Ultrasonic bone age fractionates cognitive abilities in adolescence

Adolescent development is not only shaped by the mere passing of time and accumulating experience, but it also depends on pubertal timing and the cascade of maturational processes orchestrated by gonadal hormones. Although individual variability in puberty onset confounds adolescent studies, it has not been efficiently controlled for. Here we introduce ultrasonic bone age assessment to estimate biological maturity and disentangle the independent effects of chronological and biological age on adolescent cognitive abilities. Comparing cognitive performance of female participants with different skeletal maturity we uncover the impact of biological age on both IQ and specific abilities. We find that biological age has a selective effect on abilities: more mature individuals within the same age group have higher working memory capacity and processing speed, while those with higher chronological age have better verbal abilities, independently of their maturity. Based on our findings, bone age is a promising biomarker of adolescent maturity.


Supplementary Figure 1: The association between menarche age and pubertal maturity assessed by ultrasonic bone age.
The small table shows mean menarche age for four chronological age groups between 11 to 14 years of age. Mean menarche age increases with age as a larger proportion of girls becomes postmenstrual. The bar-graph show the percentage of participants who are already postmenstrual within the bone age determined maturational groups (see Figure 1.c of the main text), for each chronological age-group independently. There is an association between bone age determined maturity and the percentage of postmenstrual girls. More advanced maturity involves a higher percentage of postmenstrual participants. Notice that bone age defined maturity levels cannot be replaced by menarche age because this information is only available for postmenstrual girls, and there is no information on those who are not menstruating yet. However, the association presented in the figure provides support for bone age being a proper proxy for pubertal maturation.

Supplementary Table 2: Correlations for the four factors of WISC-IV.
Correlations between chronological age / biological age and WISC-IV broad abilities and overall performance. 95CI = 95% confidence interval

Supplementary Table 3: Partial correlations for the subtests of WISC-IV.
Partial correlations between chronological age / biological age and WISC-IV subtests. The correlations represent the independent effect of chronological age and biological age, with the effect of the other kind of age controlled for. 95CI = 95% confidence interval. VC = Verbal Comprehension, PR = Perceptual Reasoning, WM = Working Memory, PS = Processing Speed

Supplementary Table 4: Linear regressions for the subtests of WISC-IV.
Multiple linear regressions for chronological age / biological age on WISC-IV broad abilities and overall performance, B = unstandardised coefficients, beta = standardised coefficients, 95CI = 95% confidence interval.VC = Verbal Comprehension, PR = Perceptual Reasoning, WM = Working Memory, PS = Processing Speed.

Supplementary Table 5: Relative Weight Analysis results.
Predictors have a significant effect if the confidence interval of the tests of statistical significance does not contain zero. The effects of the predictors are significantly different if the confidence interval of the predictor comparison does not contain zero. The rescaled relative weights indicate that 78.4 % of the effect of these predictors can be ascribed to BA; 21.6% of the effect of these predictors can be ascribed to CA in the case of Working Memory. With respect to Verbal Comprehension, the rescaled relative weights indicate that 66.2 % of the effect of these predictors can be ascribed to CA; 33.8% of the effect of these predictors can be ascribed to BA. In the case of overall performance BA and CA have almost identical effects (51.5% vs 48.5%, respectively). The difference between relative weights was significant in the case of Working Memory.

Supplementary Table 6: Correlations within 1-year-wide age-groups.
Correlations between chronological age / biological age and WISC-IV broad abilities and overall performance within 1-year-wide age-groups corresponding to the two orthogonal dimensions of Figure 1.a. Those rows where the MRL column has BA, correlations are at fixed CA values. Those rows where the MRL column has CA, correlations are at fixed BA values. Note that we ran correlations only for those variable combinations where multiple linear regression (see Table 2. of the main text) was significant. Mean age in central bins was 12.5, 13.5, 14.5. Checkmarks in the last column indicate results still significant after Holm-Bonferroni correction. X in the last column indicates the result that was not significant after the correction.

Supplementary Table 7: ANOVA results for the one-year-wide age-groups.
Where correlations were significant after Holm-Bonferroni correction in Supplementary Table 6., we carried out a one-way ANOVA analysis on the two extreme bins (without the middle bin) of the two independent dimensions of Figure 1.a. This analysis provides a view within narrower age-ranges at significant differences between delayed and advanced or between younger and older participants (see the line-graphs in Figures 2. and 3. in the main text).

Supplementary Table 8: Distribution of parental education in the studied sample.
To see whether parental education, as an indicator of socio-economic status, might be a contributing factor in the differences of cognitive development of the current sample, we carried out statistical analyses. The non-parametric independent-samples median tests showed that medians were the same both across bins and across the acceleratedaverage-decelerated dimensions (see Fig. 1a and Fig.1b of the main text for the definition of bins and dimensions) regarding parental education level (alpha=0.05, CI=95). The non-parametric independent-samples Kruskal-Wallis test showed that distribution of parental education was the same across bins. Independent-samples Kruskal-Wallis test also showed that parental education has the same distribution in accelerated, average and decelerated participant groups (alpha=0.05 and CI=95). Therefore, parental education does not contribute to differences in cognitive development in the current sample.

WISC subtest selection
We purported to measure each broad ability with three tests. In the case of Verbal Comprehension and Perceptual Reasoning the ten core tests of the battery already include three tests of each. For Perceptual Speed, there are two core tests in the standard battery, and they can be supplemented with a third test, Coding, which we also administered. In the case of Working Memory there are two core tests in the standard battery and a third test, Arithmetic, is available as supplementary.
However, the status of the Arithmetic subtest is controversial. Upon examining the content of the test, it appears as a complex measure that taps on various domains at the same time: quantitative knowledge, quantitative reasoning, working memory, and even verbal comprehension. In fact, the Cattell-Horn-Carroll (CHC) model recognizes two different aspects of individual differences in math-related cognition: the separate broad ability factor 'Quantitative knowledge' reflects acquired quantitative or numerical knowledge, but not reasoning with such knowledge, while 'Quantitative reasoning', a narrow factor under the broad ability factor 'Fluid reasoning' represents reasoning with numerical material 1 .
For it to be a better measure of Working memory, the publishers of the WISC-IV have substantially modified the Arithmetic test from previous versions: the math-knowledge load was reduced, and the working memory demands were increased 2 . Despite this, several studies investigating the factor structure of the WISC-IV found that it is still not a pure measure of working memory; it has been found that besides memory, it measures fluid reasoning 3,4 and/or crystallized intelligence/comprehension & knowledge 5,6 , too. Even a study that confirmed Arithmetic as a measure of Working memory found that its factor loading is much smaller than of the core working memory subtests 7 .
A study fitted a model in which the Arithmetic test was removed from the 4-factor structure of the WISC and was directly measuring the higher order g factor 8 . Such a model is equivalent to one in which a separate factor is added, of which Arithmetic is the single indicator. Therefore, such a factor is statistically redundant, but substantively it is more plausible to claim that Arithmetic is a measure of a fifth broad ability than a direct measure of g. Indeed, it appears that the main difficulty from a latent variable modelling perspective is that Arithmetic might be the single indicator of a fifth broad ability, only allowing for suboptimal models.
The manual did in fact consider such a 5-factor solution but discarded it because it did not improve model fit over the 4-factor solution. Yet, importantly from our perspective, the 5-factor solution with Arithmetic as the only indicator of the fifth factor did improve model fit in a particular age group: in 11-16-year-old children 2 . Since this is the exact age range we were targeting, after considering all the above evidence we decided against administering the Arithmetic test as a supplementary test of Working Memory.