Cortical thickness, gray matter volume, and cognitive performance: a crosssectional study of the moderating effects of age on their interrelationships

Abstract

In a sample comprising younger, middle-aged, and older cognitively healthy adults (N = 375), we examined associations between mean cortical thickness, gray matter volume (GMV), and performance in 4 cognitive domains—memory, speed, fluency, and crystallized intelligence. In almost all cases, the associations were moderated significantly by age, with the strongest associations in the older age group. An exception to this pattern was identified in a younger adult subgroup aged <23 years when a negative association between cognitive performance and cortical thickness was identified. Other than for speed, all associations between structural metrics and performance in specific cognitive domains were fully mediated by mean cognitive ability. Cortical thickness and GMV explained unique fractions of the variance in mean cognitive ability, speed, and fluency. In no case, however, did the amount of variance jointly explained by the 2 metrics exceed 7% of the total variance. These findings suggest that cortical thickness and GMV are distinct correlates of domain-general cognitive ability, that the strength and, for cortical thickness, the direction of these associations are moderated by age, and that these structural metrics offer only limited insights into the determinants of individual differences in cognitive performance across the adult lifespan.

Introduction

The effects of age on brain structure, and their role in mediating age-related cognitive decline, have long been a topic of interest. With the development of semiautomated pipelines for the analysis of structural MRI data (Fischl and Dale 2000), much of this interest has focused on the structural metric of neocortical (henceforth, cortical) thickness. Consequentially, a large literature has developed, documenting the trajectory of cortical thickness across the lifespan. As documented in a report that aggregated data across >100 crosssectional and longitudinal datasets, cortical thickness peaks in late infancy and declines monotonically with increasing age thereafter (Bethlehem et al. 2022).

In contrast to the lifelong negative association between cortical thickness and age evident after early childhood, the association between cortical thickness and cognitive ability across the lifespan has been reported to be nonlinear. Convergent findings from crosssectional and longitudinal studies indicate that cortical thickness is negatively correlated with cognitive ability in adolescence and young adulthood (Tamnes et al. 2010; Schnack et al. 2015; de Chastelaine et al. 2019; Krogsrud et al. 2021) but demonstrates a positive correlation with cognitive ability in later life (e.g. Karama et al. 2014; Salthouse et al. 2015; Sun et al. 2016; de Chastelaine et al. 2019). The findings for young individuals have been interpreted as a reflection of the beneficial effects of relatively extensive synaptic “pruning” during cortical development and maturation (Schnack et al. 2015; Tadayon et al. 2020). The findings for older adults are consistent with the intuition that “more is better” and the widely held view that degradation of brain structure is a mediator of age-related cognitive decline (a corollary of the “brain maintenance” hypothesis, Cabeza et al. 2018). The notion of cortical thinning as a mediator of individual differences in age-related cognitive decline has, however, been called into question by longitudinal evidence that childhood intelligence accounts for most of the association between intelligence and cortical thickness found some 60 years later, suggesting a lifelong association between the 2 variables (Karama et al. 2014).

The studies discussed above employed relatively undifferentiated metrics of thickness, such as the mean thickness of the entire cortical mantle. Numerous studies have examined associations at a more granular level, reporting correlations between the thickness of specific cortical regions and performance in different cognitive domains (e.g. Burzynska et al. 2012; Vonk et al. 2019). However, the reliability of such correlations has been called into question (Masouleh et al. 2022). Moreover, many studies reporting regionally specific associations between thickness and cognitive performance did not account for the strong correlations that exist between regional thickness measures. When this interdependency is accounted for, little evidence of regionally specific associations between cortical thickness and cognitive performance remains (Salthouse et al. 2015; Kranz et al. 2018; Hou et al. 2021; Krogsrud et al. 2021; but, see Lee et al. 2016). Analogously, associations between thickness and specific cognitive abilities are also largely eliminated after the variance shared across different cognitive measures is controlled for (Salthouse et al. 2015; Tsapanou et al. 2019; Hou et al. 2021; but, see Krogsrud et al. 2021). Thus, findings indicative of selective associations between regional cortical thickness and domain-specific cognitive performance seem largely to reflect a single association between global cortical thickness and mean cognitive ability.

The employment of cortical thickness as a structural brain metric is a relatively recent development. A variety of measures related to brain size or volume have a longer, and somewhat independent, history as predictors of cognitive ability, mainly the prediction of general intelligence. Volumetric measures decline monotonically with increasing age from midchildhood (Bethlehem et al. 2022), and they have consistently been reported to correlate with cognitive ability seemingly independently of sex or age (for reviews and meta-analyses, see Pietschnig et al. 2015, 2022). Findings from 1 recent large-scale study (Nave et al. 2019) indicated that the principal driver of the relationship between brain volume and intelligence is gray matter volume (GMV). As with cortical thickness, associations between GMV and cognition at the regional level are weak, if reliable at all (Cox et al. 2019; Masouleh et al. 2019), and largely reduce to an association captured by whole-brain measures.

