Saturday, June 9, 2018

What predicts achievement better: g or broad abilities?



McGill, R. J., & Busse, R. T. (2015). Incremental validity of the WJ III COG: Limited predictive effects beyond the GIA-E. School Psychology Quarterly30(3), 353.  https://pdfs.semanticscholar.org/f5b5/d70077a1b7747a31bbcd5fb7b7dfcc38c2a3.pdf

What predicts achievement better:  g or broad abilities?

Well, it depends on the model of intelligence that you're using, and on the way the statistical analysis was done in the research on which you base your conclusion.

The WJ III COG examiner manual encourages primary interpretation at the broad ability level (e.g., CHC-related cluster scores). Because linking performance in reading/writing/math to the state of the child's cognitive abilities is a major use of intelligence tests, examining relationships between WJ III COG cluster scores and external achievement measures is important. These examinations are also critically important for evaluating the tenability of several models that have been proposed for use in the identification of specific learning disabilities (SLD) in children and adolescents. These and similar models utilize lower-order scores, such as the WJ III COG broad ability clusters, as a critical component for determining whether or not an individual has a learning disability.

According to Flanagan's model, a child is learning disabled if: A.  he has significantly poor reading/writing/math achievement.  B.  he has one (or two) significantly low broad ability scores.  C.  the low broad ability scores can explain the child's poor performance in reading/writing/math.  D.  the child's other broad abilities are average or above average.  E.  excluding factors (like insufficient or inappropriate instruction or emotional problems) are not better explanations of the poor reading/writing/math achievement.

Shortly after the publication of the WJ III COG, McGrew and his colleagues utilized multiple regression to examine predictive relationships between WJ III COG CHC clusters and standardized reading, writing and math measures. Their analyses provided evidence for differential predictive effects across the age span for specific CHC clusters.  I'd reviewed one of McGrew's studies in a prior post.  Here are slides from that post (click to enlarge). 









These slides contain valuable information that can help us plan the diagnostic process, identify the child's difficulties and plan an intervention.

But  McGill and Busse argue that these studies did not control for potential effects of the common variance shared by mental measures.  The common variance is a manifestation of g.  Each subtest and broad ability measures both g and the specific thing it is supposed to measure.  If a broad ability is "saturated" with g, it is a good measure of g but a poorer measure of the unique construct it's supposed to measure.
  
To investigate the tenability of the recommendation for practitioners to interpret primarily at the broad ability level, it is necessary to examine the incremental predictive validity provided by the broad abilities after controlling for the effects of variance already accounted for by the full scale score. That is, the extent to which broad abilities add meaningful information beyond what we get from the full scale IQ score.

The participants in this study were children and adolescents ages 6–0 to 18–11 (n = 4,722) drawn from the standardization sample for the WJ III COG and the WJ III ACH.  The WJ III COG measures 7 broad abilities: Comprehension Knowledge, Fluid Reasoning, Long Term Storage and Retrieval, Short Term Memory, Visual Processing, Auditory Processing and Processing Speed. The General Intellectual Ability – Extended (GIA-E) score is composed of 14 subtests, 2 subtests for each broad ability.

McGill and Busse used Hierarchical multiple regression analysis to analyze the data.   In this procedure, the full scale score is entered first into a regression equation followed by the lower order factor or cluster scores to predict a criterion achievement variable. This entry technique allows for the predictive effects of the cluster scores to be assessed while controlling for the effects of the full scale score.

The authors found that GIA-E accounted for statistically significant portions of each of the WJ III ACH cluster scores: Broad Reading, Basic Reading, Reading Comprehension, Broad Mathematics, Math Calculation, Math Reasoning, Broad Written Language, Basic Writing, Written Expression, Oral Expression, and Listening Comprehension (I'm not sure oral expression and listening comprehension are really achievement areas.  I think they belong more to the comprehension knowledge cluster and they are also affected by fluid reasoning).  The GIA-E accounted for 29% (Math Calculation Skills) to 56% (Listening Comprehension; Mdn 46%) of the criterion variable variance. 

CHC clusters entered jointly into the second block of the regression equations accounted for 2% (Math Reasoning) to 23% (Oral Expression; Mdn 5%) of the incremental variance.   The 23% incremental variance in Oral Expression was predicted by the Comprehension Knowledge cluster.  But even this unique outcome is dubious:  oral expression tests measure linguistic competency and vocabulary, that is – Comprehension Knowledge.  So it's hardly surprising that Comprehension Knowledge predicts Comprehension Knowledge…None of the other broad abilities accounted for more than 5% of achievement variance beyond GIA-E.   Although the CHC clusters contributed significant portions of incremental achievement variance beyond the effects of the GIA-E, effect size estimates were negligible.  The results from the current study indicate that practitioners who interpret CHC cluster scores on the WJ III COG, without accounting for the effects of the GIA-E risk overestimating the predictive effects of various CHC-related abilities

These results are fairly consistent with those that have been obtained from other cognitive measures like the Wechsler test.  

But how does all this fit with McGrew's results presented in the slides above?

McGill and Busse write that reverse entry of the independent variables, in this case entering the broad ability clusters first, would result in the clusters accounting for approximately the same variance proportions that were attributed to the GIA-E in this study. Consequently, the GIA-E would provide little incremental prediction. Order of entry arbitrarily determines whether scores such as the GIA mean everything or nothing.  However, order of entry is not an arbitrary process and must be determined a priori according to expected theoretical relationships between the variables and causal priority. Contemporary intelligence theory (e.g., CHC) and the WJ III COG structural model support entering the GIA-E before the clusters because the cluster scores are both theoretically and statistically subordinate to the GIA-E. Reverse entry conflicts with existing intelligence theory and violates the scientific law of parsimony (if you can predict something with one variable with the same level of success as with many variables, prefer the one over the many).

The CHC model is an integration between the Cattell and Horn model and Carroll's model.  Cattell and Horn presented a model of intelligence that included broad abilities but did not include g.  Carroll presented a model that included both g and broad abilities. To this day, many questions remain as to whether g reflects an actual latent ability or is merely a statistical artifact resulting from the tendency for all tests of mental ability to be positively correlated.

So, if I believe there's no g (like Cattell and Horn), the first thing I'll enter into the regression equation would be the broad abilities, and the results I'd get would be that the broad abilities predict achievement pretty well.  On the other hand, If I believe in g (like Carroll), I'll enter g first and find that it's the best predictor of achievement and that broad abilities do not add any meaningful incremental value…

What's more, it's obvious that FSIQ predicts achievement.  Obviously, it predicts achievement better than any broad ability, since it embodies the influence of all broad abilities.  But this is not the interesting question.  What we want to know is to what extent do the components of g, that is the broad abilities, predict achievement.  I think this is a good theoretical reason to enter broad abilities first into the regression equation whether we believe in g or not.  We don't want to use g to predict achievement because it's too broad and doesn't lead to meaningful interventions.

McGill and Busse point out that incremental validity researchers have largely relied on archived standardization data to assess the predictive effects of cognitive test scores. This is problematic given that the two incremental validity studies that have been conducted using data obtained from clinical samples have found significantly diminished effects associated with the general factor with greater portions of achievement variance accounted for by factor-level scores.

Additionally whereas the cognitive variables consistently accounted for large portions of achievement variance, approximately half of the variance in the WJ III ACH variables was left unpredicted in this study. What explains this additional variance?  Maybe noncognitive variables like motivation or effort.


No comments:

Post a Comment