how supervision of basic pro- put too much stock in these analyses without recognizing their
cesses and attention-control capabilities differ from attention. indeterminate nature (i.e., the fact that an infinite number of
The lack of clarity was highlighted in a recent review of WM models exist that will fit the data equally well in an SEM analysis
measurement, in which Oberauer (in press) stated that there is no with more than one factor). The indeterminate nature of SEM
clear picture as to what the good WMTs have in common theo- analysis seems especially problematic because structural coeffi-
retically and that the definition advocated by others in the field cients are meant to imply causal relations and because only rarely
that working memory capacity is the ability to control attention in do researchers acknowledge the existence of alternative models.
the face of distraction seems . . . too broad given that some Although researchers are routinely willing to consider small mod-
executive control tasks are only weakly related to other measures ifications of their original models (e.g., nested models), rarely do
of working memory. they explore models that are qualitatively different (i.e., research-
Oberauer et al. (2005) also suggested that the correlation be- ers are generally happy to accept the explanation of the data if the
tween WM and g is underestimated in Ackerman et al. (2005) model fits well; MacCallum & Austin, 2000). Although some WM
because of paradigm, content, and method variance. They asserted researchers have more recently explored alternative models (see,
that these are unwanted sources of variance that usually diminish e.g., Kane et al., 2004), in our view, historically, much WM SEM
the measured correlation with the criterion variables (i.e., if they analysis has suffered from this type of confirmation bias.
do not affect the criterion variable as well) (Oberauer et al., 2005, Oberauer et al. (2005) provided an alternative to the model
p. 62). We disagree. In fact, many reported relations between WM shown in Figure 2 of Ackerman et al. (2005) in their comment
and ability measures are likely inflated by overlapping paradigms, article. Our rationale for setting the loadings on the basis of the
contents, and methods (as we discussed in Ackerman et al., 2005, initial analysis was grounded in extant research (e.g., Carroll,
and as can be seen in the cited literature). As such, we assert that 1993). We chose this method as opposed to allowing the program
it is far more likely that the correlations are overestimated in such to estimate these relationships, as did Oberauer et al. We disagree
cases rather than underestimated. with the point that one SEM model that has a similar fit to the data
has more verisimilitude (for WM and g) when the first-order
Methodological Issues loadings do not conform to the existing data on the structure of
abilities (e.g., Carroll, 1993). Also, it is important to point out that
The fixed-effects model in meta-analytic methods is designed to the procedure for conducting the meta-analysis published in
take into account the universe of correlations and report an average Viswesvaran and Ones (1995) instructed use of the harmonic mean
74 BEIER AND ACKERMAN

Table 1
N for Each Cell of the Correlation Matrix Used to Calculate Harmonic Mean in the Structural Equation Models in Ackerman, Beier,
and Boyle (2005)

WM/AB measures 1 2 3 4 5 6 7 8 9 10 11 12 13

1. WM: verbal
2. WM: verbal with numerical 1,037
3. WM: numerical 3,314 2,475
4. WM: numerical with verbal 1,311 1,490 1,821
5. WM: numerical with spatial 710 610 261 344
6. WM: spatial 792 296 698 344 472
7. AB: verbal 3,698 2,948 3,426 2,336 1,307 719
8. AB: numerical 2,220 1,819 2,957 2,213 1,024 589 3,964
9. AB: spatial 870 562 906 168 460 405 353
10. AB: reasoningverbal 668 662 1,414 139 714 562 549
11. AB: reasoningnumerical 285 961 1,194 399 139 1,055 1,055 94 701
12. AB: reasoningspatial 610 392 680 392 233 486 392 377 680 139
13. AB: Raven 708 253 708 97 940 183 736 684 227 52
14. AB: knowledge 1,154 1,990 2,388 1,970 904 296 3,224 3,043 726 961 556 1,810

Note. WM working memory, AB ability. A dash indicates an empty cell corresponding to an imputed value in the correlation matrix shown in Table
4 in Ackerman et al. (2005).