Here, we took advantage of structural MRI and neuropsychological test data obtained from 375 participants, falling mainly into the younger (18–30 years) and older adult (ca. 65–75 years) age ranges along with a smaller middle-aged group (45–55 years), to examine the interrelationships between cortical thickness, GMV, and cognitive performance. As noted above, there is little evidence that measures of cortical thickness or GMV demonstrate reliable associations with cognitive ability at the regional level. Accordingly, we focus here on whole-brain metrics of these variables. We address 4 principal questions. First, with the additional power afforded by a larger sample, do we replicate our prior findings (de Chastelaine et al. 2019) that the relationship between cortical thickness and cognitive ability is moderated by age such that a negative correlation is found in young adulthood, while a positive relationship is evident in later life? Second, can we replicate prior findings (Salthouse et al. 2015; Hou et al. 2021) that cortical thickness is associated with general rather than domain-specific aspects of cognitive ability, and do these findings extend to GMV? Third, is the relationship between GMV and cognitive ability moderated by age? Finally, do cortical thickness and GMV explain unique fractions of variance in cognitive performance, that is, are they distinct or redundant correlates of cognitive ability? To our knowledge, this last question has rarely been addressed (but, see Ritchie et al. 2015; Hedden et al. 2016).

Materials and methods

Participants

A total of 375 cognitively healthy adults contributed to the analyses presented below. The participants were recruited from UT Dallas and its surrounding communities and comprised 195 younger (18–30 years) adults, 145 older (63–76 years) adults, and 35 middle-aged (43–55 years) individuals. Demographic details (along with neuropsychological test scores—see below) are summarized in Table 1. Data from an additional 18 participants who met the inclusion criteria described below (5 younger, 3 middle-aged, and 10 older) were rejected because of the low quality of their structural brain images.

Table 1.

Demographic information and neuropsychological test scores (sum or mean [SD]) in young, middle-aged, and older adults.

	Young	Middle	Older
N	195	35	145
Sex (M/F)	97/98	15/20	67/78
Age at neuropsychological test (years)	22.79 (3.13)	49.60 (3.45)	68.50 (3.44)
Years of education	15.71 (2.10)	16.49 (2.48)	16.84 (2.27)
MMSEa	29.30 (0.87)	29.34 (0.77)	29.21 (0.88)
CVLT short delay free recallb	12.89 (2.26)	11.57 (2.38)	10.69 (3.02)
CVLT long delay free recallb	13.33 (2.34)	12.29 (2.43)	11.27 (2.92)
CVLT short delay cued recallb	13.44 (2.17)	13.14 (1.99)	12.07 (2.60)
CVLT long delay cued recallb	13.73 (2.14)	12.97 (1.95)	12.18 (2.49)
CVLT recall composite	53.39 (8.39)	49.97 (8.14)	46.13 (10.42)
CVLT recognition hitsc	15.36 (1.02)	15.29 (0.83)	14.83 (1.34)
CVLT recognition FAsd	0.83 (1.15)	1.91 (2.16)	2.25 (2.34)
WMS logical memory Ie	30.67 (6.39)	29.00 (6.25)	28.53 (4.88)
WMS logical memory IIe	28.15 (6.60)	25.49 (5.90)	25.63 (5.78)
WMS composite	58.82 (12.58)	54.49 (11.76)	54.17 (10.17)
F-A-Sf	45.62 (11.47)	47.11 (11.72)	45.30 (11.71)
SDMTg	63.14 (11.26)	55.17 (7.60)	48.98 (8.40)
Trails Ah	21.20 (7.29)	23.97 (6.29)	31.03 (10.85)
Trails Bh	46.55 (18.30)	51.89 (16.89)	68.90 (26.66)
Forward/backward digit spani	19.11 (4.22)	17.97 (3.48)	18.16 (3.87)
Category fluency testj	24.52 (5.33)	23.60 (5.92)	21.76 (5.19)
TOPF/WTARk	108.60 (5.72)	111.34 (4.95)	111.05 (5.70)
Raven’s progressive matricesl	11.08 (1.02)	10.20 (1.68)	9.33 (2.12)

^aNumber of correct items out of 30.

^bNumber of recalled items out of 16.

^cNumber of recognized items out of 16.

^dNumber of incorrect items out of 32.

^eNumber of recalled items out of 50.

^fTotal number of correctly generated words.

^gNumber of correct items out of 110.

^hSeconds.

ⁱNumber of correct items out of 30.

^jTotal number of correctly generated words.

^kScaled score referenced to age 43.

^lNumber of correct items out of 12.

The participants were recruited into ≥1 of the fMRI studies conducted in our laboratory between 2011 and 2019. Participants were right-handed, had normal or corrected-to-normal vision, and were fluent English speakers before the age of 5. Additionally, participants had no history of neurological or psychiatric disease, substance abuse, diabetes, or current or recent use of prescription medication affecting the central nervous system. The participants were accepted into a study according to a common set of inclusion and exclusion criteria that were intended to minimize the likelihood of including individuals with cognitive impairment arising from neuropathology, as determined by our neuropsychological test battery (see Neuropsychological test battery section below). Participants gave informed consent according to the procedures approved by the UT Dallas and UT Southwestern Institutional Review Boards and were compensated at the rate of $30 per hour.

