In June 2016 I published a post which
discussed a paper by Prof. Ryan McGill.
In that paper McGill argued against the PSW operational definition of
learning disability (usually using CHC model). I
disagreed with him on a few points, and was delighted that he read the post and responded to it.
This post discusses a new paper by Prof.
McGill. In this paper he deals with the
question of what predicts reading better – the general ability (the IQ score)
or broad CHC abilities. Before I present
his arguments (and my response to them), I think it's time to introduce Prof.
McGill, especially since he is our (educational psychologists)
colleague.
Prof. Ryan J. McGill
Ryan J. McGill is Assistant Professor and Director of
the School Psychology Program at the College of
William & Mary. His research and
teaching interests include psychological assessment and measurement, judgement
and decision-making in school psychology, assessment and identification of
specific learning disability, and quantitative psychology. Previously, he was
a faculty member in the school psychology programs at Texas Woman’s University
(2014-2016) and Chapman University (2011-2014). Additionally, he was a
practicing school psychologist in Southern California from 2009-2014.
These are the main points
in McGill's paper as I understood them:
With regard to CHC cognitive-achievement
relationships, a series of highly cited predictive validity investigations
provide evidence of differential predictive effects for CHC-related broad
abilities for reading, mathematics, and written language achievement across the
age span. With regard to reading achievement, Evans and colleagues (2001) examined
relations between CHC broad and narrow abilities and various reading abilities
across 14 age groups. Their findings suggest that several abilities (i.e.,
Phonological Awareness, Processing Speed [Gs], and Long-Term Retrieval [Glr])
consistently accounted for significant effects in reading achievement across
the age span. As a result, they encouraged primary consideration of these and
other related CHC broad and narrow abilities when investigating reading skill
development. However, Evans et al. (2001) did not to include an estimate of
general intelligence in their prediction models, a variable that has a rich
history of accounting for meaningful levels of academic achievement variance.
Thorndike (1986) noted that 85% to 90% of
predictable variance in measures of achievement may be accounted for by a
single general score (i.e., FSIQ), that is thought to estimate general
intellectual ability. As a consequence of Evans et al.'s (2001) omission of an estimate
of general intelligence, the unique contributions of CHC broad and narrow
abilities in predicting reading abilities above and beyond a more parsimonious
general intelligence dimension is unclear.
Many researchers debated which factor has
more influence on reading – the general ability or broad abilities – and found
evidence for both sides of the argument.
Floyd, Keith, Taub, and McGrew (2007) examined the latent predictive
effects of CHC-related abilities on reading decoding. In contrast, to Evans et
al. (2001), the authors chose to model a general intelligence factor and found
that its influence on reading was mediated through the Stratum II broad
abilities. Utilizing the same methodological approach, Benson (2007) later
concluded that g had direct and significant effects on reading abilities and
that the effects associated with the broad and narrow abilities was mostly
small. Alternatively, Floyd, Meisinger, Gregg, and Keith (2012) suggested that
an integrative model with both direct and indirect effects from g best
predicted reading comprehension across development.
Given the multidimensionality inherent in
contemporary measures of CHC-related abilities, researchers have consistently
recommended that the effects of the higher-order g-factor should be partialed
out or controlled for prior to making inferences regarding the relative
importance of lower-order cognitive variables. Failure to do so, may risk
overestimating the effects of lower-order variables at the expense of the higher-order
dimension.
McGill attempted to measure the influence
of broad abilities on reading above and beyond the influence of the general
ability. The participants were children
and adolescents ages 6-0 to 18-11 (N = 4,722) drawn from the standardization
sample for the WJ III. The WJ III COG is
a multidimensional test of general intelligence for ages 2 to 90 years. The
measure is comprised of 30 subtests, 14 of which contribute to the measurement
of seven CHC-based Stratum II broad cluster scores: Comprehension-Knowledge
(Gc), Fluid Reasoning (Gf), Auditory Processing (Ga), Visual-Spatial Thinking
(Gv), Short-Term Memory (Gsm), Long-Term Retrieval (Glr), and Processing Speed
(Gs). Additionally, six clinical cluster scores (Phonemic Awareness, Working
Memory, Broad Attention, Cognitive Fluency, Executive Processes, and Delayed
Recall) thought to reflect more narrow CHC dimensions are also available
through different configurations of the subtests. McGill wanted to know to what extent each of
these cognitive measures predicts reading.