of the sample sizes in each cell of the correlation matrix, as takes a much more skeptical view of latent correlations when the
opposed to the harmonic mean of the sample size of each study. raw data suggest that the relations are weak.
The arithmetic mean of 114 subjects calculated by Oberauer et al.
was the mean number of participants per study, not the mean
number of the sample sizes reported in each cell of the intercor- References
relation matrix. Table 1 shows the number of participants per cell
Ackerman, P. L., Beier, M. E., & Boyle, M. O. (2005). Working memory
used to calculate the harmonic mean.
and intelligence: The same or different constructs? Psychological Bul-
Although both comments emphasize the importance of the SEM letin, 131, 30 60.
analysis, neither take into account that the proof of the utility of Ackerman, P. L., Bowen, K. R., Beier, M. E., & Kanfer, R. (2001).
these constructs is not in the abstract realm of latent constructs but Determinants of individual differences and gender differences in knowl-
rather in the degree of prediction that can be attained with real edge. Journal of Educational Psychology, 93, 797 825.
data. Oberauer et al. (2005) came close to this, by appealing to the Beier, M. E., & Ackerman, P. L. (2004). A reappraisal of the relationship
use of aggregation methods (which we also agree with and dis- between span memory and intelligence via best evidence synthesis.
cussed extensively in Ackerman et al., 2005). Aggregation is an Intelligence: A Multidisciplinary Journal, 32, 607 619.
excellent idea for predictive purposes, but ultimately the aggrega- Carroll, J. B. (1993). Human cognitive abilities: A survey of factor-analytic
tion of different WM composites yields a result that is similar to studies. New York: Cambridge University Press.
aggregating tests to yield a g composite. What is gained by Engle, R. W. (2002). Working memory capacity as executive attention.
Current Directions in Psychological Science, 11, 19 23.
aggregation has to be balanced against the loss of construct clarity
Harman, H. H. (1976). Modern factor analysis (3rd ed.). Chicago: Uni-
for WM (which is what Oberauer et al., 2005, claimed is the
versity of Chicago Press.
limitation of g theory in their comment). However, many adherents Humphreys, L. G. (1962). The organization of human abilities. American
of g endorse the idea of aggregation, in that general intelligence is Psychologist, 17, 475 483.
thought to represent the entire repertoire of knowledge, skills, and Hunter, J. E., & Schmidt, F. L. (2000). Fixed effects vs. random effects
so on. Where construct clarity is gained with aggregation for g, it meta-analysis models: Implications for cumulative research knowledge.
is lost for WM, because WM theories are predicated on specific International Journal of Selection and Assessment, 8, 275292.
processes. Kane, M. J., Hambrick, D. Z., & Conway, A. R. A. (2005). Working
On the central question of whether WM and g are isomorphic, memory capacity and fluid intelligence are strongly related constructs:
there appears to be general agreement that they are indeed different Comment on Ackerman, Beier, and Boyle (2005). Psychological Bulle-
constructs (although there remains some disagreement about the tin, 131, 66 71.
Kane, M. J., Hambrick, D. Z., Tuholski, S. W., Wilhelm, O., Payne, T. W.,
size of the relationship). However, we disagree with our commen-
& Engle, R. W. (2004). The generality of working memory capacity: A
tators that the choice of a subset of tasks and studies with arguable
latent-variable approach to verbal and visuospatial memory span and
representativeness and, in our view, dubious identification pro- reasoning. Journal of Experimental Psychology: General, 133, 189
vides a more accurate estimate of the relation between WM and g 217.
than that reported in Ackerman et al. (2005). Contrary to what the Kyllonen, P. C., Lohman, D. F., & Snow, R. E. (1984). Effects of aptitudes,
commentators may have asserted, we do think there is some value strategy training, and task facets on spatial task performance. Journal of
in exploring the WM construct. However, the choice for research- Educational Psychology, 76, 130 145.
ers is between an approach with a confirmation bias and one that Lohman, D. F. (1979). Spatial ability: A review and reanalysis of the
WORKING MEMORY AND INTELLIGENCE: REPLY 75

correlational literature (Tech. Rep. No. 8). Stanford, CA: Stanford ment on Ackerman, Beier, and Boyle (2005). Psychological Bulletin,
University, School of Education. 131, 61 65.
MacCallum, R. C., & Austin, J. T. (2000). Applications of structural Viswesvaran, C., & Ones, D. S. (1995). Theory testing: Combining psy-
equation modeling in psychological research. Annual Review of Psychol- chometric meta-analysis and structural equations modeling. Personnel
ogy, 51, 201226. Psychology, 48, 865 885.
Oberauer, K. (in press). The measurement of working memory capacity. In
O. Wilhelm & R. W. Engle (Eds.), Handbook of understanding and
measuring intelligence. Thousand Oaks, CA: Sage. Received August 10, 2004
Oberauer, K., Schulze, R., Wilhelm, O., & Su, H.-M. (2005). Working Revision received September 4, 2004
memory and intelligenceTheir correlation and their relation: Com- Accepted September 8, 2004

Das könnte Ihnen auch gefallen