The same MRI scanner and closely similar acquisition sequences were employed to acquire a T1-weighted structural MRI scan from each participant. For those participants who contributed to multiple studies, we employed the MRI scan that was acquired closest to the time they first undertook the neuropsychological test battery. The mean time between the test and scan sessions was 3.1 weeks (range: −41 to +43 weeks) and did not differ reliably according to age group (means of 3.2, 2.9, and 3.1 weeks for the young, middle-aged, and older groups, respectively, P > 0.9). We could find no evidence that test-scan interval mediated or moderated any of the associations between the structural and cognitive variables described below. The same held true for the factor of study cohort.

Structural brain data in the form of cortical thickness estimates have been reported previously for a subset of the present sample (36 of the younger participants, the 35 middle-aged participants included here, and 62 of the older participants; see de Chastelaine et al. 2019 and Hou et al. 2021). Neither the analyses of cortical thickness estimates from the full sample nor the analyses of GMV have been reported previously.

Neuropsychological test battery

With the exception of 1 younger adult, participants completed the neuropsychological test battery on a separate day from the MRI session. The battery comprised the Mini-Mental State Examination (MMSE), the California Verbal Learning Test-II (CVLT; Delis et al. 2000), Wechsler Logical Memory (Tests 1 and 2; Wechsler 2009), the Symbol Digit Modalities Test (SDMT; Smith 1982), Trail Making (Tests A and B; Reitan and Wolfson 1985), the F-A-S subtest of the Neurosensory Center Comprehensive Evaluation for Aphasia (Spreen and Benton 1977), the Wechsler Adult Intelligence Scale–Revised (Forward and Backward digit span subtests; Wechsler 1981), Category Fluency test (Benton 1968), and Raven’s Progressive Matrices (List 1; Raven et al. 2000). In addition, participants completed either the Wechsler Test of Adult Reading (WTAR; Wechsler 2001) or its revised version, the Wechsler Test of Premorbid Functioning (TOPF; Wechsler 2011); 348 participants completed the WTAR, and 27 completed the TOPF. The TOPF and WTAR scores were converted to commensurate measures by scaling the raw scores according to the 2 tests’ respective age-scaled norms for age 43 years (the mean age of our participants). Scores on each of the neuropsychological tests are summarized in Table 1 for each age group. To minimize the likelihood of including participants with mild cognitive impairment or early dementia, potential participants were excluded if they performed >1.5 SD below age norms on ≥2 nonmemory tests, >1.5 SD below the age norm on at least 1 memory test, or if their MMSE score was <26. Because the 4 CVLT recall scores were highly correlated across participants (min r = 0.81), they were summed to give a single composite CVLT recall score prior to further analysis. For the same reason, the 2 WMS Logical Memory Scores (r = 0.86) were summed to give a single composite score. By contrast, the correlation between the 2 scores (A and B) comprising the Trails test was markedly lower (0.51); hence, these scores were not averaged prior to entry in the PCA (see next section).

Principal components analysis

The composite CVLT and Logical Memory recall scores, along with all other test scores apart from the MMSE, were subjected to principal components analysis (PCA) to identify the constructs underpinning performance on the test battery (cf. de Chastelaine et al. 2019; Hou et al. 2021). The MMSE was excluded from the PCA because of score compression, with >80% of the sample scoring at (30) or near (29) ceiling. The PCA was performed as follows: First, the test scores for each age group were z-transformed within-group (to obtain a solution that reflects relative test performance independently of age group). Second, the resulting z-scores were combined into a single matrix and subjected to a PCA as implemented in IBM SPSS Statistics v.28. Components were retained if their eigenvalues exceeded (or, for 1 component, closely approached) 1 (see Table 2). Third, the retained components were subjected to varimax rotation to simplify the solution space. In a final step, we computed participant-specific component scores to reflect the performance of a given participant on each of the retained principal components. To do this, we z-scored the raw test scores across age groups and, for each of the retained components, multiplied the z-scored data by the component loading and summed the resulting scores.

Table 2.

Varimax-rotated factor loadings from the PCA of the neuropsychological test battery.

	Memory	Speed	Fluency	Crystallized intelligence
CVLT recall composite	0.85	0.11	0.24	0.01
CVLT recognition hits	0.70	0.01	−0.10	0.33
CVLT recognition FAs	−0.65	−0.11	−0.24	0.28
WMS composite	0.63	0.16	0.05	0.07
F-A-S	−0.03	0.17	0.61	0.33
SDMT	0.18	0.72	0.28	0.06
Trail A	−0.07	−0.75	−0.14	0.01
Trail B	−0.13	−0.73	−0.04	−0.28
Forward/backward digit span	0.06	0.22	0.07	0.75
Category fluency	0.11	0.21	0.76	−0.08
TOPF/WTAR	0.07	0.01	0.46	0.64
Raven’s	0.19	0.07	0.53	0.13
Eigenvalues (% variance)	3.37 (28.08%)	1.56 (13.02%)	1.08 (9.01%)	0.97 (8.10%)

MRI acquisition

A Philips Achieva 3T MR scanner (Philips Medical System, Andover, MA, United States) equipped with a 32-channel head coil was used to acquire the anatomical images. T1-weighted anatomical images were acquired with a MPRAGE pulse sequence with the following range of parameters (due to incremental software upgrades, there were small differences in TR, TE, and slice number across experiments): TR = 8.05–8.38 ms, TE = 3.68–3.90 ms, reconstructed FOV = 256 × 256, voxel size = 1 × 1 × 1 mm, 160/176 slices, sagittal acquisition.

