Preliminary Research on Adult LD Identification with the WJ III

Preliminary research evidence on potentially important diagnostic patterns of WJ III scores in adult university subjects was recently extracted from the Gregg/Hoy LD/Normal university sample previously described in this chapter. The results, briefly summarized here, demonstrate the potential of the WJ III in the identification of adolescents and adults with or without learning disabilities.

Select WJ III COG and ACH data from the Gregg/Hoy university sample was subjected to the Classification and Regression Tree (CART) program, a robust set of decision-tree procedures for "data mining" and predictive modeling (Sal-ford Systems, 1999, 2000). Briefly, CART uses computer-intensive and complex data-searching algorithms to identify important patterns and relations in data. CART can uncover hidden structure in very large and highly complex data, even data that may be difficult to analyze with traditional statistical methods (e.g., when a set of variables are highly multicollinear).

Using the dependent categorical variable of LD versus Not-LD (Normal), the CART analyses "grew" a large decision-making tree that resulted in the optimal classification of 204 subjects with scores from the complete set of WJ III COG and ACH clusters. The initial LD/Not-LD decision-tree was set aside by CART and a tenfold internal cross-validation procedure produced 10 new independent trees that were then combined and used to "prune" the original tree. The complete sample was then classified based on the final pruned tree. The results of these analyses, which identify two Normal (Not-LD) and three LD classification groups (called terminal nodes), are presented in Figure 14.4.

The first decision point in the tree indicates that the WJ III Academic Fluency cluster is the single most important variable in differentiating LD from Not-LD university students in this sample. Subjects with WJ III Academic Fluency scores greater than 107 are most likely not LD. Terminal Node 5 includes 81 normal subjects and 13 LD subjects. Thus, this first decision rule results in 86.2% of these 94 subjects correctly classified as Not-LD, and 13.8 % of the LD subjects misclassified (included in Terminal Node 5). Conversely, a WJ III Academic Fluency score less than or equal to 107 produces three other decision-making points and four other terminal nodes (Nodes 1 through 4). Nodes 1 through 4 have classification accuracy figures ranging from 81.0% (Node 2) to 95.4% (Node 1). The complete decision tree indicates the following for university subjects:

• A subject with a WJ III Academic Fluency score greater than 107 is most likely not previously diagnosed as LD. This classification was accurate 86.2% of the time.

• A subject with a WJ III Academic Fluency score less than 107 and a WJ III Basic Writing Skills cluster score less than or equal to 96 is most likely LD (95.4% accuracy).

• A subject with a WJ III Academic Fluency score less than or equal to 107, a WJ III Basic Writing Skills cluster score greater than 96, and a WJ III Verbal Ability-Std cluster (same as the Verbal Comprehension test) score less than or equal to 104 is most likely LD (81.0% accuracy).

Is Verbal

Figure 14.4

WJ III university subject LD/Not-LD classification and decision tree (Gregg/Hoy study, 2001)

Figure 14.4

WJ III university subject LD/Not-LD classification and decision tree (Gregg/Hoy study, 2001)

• A subject with a WJ III Academic Fluency score less than or equal to 107, a WJ III Basic Writing Skills cluster score greater than 96, a WJ III Verbal Comprehension test score greater than 104, and a WJ III Phonemic Awareness (Ga) cluster score less than or equal to 106 is most likely LD (87.5% accuracy).

• Finally, a subject with a WJ III Academic Fluency score less than or equal to 107, a WJ III Basic Writing Skills cluster score greater than 95.5, a WJ III Verbal Comprehension test score greater than 104, and a WJ III Phonemic Awareness (Ga) cluster score greater than 106 is most likely not LD (87.5% accuracy).

Using the above WJ III-based decision-tree rules, the cross-validation classification table revealed classification accuracy rates of 81% (LD) and 85% (Normal). Given that the initial sample was almost equally divided between LD (n = 101) and Not-LD (n = 103) subjects, this classification agreement rate suggests that the WJ III COG and ACH variables included in Figure 14.3 can improve over a 50% chance base-rate classification of newly referred university subjects by approximately 30%. This indicates that the WJ III COG and ACH batteries include measures that may be particularly helpful in the identification of adults with learning difficulties.

Although CART procedures are atheoretical and empirically driven, post-hoc interpretation of the results in Figure 14.4 presents a number of interesting theoretical hypotheses. First, the pivotal WJ III Academic Fluency cluster, which is comprised of the WJ III Reading, Math, and Writing Fluency tests, suggests that, if an individual reaches a state of automaticity in basic academic functioning, this may be a strong indicator that the person does not have any specific disabilities. Academic fluency can be considered the "end state" of academic performance that results from the successful acquisition and integration of basic academic and related cognitive abilities, much like the expert state in the novice/expert cognitive psychology literature. Second, three subgroups of adults with LD may be identified via the inspection of domain-specific constellations of cognitive and achievement abilities. One group (Terminal Node 1) may be subjects with poor automaticity in general academic functioning and low basic skills in writing, with no apparent associated cognitive deficits. The second group (Terminal Node 2) may display poor academic automaticity associated with a weakness in general verbal knowledge or comprehension (Gc). The third group also displays poor academic automaticity, but, instead of relative Gc deficits, they have associated cognitive problems in phonemic awareness (Ga).

An additional source of useful information regarding the WJ III COG and ACH clusters in this study is the CART variable importance output table (Table 14.13). The most important variable is assigned an index score of 100, and all other variables are scaled in relative terms to this anchor point. Variable importance is related to both the potential and actual splitting behavior of a variable, and it is possible for a variable to be very important but not included in the final decision tree (the variable may be a constant "bridesmaid" at each splitting point in the data). This information helps identify possible explanations for the structure in the data and also identifies "surrogate" variables that may be used to make decisions at critical points when a subject is missing data on the decision variable.

The most obvious conclusion from an inspection of Table 14.13 is that a number of variables were lurking just below the surface at critical decision points and, if used as replacements for the critical WJ III variables, may produce similar classification accuracy. A number of substantive conclusions are gleaned from Table 14.13. First, deficient achievement in the language arts domains (reading and writing) is the most obvious characteristic that differentiates university subjects with and without LD. Second, the cognitive and achievement domains that differentiate LD and Not-LD university subjects are primarily those dealing with auditory-linguistic achievement (WJ III Broad Reading, Basic Writing Skills, Basic Reading Skills) and cognitive abilities (WJ III Phoneme/Grapheme Knowledge, Phonemic Awareness, Auditory Processing), and efficient/automatic cognitive and achievement functioning (WJ III Academic Fluency, Cognitive Efficiency, Processing Speed, Short-Term Memory, and Working Memory). This information, in addition to the practical decision-rule tree, suggests that the WJ III COG and ACH clusters that tap these abilities should receive significant attention when evaluating adults for possible learning difficulties. The high level of accuracy achieved via these analyses is most likely a function of the power of the CART procedures combined with a battery of cognitive and achievement measures (WJ III) that cover a wide

TABLE 14.13 WJ-III relative importance variable ratings for LD/Not-LD university sample CART analysis

WJ III Cluster

CHC Ability Domain

Relative Importance

Broad Reading

Grw

Was this article helpful?

0 0
Adult Dyslexia

Adult Dyslexia

This is a comprehensive guide covering the basics of dyslexia to a wide range of diagnostic procedures and tips to help you manage with your symptoms. These tips and tricks have been used on people with dyslexia of every varying degree and with great success. People just like yourself that suffer with adult dyslexia now feel more comfortable and relaxed in social and work situations.

Get My Free Ebook


Post a comment