The scores of each child fell in the normal range in all cognitive measures (Table 1).
In the WISC-IV FSIQ the good readers significantly outperformed both the dyslexics (p < 0.001) and the preterms (p < 0.031), and the same was true for the VCI (p = 0.032, p = 0.047) and the PeRI (p < 0.002, p < 0.013). In the WMI and the PrSI the good readers had a significant advantage only over the dyslexics (p < 0.009, p < 0.013). In most of the WISC-IV measures there was no difference between the preterms and the dyslexics with the only exception of the PrSI in which case the preterms scored better (p < 0.05).
According to the GLMMs the groups had a significant effect on all the measures (Table 2).
The good readers performed significantly better than the dyslexics on each Rey measures (CT: p = 0.004. CS: p = 0.001. MT: p = 0.027. MS: p < 0.001). The advantage of the good readers over the preterms was significant only in the CS (p < 0.024) and marginal in the MS (p = 0.054) in which the difference between the preterms and dyslexics was also marginally significant (p = 0.064).
Reading and spelling
In each reading and spelling measure both the good readers and the preterms performed significantly better than the dyslexics (in all comparisons: p < 0.001). The scores of the good readers and the preterms did not differ significantly.
According to the GLMM the groups had significant effects on all the 3 measures.
In order to reveal the cognitive background of the reading and spelling abilities a 3-step analyses were performed. Considering the few measurement points and the relatively great number of independent variables, in order to select the meaningful explanatory variables first Random Forests (Hothorn et al. 2006a) were applied. RF is an ensemble of Classification and Regression Trees (see in the next paragraph). The trees are built using random subsets of the data and random subsets of variables chosen at branching point of the trees. The importance of the variables can be estimated by this method (Hothorn et al. 2006b; Strobl et al. 2008).
In the second step a Classification and Regression Tree (CART) was fitted using the previously selected important variables to refine the selection of the explanatory variables and discover the potential interaction; CARTs are built using a nonparametric regression approach. Both numerical and categorical variables can be used to build a tree. The general rule is to split the observations into two parts based on a predictor variable (root) and then to split the subset further based on another or the same variable on a recursive way (Hothorn et al. 2006b).
Finally, general linear mixed models (GLMM) were fitted with the previously selected explanatory variables to prove their significance. The linear mixed model made it feasible to take into consideration the dependency of the data (the children attending the same school cannot be considered independent).
For reading accuracy the following explanatory variables were selected by the Random Forest: Rey MS, FSIQ, Rey CT, and WMI. On the ground of the intercorrelations, the CART eliminated the FSIQ and the Rey CT. The most powerful predictor was the recall accuracy of the Rey figures. The accurate readers had Rey MS > 16, but subjects with Rey MS ≤ 16 had chances to be accurate readers if scored > 106 on WMI. The WMI moderated the contribution of the Rey MS to reading accuracy (Fig. 1).
For reading fluency the Rey MS, PrSI, and WMI came out as important variables at the first step which was corroborated by the CART technique. The most powerful predictor of fluent reading was the recall accuracy of the Rey figures (Rey MS). Working memory and processing speed had mediator roles: Children with lower Rey MS (≤ 14) but having relatively high WMI (> 106) and/or PrSI (> 97) could as well be fluent readers (Fig. 2).
For spelling accuracy RF identified by FSIQ, Rey MS, and VCI as explanatory variables. According to the CART IQs > 105 provide suitable bases for accurate spelling. The role of the FSIQ is moderated by the recall accuracy of the Rey figures and the working memory (VCI). The ideal background structure for spelling accuracy is Rey MS, if FSIQ > 105 and VCI > 121, or if FSIQ < 105 and Rey MS > 21.5 (Fig. 3).
At the first step for spelling speed from among the heavily intercorrelated independent variables the PrSI was eliminated by the Random Forest; then, CART selected FSIQ and VCI as significant predictors. FSIQ was the main explanatory variable and VMI was a mediator: In order to react quickly, children either needed high FSIQ (> 111) or at least relatively high VCI (> 93) (Fig. 4).
As it was noted earlier, the aim of the study did not require the groups to be matched on IQ.
Nevertheless, we checked a potential bias in the data resulted by the higher IQs in the good reader group and lower IQs in the other two groups. We made subsets from the groups by filtering out children with high and low IQs. According to the repeated analysis using this restricted data set (preterms n = 13, dyslexics n = 17, good readers n = 20), the mean values hardly changed after the removal of children with high and low IQs. The results of the comparisons of the cognitive measures in the three groups, using the same general linear mixed models as in case of the total sample, were as follows: No significant difference was found in FSIQ (p = 0.8454), VCI (p = 0.4849), PRI (p = 0.2673), WMI (p = 0.0954), PSI (p = 0.3002), Rey CT (p = 0.0675), Rey_MT (p = 0.0846). However, the differences remained significant in reading accuracy (p < 0.0001), reading fluency (p < 0.0001), spelling accuracy (p = 0.0066), and spelling speed (p = 0.0042), as well as in Rey CS (p = 0.0023) and Rey MS (p = 0.046). These results suggested that the group differences in IQ were very unlikely to influence the results of the analyses (which would not have been feasible with the small data set).