Estimation of cortical thickness and brain volume

Cortical thickness and the different brain volume estimates were derived from each participant’s T1-weighted image using Freesurfer’s (v6.0) semiautomatic processing pipeline (http://surfer.nmr.mgh.harvard.edu/fswiki; Dale et al. 1999; Fischl and Dale 2000; Fischl et al. 2002). After skull-stripping and intensity normalization, each 3D T1 volume was reconstructed as an inflated 2D surface so that white matter and pial surfaces could be identified for each hemisphere. Using these surface estimates, cortical thickness could then be estimated.

After the completion of this first step, white matter and pial surfaces were visually inspected by 1 of 6 trained raters who were blind to the participants’ age. Although the automated analysis generally succeeded in identifying white and gray matter boundaries, some brain regions (e.g. orbito-frontal, insula, and temporal regions, and boundaries between gray matter and the pial surface and the cerebellum) frequently required further manual edits. Tissue classification errors were corrected by editing the white matter and brain masks returned by the first pass through the Freesurfer pipeline. In addition, “control points” were added to voxels on, or close to, what were determined to be white matter paths that had been excluded by the automated analysis. Manual edits were repeated as necessary. Once completed, each edited scan was checked by a second rater for quality control purposes. Reliability of mean cortical thickness estimates across the 6 raters was assessed for 5 randomly selected brain images (2 younger, 2 older, and 1 middle-aged participant). The reliability coefficient (ICC[2]; Shrout and Fleiss 1979) was 0.88.

Cortical thickness was measured as the distance from the gray/white matter boundary to the pial surface on a vertex-by-vertex basis across the entire cortical mantle of each hemisphere. Mean whole-brain cortical thickness was calculated by averaging cortical thickness values generated for the 2 hemispheres. GMV was indexed by the “Total Gray” statistic returned by the Freesurfer package. This statistic is an aggregate of surface-based volume estimates (derived from the parcellated cortical ribbon including the hippocampus) and gray matter voxel counts from the segmented subcortical structures. We also extracted estimates of cortical surface area and cortical, total brain, and white matter volumes (WMV). As reported in the Supplemental Materials, findings for cortical and total brain volumes (see Tables S4 and S6) were somewhat weaker than, but very similar to, those for GMV, while the findings for surface area and WMV (see Tables S5 and S7) were weaker still.

Estimation of within-scan head motion

It has been reported that within-scan head motion results in the underestimation of Freesurfer derived cortical thickness and GMV values (Reuter et al. 2015). It has also been reported previously that head motion estimates obtained from a functional scan acquired proximal to a T1-weighted structural scan can be employed as a proxy for within-scan motion occurring during the structural scan (Savalia et al. 2017; de Chastelaine et al. 2019). In light of these prior reports, here, we obtained head motion estimates in the form of frame displacement (FD) values (Power et al. 2012; de Chastelaine et al. 2019) from the functional time series acquired closest in time to each participant’s structural scan (functional data, and hence FD values, were unavailable for 1 younger and 1 older participant).

Statistical analyses

Sex and age group differences in cognitive and neural metrics were evaluated with 3 (age group) by 2 (sex) factorial ANOVAs. Significant age group effects were followed up as necessary with post hoc t-tests. Relationships between brain metrics and performance in each of the 4 cognitive domains identified by PCA of the neuropsychological test scores (see below) were evaluated through a 2-step hierarchical regression approach. The first step included age group, sex, and years of education as predictor variables. The second step added the brain metric of interest (either thickness or GMV) and its interaction with age group as additional predictors. A significant increase in the variance explained by the model after the inclusion of these predictors indicated that the brain metric accounted for unique variance in performance in the cognitive domain in question. In cases where the interaction was nonsignificant, the full model was rerun after dropping the term. If the interaction term was significant, group-wise partial correlations between the brain and performance metrics, controlling for sex, years of education and chronological age, were computed. The same approach was taken for the “mean cognitive ability” metric formed by summing each participant’s 4 component scores (see below).

The Bonferroni procedure was employed to correct the nominal significance level (P < 0.05) of the above-described regression and correlational analyses for multiple comparisons. The correction was applied separately for each brain metric (that is, the 2 metrics were treated as separate test families). For any follow-up group-wise correlations, the correction was also applied separately for each age group. Because the 5 cognitive performance scores were highly correlated (rs between 0.378 and 0.905), we employed the “M_eff correction” (Nyholt 2004; Derringer 2018) to estimate the Bonferroni divisor. This yielded a corrected significance level of P < 0.017.

To examine the specificity of any relationship between a brain metric and a given component score, we reran the hierarchical regression analyses that identified the relationship after including the other 3 component scores as additional step 1 predictor variables (cf. Salthouse et al. 2015). The relationship was deemed specific if the addition of the brain metrics in step 2 of the analyses still led to a significant increase in explained variance.

