The neurocognitive architecture of fluid ability in children and adolescents

Fluid ability is the capacity to solve novel problems in the absence of task-specific knowledge, and is highly predictive of outcomes like educational attainment and psychopathology. Here, we modelled the neurocognitive architecture of fluid ability in two cohorts: CALM (N=551, aged 5-17) and NKI-RS (N=335, aged 6-17). We used multivariate Structural Equation Modelling to test a preregistered ‘watershed model’ of fluid ability. We found that the watershed model fit the data well for both samples: White matter contributed to working memory and processing speed, which, in turn, contributed to fluid ability (R2(CALM)=51.2%, R2(NKI-RS)=78.3%). The relationship between cognitive abilities and white matter differed with age and showed a dip in strength around ages 7-12 years. Speculatively, this age-effect may reflect a reorganization of the neurocognitive architecture around pre- and early puberty. Overall, these findings highlight that fluid ability is part of a complex hierarchical system of partially independent effects.

Working memory, g f and processing speed are separable constructs. 126 2.
Individual differences in g f are predicted by working memory and processing speed. The Single-Factor Model ( Figure 2) showed acceptable absolute fit (Table 1) (Table 1) but also approximately 0% probability of being data-generating model. Two-Factor 154 Model B (Figure 2, treating working memory and g f as a unitary factor) fit the data well (Table   155 1). Two-Factor Model B showed low probability of being the data-generating model for NKI-156 RS (AIC weight = 3.40%) but a high probability of being the data-generating model for CALM  Even though working memory, processing speed and g f were highly correlated in both 164 samples (Table 2), this pattern of results indicates that processing speed formed a clearly 165 separable factor from working memory and g f in both samples. Working memory and g f were 166 also clearly separable in NKI-RS, but not in CALM. To facilitate comparison across samples and 167 in accordance with our preregistered analysis plan we used the Three-Factor Model (  . We therefore modelled each of the ten white matter tracts separately in 178 all subsequent models. .030). To understand whether working memory and processing speed made significant joint 186 and independent contributions to g f , we compared a freely-estimated MIMIC model to one in 187 which regression paths from processing speed and working memory to g f were constrained 188 to zero. The freely-estimated MIMIC model fit better than a model in which these regression 189 paths were constrained to zero (CALM: ∆χ 2 (2) = 250.20, p < .001; NKI-RS: ∆χ 2 (2) = 199.67, p < 190 .001). This indicates that working memory and processing speed jointly contributed to g f . In 191 CALM (∆χ 2 (1) = 15.53, p < .001), but not NKI-RS (∆χ 2 (1) = 3.25, p = .072), the freely-estimated 192 MIMIC model also fit better than a model in which paths from processing speed and working 193 memory to g f were constrained to be equal. Working memory showed a greater effect (as 194 indicated by standardized path estimates) than processing speed in both samples, but the 195 difference was more pronounced in CALM (Table 3). Contrary to our prediction, therefore, 196 processing speed seemed to make a negligible contribution to g f above and beyond working 197 memory. speed and g f . We examined path estimates and alternative configurations of the model to 202 assess whether the main prediction from the watershed model holds, namely, that white 203 matter contributes to working memory capacity and processing speed, which, in turn, 204 contribute to g f . 205 We found largely converging results across samples. The watershed model showed good fit in turn, explained even more variance in g f (R 2 CALM = 51.2%; R 2 NKI-RS = 78.3%).

211
Comparing the freely estimated watershed model to alternative, constrained, models 212 showed that white matter contributed significantly to memory and processing speed.

278
Using SEM trees, we investigated whether the associations among cognitive and neural 279 dimensions varied with age. As per our preregistration, we investigated potential age-280 differences in the relationships between g f , working memory and processing speed. In 281 additional, exploratory, analyses we investigated potential age differences in significant 282 watershed paths between white matter and cognitive endophenotypes (Figure 3 and 4).  As shown in Table 4, we observed age-related shifts in brain-behaviour associations 305 throughout childhood and adolescence. For both samples and all but one path, there was an 306 initially strong relationship between variables, then a dip around ages 7 -9 for CALM and age

390
On a more general level, this result highlights the existence of potential non-linear changes in 391 brain-behaviour mapping during childhood and adolescence and underlines the value of 392 modern statistical approaches, such as SEM Trees, for the study of age-related differences.

393
Nonetheless, it is worth noting that this exploratory result will need to be replicated in future 394 confirmatory studies with fine-grained data on puberty and larger sample sizes. The latter will 395 also allow for detailed investigations of potential gender differences. were of different sample sizes. It will therefore be necessary to replicate these findings in 406 large typical and atypical cohorts collected in the same setting. Another limitation of our 407 study is that our samples were cross-sectional, and not longitudinal. This means that were 408 able to investigate individual differences in g f , but we could not assess intra-individual 409 changes during childhood and adolescence. Although the relatively narrow age range makes  We included cognitive tasks measuring the domains of g f , working memory or processing 463 speed for CALM and NKI-RS. See Table 6 for the complete list of tasks used and the

Data Processing and Structural Equation Modelling
479 Covariance matrices and scripts replicating key analyses can be obtained from: 480 https://github.com/df1234/gf_development. We modelled raw scores for g f and working 481 memory tasks, as preregistered. Raw scores on processing speed tasks were transformed, 482 which was not preregistered. First, we inverted response time scores to obtain more intuitive 483 measures of speed for all but the CNB Motor Speed task, for which raw scores were already a 484 measure of speed. To satisfy SEM assumptions of normality, we additionally applied a log-transformation to reaction time tasks. For the CNB Motor Speed task only, we removed 486 values ± 2 SD of the mean (N = 6) because of the presence of outliers. 487 We modelled the associations between cognition and white matter microstructure using 488 Confirmatory Factor Analysis and SEM in R (R core team, 2015) using the package lavaan 489 (Rosseel, 2012). All models were fit using maximum likelihood estimation with robust Huber-