Reading was assessed with the
WJ3ACH. The WJ III-ACH is a
comprehensive academic assessment battery designed to measure five academic
domains: Reading, Written Language, Mathematics, Oral Language, and Academic
Knowledge. The WJ III ACH is comprised of 22 subtests that combine to provide
17 broad clusters and a total achievement composite score.
McGill
found that the general ability score (GIA) accounted for statistically
significant (p < .05) portions of the Basic Reading scores in all of the age
brackets that were assessed. Across the 13 regression models utilized to
predict Basic Reading, the GIA accounted for 40% (age 12) to 63% (age 17; M =
49%) of the criterion variance. Broad
clusters entered jointly into the second block of the regression equations
accounted for 2% (age 17) to 6% (ages 6; M = 4%) variance beyond g. Narrow clusters entered jointly into the
second block of the regression equations accounted for 1% (ages 8-10, 16, 17)
to 7% (age 7; M = 2%) additional variance beyond g.
The
GIA accounted for statistically significant (p < .05) portions of the
Reading Comprehension scores in all of the age brackets that were assessed.
Across the 13 regression models utilized to predict Reading Comprehension, the
GIA accounted for 46% (age 6) to 67% (age 17; M = 61%) of the criterion
variance. Broad clusters entered jointly
into the second block of the regression equations accounted for 3% (age 6) to
10% (ages 17; M = 8%) additional variance beyond g. The incremental variance coefficients
attributed to individual WJ III COG broad clusters ranged from 0% to 9%. With
only the variance coefficient associated with the Comprehension-Knowledge
cluster at age 17 ( = .09) accounting for meaningful amounts of achievement
variance on its own Narrow clusters entered
jointly into the second block of the regression equations accounted for 0% (age
14) to 9% (age 7; M = 4%) additional variance beyond g.
In
light of these findings, McGill argues that the failur
e to replicate the
broader results produced by Evans et al. (2001), suggest that a more
circumspect appraisal of the importance of CHC dimensions in relationship to
the development of reading skills may be needed in the professional literature.
McGill notes that, due to the
hierarchical structure of the measurement instrument, the importance of order
of entry when utilizing Hierarchical Multiple Regression Analysis (HMRA) to
assess the incremental effects of independent variables must also be
considered. Hale, Fiorello, Kavanaugh, Holdnak, and Aloe (2007) demonstrated
that by entering the first-order factor scores from a previous iteration of the
Wechsler Intelligence Scale prior to entering the FSIQ score, the predictive
effects of FSIQ were diminished to the point of being inconsequential. As a
result, Hale and colleagues argued that order of entry arbitrarily determines
whether scores such as the GIA mean everything or nothing due to the long
established fact that variables entered first into a regression equation
capture greater criterion variance than variables entered later. However, order
of entry is not an arbitrary process and must be determined a priori according
to the expected theoretical relationships between variables. The proposed indirect hierarchical structural
model for the WJ III COG (as per CHC theory) support entering the GIA score
prior to the broad and narrow clusters due to the fact that these scores are
subordinate to the GIA. Further, reverse entry conflicts with CHC theory and
constitutes a violation of the scientific law of parsimony.
These are McGill's arguments.
But
I think this seems trivial.
There
is independent evidence that broad and narrow abilities contribute to reading. For instance it's an established fact that
phonological awareness contributes to reading.
It's also known that rapid naming contributes to reading. I think it's indisputable that Comprehension-Knowledge
contributes to basic reading as well as to reading comprehension. Since g is comprised of these and other abilities
that contribute to reading, it's obvious that g will be the best predictor of
reading.
This
doesn't mean that from now on we'll use only the general ability as a predictor
of reading. This is of no interest to us. What we want to know is
which broad abilities within the general ability predict reading. This
is the knowledge that will lead us to better interventions. We can learn that
by not including an estimate of general intelligence in the prediction
model. We don't need these broad abilities to be better predictors
of reading than the general ability. But if some broad abilities
contribute more to reading than others, we want to know that.