Results

Neuropsychological test scores

Performance on the neuropsychological test battery is reported in Table 1.

Principal components analysis

The varimax-rotated PCA solution returned 4 components that, between them, accounted for 58.21% of the variance in the neuropsychological test scores. The loadings of each test on the 4 rotated components, as well as their eigenvalues and proportion of variance explained, are reported in Table 2. It can be seen from the table that the components roughly correspond to the constructs of “memory,” “speed,” “fluency,” and “crystallized intelligence” (a closely similar pattern of loadings was reported by de Chastelaine et al. 2019, following PCA of a subset of the present data). As in Hou et al. (2021), we generated a mean cognitive ability score for each participant by summing the 4 component scores. We also estimated general cognitive ability by using the loadings associated with the first unrotated principal component extracted from the PCA of the neuropsychological test data (cf. Nave et al. 2019). Controlling for sex, age group, and years of education, the 2 estimates of overall ability were almost perfectly correlated (r = 0.997). Hence, below, we only report the findings for the first of these metrics. We interpret it as a rough proxy for general intelligence.

Outlier identification

We conservatively defined outlying data points as +/− >3.5 SDs from the mean for each age group. This criterion identified 1 younger participant with excessively low memory, fluency, and mean cognitive ability scores, and 1 older participant with an excessively low speed component score. No other behavioral or brain measures met the criterion. The outlying scores were dropped from all analyses reported below. Thus, analyses were conducted on the full sample (n = 375) in the case of the crystallized intelligence component scores, and on a sample size of n = 374 for the remaining 3 component scores and mean cognitive ability.

Cognitive component scores

Component scores are summarized by age group and sex in Table 3. Also summarized in the table are the mean cognitive ability scores (see above). Fixed-effects ANOVAs were employed to examine the effects of age group (younger, middle-aged, and older) and sex on each component. The outcomes of the ANOVAs are summarized in Table 4 along with the results of post hoc pairwise contrasts when these were indicated. As is evident from the table, memory scores varied robustly and additively according to the factors of sex and age group such that females outperformed males, and there was a graded effect of age group, with younger adults demonstrating the highest scores and older adults the lowest. By contrast, for the speed and fluency contrasts, only a graded age group effect was evident. For crystallized intelligence, the age group factor was nonsignificant, but a small sex effect was evident, with males demonstrating the higher scores. Finally, as in the case of speed and fluency, mean cognitive ability varied with age group but not sex. In light of the sex effects that were evident for 2 of the 5 cognitive constructs, we included sex as an independent variable in all of the regression and correlational analyses described below.

Table 3.

Component and mean cognitive ability scores (mean (SD)) stratified by age group and sex.

	Young	Middle	Older
	M F	M F	M F
Memory	0.44 (2.10) 1.55 (1.66)	−0.97 (1.54) 0.37 (1.75)	−2.00 (2.26) −0.69 (2.56)
Speed	1.13 (1.83) 1.20 (1.51)	−0.05 (1.34) 0.12 (1.64)	−1.71 (1.82) −1.47 (2.39)
Fluency	0.71 (1.95) 0.61 (1.52)	0.05 (1.72) 0.12 (1.82)	−1.07 (1.85) −0.77 (2.21)
Cryst intell	0.23 (1.61) −0.01 (1.30)	0.60 (1.34) −0.10 (1.36)	−0.01 (1.53) −0.36 (1.66)
Mean cog	2.75 (6.17) 3.36 (4.88)	−0.37 (4.21) 0.50 (5.71)	−4.79 (5.94) −3.29 (7.68)

M, male; F, female; Mean cog, mean cognitive ability; Cryst intell, crystallized intelligence.

Table 4.

Results of fixed-effects ANOVAs examining the effects of age group (young, middle-aged, and older) and sex on each cognitive component score along with the outcomes of post hoc pairwise contrasts (bold rows denote significance at P < 0.05).

Memory
Age group	F(2,368) = 55.27, P < 0.001, partial η  2 = 0.23
Sex	F(1,368) = 19.28, P < 0.001, partial η  2 = 0.05
Age group × sex	F(2,368) = 0.21, P = 0.811, partial η2 = 0.00
Independent samples’ t-tests
Young versus middle	t(227) = 3.61, P < 0.0001, Cohen’s d = 0.66
Young versus older	t(258) = 9.41, P < 0.001, Cohen’s d = 1.07
Middle versus older	t(71) = 3.00, P < 0.005, Cohen’s d = 0.46
Speed
Age group	F(2,368) = 91.00, P < 0.001, partial η  2 = 0.33
Sex	F(1,368) = 0.23, P = 0.630, partial η2 = 0.00
Age group × sex	F(2,368) = 0.01, P = 0.986, partial η2 = 0.00
Independent samples’ t-tests
Young versus middle	t(228) = 3.71, P < 0.001, Cohen’s d = 0.68
Young versus older	t(271) = 12.95, P < 0.001, Cohen’s d = 1.47
Middle versus older	t(177) = 4.27, P < 0.001, Cohen’s d = 0.81
Fluency
Age group	F(2,368) = 31.61, P < 0.001, partial η  2 = 0.15
Sex	F(1,368) = 0.09, P = 0.771, partial η2 = 0.00
Age group × sex	F(2,368) = 0.67, P = 0.513, partial η2 = 0.00
Independent samples’ t-tests
Young versus middle	t(227) = 1.94, P = 0.054, Cohen’s d = 0.36
Young versus older	t(274) = 7.66, P < 0.001, Cohen’s d = 0.87
Middle versus older	t(178) = 2.65, P = 0.009, Cohen’s d = 0.50
Crystallized intelligence
Age group	F(2,369) = 2.08, P = 0.127, partial η2 = 0.01
Sex	F(1,369) = 4.48, P = 0.035, partial η  2 = 0.01
Age group × sex	F(2,369) = 0.34, P = 0.710, partial η2 = 0.00
Mean cognitive ability
Age group	F(2,368) = 56.00, P < 0.001, partial η  2 = 0.23
Sex	F(1,368) = 1.47, P = 0.227, partial η2 = 0.00
Age group × sex	F(2,368) = 0.22, P = 0.800, partial η2 = 0.00
Independent samples’ t-tests
Young versus middle	t(227) = 2.91, P < 0.005, Cohen’s d = 0.53
Young versus older	t(337) = 10.36, P < 0.001, Cohen’s d = 1.14
Middle versus older	t(178) = 3.29, P < 0.001, Cohen’s d = 0.62

Years of education

ANOVA (factors of age group and sex) gave rise solely to a significant effect of age group (F(2, 369) = 12.84, P < 0.001, partial η² = 0.07). Post hoc tests revealed that years of education were significantly greater for the older than the younger group, with the middle-aged adults differing significantly from neither of the other groups. These findings motivated the inclusion of years of education as a control variable in the across-group regression and correlational analyses. However, all findings reported below were unaltered when this variable was omitted from the analyses.

Cortical thickness

Whole-brain cortical thickness estimates are summarized in Supplementary Table S1A and illustrated in Fig. 1A. ANOVA of these data (factors of age group and sex) gave rise to a significant effect of age group (F(2,369) = 198.20, P < 0.001, partial η² = 0.52), but there was no evidence of a sex effect (P = 0.402) or a sex × age group interaction (P = 0.167). Post hoc contrasts revealed that thickness was graded across the age groups such that the estimates for younger adults exceeded those for the middle-aged group (t(228) = 6.77, P < 0.001, Cohen’s d = 1.24), while, in turn, the middle-aged group demonstrated larger estimates than the older age group (t(178) = 4.98, P < 0.001, Cohen’s d = 0.94).

Fig. 1.

Distributions of estimates of (A) cortical thickness and (B) GMV, stratified by sex and age group. Black dots and bars represent means +/−95% confidence intervals.

Gray matter volume

Estimates for GMV are summarized in Supplementary Table S1B and illustrated in Fig. 1B. ANOVA gave rise to significant effects of age group (F(2, 369) = 130.98, P < 0.001, partial η² = 0.42) and sex (F(1,369) = 75.46, P < 0.001, partial η² = 0.17). The interaction between the factors was not significant (P = 0.076). As for cortical thickness, post hoc contrasts indicated that GMV was lower in the middle-aged than the younger group (t(228) = 4.15, P < 0.001, Cohen’s d = 0.76) and lower in the older than the middle-aged group (t(178) = 4.24, P < 0.001, Cohen’s d = 0.80). As is evident from Fig. 1B, the main effect of sex was driven by greater GMV in males.

Relationships between chronological age, cognitive performance, and structural metrics

We used partial correlations (controlling for sex and education) to examine relationships between chronological age and cognitive and brain structural metrics separately for each age group. The results of these analyses are summarized in Table 5, where it is evident that in the younger and older age groups cortical thickness was significantly and negatively correlated with age (albeit more strongly in the older group; difference between correlation coefficients P < 0.05). By contrast, a reliable negative correlation between age and GMV was present in the older group only (while nonsignificant, a modest negative correlation was also evident in the middle-aged group). Turning to the cognitive component scores, no score demonstrated a reliable correlation with age in the younger age group. By contrast, except for crystallized intelligence, all cognitive scores correlated reliably with age in the older group. Only in the case of the speed component, however, did the corresponding correlations differ significantly between the younger and older age groups (P < 0.05). Again, although they were nonsignificant, the pattern of correlations in the middle-aged group roughly mirrored that in the older group.

Table 5.

Correlations between chronological age, cognitive performance, and structural metrics for each age group, after controlling for sex and years of education (P values in parentheses - bold values denote significance at the P < 0.05 level).

	Chronological age
	Young	Middle	Older
Mean cortical thickness	−0.20 (0.007)	−0.13 (0.468)	−0.41 (<0.001)
Gray matter volume	−0.13 (0.082)	−0.20 (0.264)	−0.36 (<0.001)
Memory	−0.04 (0.594)	−0.19 (0.296)	−0.22 (0.009)
Speed	−0.12 (0.111)	−0.16 (0.359)	−0.34 (<0.001)
Fluency	−0.07 (0.306)	−0.16 (0.377)	−0.20 (0.017)
Crystallized intelligence	−0.03 (0.634)	0.05 (0.783)	−0.06 (0.458)
Mean cognitive ability	−0.08 (0.292)	−0.15 (0.391)	−0.25 (0.002)

Relationships between cortical thickness and cognitive performance

As was discussed in the Introduction, it has previously been reported that the relationship between performance in a specific cognitive domain and cortical thickness is substantially attenuated when performance in other domains is controlled for. This was the case here: When we ran the hierarchical regression analyses predicting performance in each cognitive domain after including the other component scores as additional predictors, cortical thickness and its interaction with age group failed to explain variance in any of the scores (all Ps > 0.336; see Supplementary Table S2 for the outcomes of the regression models predicting each of the component scores). The outcome of the hierarchical regression model predicting mean cognitive ability is summarized in Table 6. As is evident from the table, sex, age group, and years of education were each significant predictors of mean cognitive ability. Crucially, the inclusion of the thickness and thickness by age interaction terms in the model led to a significant increase in explained variance (P < 0.001). In light of the significant age group by thickness interaction term, we computed partial correlations between thickness and mean cognitive ability separately for each age group (in each case controlling for sex, education, and chronological age). Group-wise partial correlations (Table 7A) revealed that the correlations only reached significance in the older age group. The scatter plot illustrating the correlation in the older and, for comparison, the younger age group is illustrated in Fig. 2A.

Table 6.

Hierarchical linear regression results for mean cortical thickness predicting mean cognitive ability (P values in parentheses- bold values denote significance at the P < 0.05 level).

Mean cortical thickness predicting mean cognitive ability
Parameter	β	t (P)	F (P)	ΔF (P)	R  2	ΔR2
Step 1			41.19 (<0.001)		0.25
Intercept		−2.74 (0.006)
Age group	−0.51	−11.05 (<0.001)
Sex	−0.09	−1.99 (0.048)
Education	0.13	2.81 (0.005)
Step 2			29.74 (<0.001)	9.66 (<0.001)	0.29	0.04
Intercept		−2.91 (0.004)
Age group	−4.19	−3.87 (<0.001)
Sex	−0.07	−1.61 (0.109)
Education	0.13	2.74 (0.006)
Mean cortical thickness	0.16	2.47 (0.014)
Mean cortical thickness × age group	3.79	3.51 (<0.001)

Table 7.

Correlations between mean cognitive ability and (A) cortical thickness and (B) GMV for each age group after controlling for chronological age, sex, and years of education (P values in parentheses- bold values denote significance at the P < 0.05 level).

	A. Mean cortical thickness			B. GMV (adjusted for sex)
	Young	Middle	Older	Young	Middle	Older
Mean cognitive ability	−0.10 (0.183)	0.14 (0.457)	0.23 (0.006)	0.06 (0.410)	−0.02 (0.907)	0.22 (0.008)

Fig. 2.

Scatter plots of the relationships between: A) cortical thickness and mean cognitive ability; B) GMV and mean cognitive ability; and C) GMV and speed in younger and older age groups, controlling for chronological age, years of education, and sex.

Relationships between GMV and cognitive performance

As was reported above, GMV differed robustly according to sex. Accordingly, we adjusted the GMV estimates for sex prior to submitting them to analyses. In so doing, we controlled for the variance in GMV that was directly attributable to sex, while allowing variance attributable to other factors to remain. If this residual variance is shared with variance in cognitive performance, then the sex adjusted GMV metric will be identified as a significant predictor of performance (see Lee et al. 2019 for a closely analogous approach to controlling for sex differences in volumetric data). Consistent with prior studies examining relationships between brain volumetric measures and cognitive ability (e.g. Cox et al. 2019; Lee et al. 2019; Nave et al. 2019), we did not control for individual differences in intracranial volume.

As with cortical thickness, associations between GMV and the component scores for memory, fluency, and crystallized intelligence were nonsignificant after controlling for the remaining component scores (Ps > 0.089; see Supplementary Table S3 for the outcomes of the regression models predicting the memory, fluency, and crystallized intelligence scores). By contrast, after controlling for the other scores, GMV continued to explain a significant (P = 0.001) fraction of the variance in the speed component scores. The outcomes of the hierarchical regression analyses predicting mean cognitive ability and speed from GMV and its interaction with age are summarized in Table 8, where it can be seen that, in both cases, the interaction terms were significant. Group-wise partial correlations (Table 7B) revealed that the correlations between mean cognitive ability were largest and only reached significance in the older age group. The same held true for the correlations with speed (r = 0.12 [P = 0.096], r = 0.11 [P = 0.557], and r = 0.26 [P = 0.002] for the younger, middle-aged, and older groups, respectively). The corresponding scatter plots are illustrated in Fig. 2B and C.

Table 8.

Hierarchical linear regression results for GMV (adjusted by sex) predicting mean cognitive ability and speed (P values in parentheses- bold values denote significance at the P < 0.05 level, results of step 1 models are listed in Table 6).

GMV predicting mean cognitive ability
Parameter	β	t (P)	F (P)	ΔF (P)	R  2	ΔR2
Step 2			28.91 (<0.001)	8.11 (<0.001)	0.28	0.03
Intercept		−2.65 (0.008)
Age group	−0.36	−6.08 (<0.001)
Sex	−0.08	−1.70 (0.090)
Education	0.14	3.05 (0.002)
GMV	0.22	3.65 (<0.001)
GMV × age group	0.12	2.54 (0.012)
GMV predicting speed
Parameter	β	t (P)	F (P)	ΔF (P)	R  2	ΔR2
Step 2			46.21 (<0.001)	13.37 (<0.001)	0.39	0.04
Intercept		−2.56 (0.011)
Age group	−0.42	−7.58 (<0.001)
Sex	−0.02	−0.55 (0.585)
Education	0.12	2.80 (0.005)
GMV	0.27	4.92 (<0.001)
GMV × age group	0.12	2.73 (0.007)

Cortical thickness and GMV as copredictors of cognitive performance

Thickness and GMV were moderately correlated (controlling for sex, years of education, and age group, r = 0.36, P < 0.001). Therefore, we examined whether the 2 metrics explained independent fractions of variance in the 3 cognitive scores—speed, fluency, and mean cognitive ability—with which they were independently associated (see above and Supplementary Tables S2 and S3). We constructed regression models in which cortical thickness, GMV, and their interactions with age group were employed to predict each component score (along with the control variables of age group, sex, and years of education). In the initial iteration of these models, the GMV by age interaction term did not approach significance and was therefore dropped. The GMV and the thickness × age group interaction terms were significant in the models predicting speed (both Ps < 0.001), fluency (P = 0.014 and P = 0.022 for the GMV and thickness × age terms, respectively) and mean cognitive ability (P = 0.021 and P < 0.001 for GMV and thickness × age group, respectively). Thus, the 2 structural metrics accounted for a unique fraction of the variance in all 3 of the cognitive scores that were examined.

Effects of head motion

Consistent with prior findings (Savalia et al. 2017; de Chastelaine et al. 2019), estimates of head motion (FD) demonstrated an age-invariant correlation with thickness estimates (r = −0.283, P < 0.001). Head motion also correlated in an age-invariant manner with estimates of GMV (r = −0.177, P < 0.001). Therefore, all of the analyses reported above were repeated after including FD (and, where appropriate, its interaction with age group) as an additional covariate. In no case did the outcome of the analyses differ materially from the outcomes described above, nor did FD explain a significant fraction of the variance in any metric of cognitive performance.

Discussion

We examined associations between mean cortical thickness, GMV, and cognitive performance in relatively large samples of younger and older adults and a smaller sample of middle-aged individuals. In agreement with prior reports (see Introduction), age group moderated the relationship between cortical thickness and cognitive performance across each of the cognitive domains we examined. Also, in agreement with prior findings, none of these relationships were reliable after controlling for variance in performance that was shared across domains. In the case of GMV, robust relationships with performance were evident for speed, fluency, and mean cognitive ability, and the relationships were moderated by age group in the cases of speed and mean cognitive ability. In contrast with the findings for cortical thickness, in addition to demonstrating a robust correlation with mean cognitive ability, GMV explained unique variance in speed component scores. When entered as predictors into a common regression model, thickness and GMV explained independent fractions of variance in mean ability and in the speed and fluency component scores. Below, we discuss these findings in relation to the 4 questions that motivated the study.

Turning first to the findings for cortical thickness, we were able to confirm our prior findings (de Chastelaine et al. 2019) that positive associations with cognitive performance are evident in older but not in younger adults (we discuss the findings for the middle-aged sample in a separate section below). We also replicated previous reports (e.g. Salthouse et al. 2015; Tsapanou et al. 2019; Hou et al. 2021) that associations between cortical thickness and performance in different cognitive domains are largely mediated by a mean ability component reflecting variance shared across the different domains. However, with the enlarged younger adult sample available here (>5 times the size of our original sample), we failed to confirm our prior report of a negative association between thickness and cognitive ability in this age group.

A possible explanation for this null finding stems from the age distribution of our younger sample, which ranged from 18 to 30 years. According to Schnack et al. (2015), the negative association between cortical thickness and intelligence peaks around age 20 years and declines over the following decade or so before reversing in direction in middle age. Given these findings, we split our younger adult sample at the median age (22 years) into younger (18–22 years; n = 101) and older (23–30 years; n = 94) subgroups. The correlation (controlling for sex, chronological age, and years of education) between mean cognitive ability (our proxy for general intelligence) and cortical thickness was small and nonsignificant in the older subgroup (r = 0.04), whereas the correlation was substantially larger and attained significance in the younger subgroup (r = −0.21, P = 0.036); the correlation with the thickness of the left hemisphere (the hemisphere for which Schnack et al. 2015 reported their strongest effects) was even larger, r = −0.28, P = 0.005. These findings, albeit from an unplanned and post hoc analysis, are consistent with those of Schnack et al. in suggesting that the negative association between cortical thickness and cognitive performance dissipates during the third decade of life.

GMV demonstrated significant associations with cognitive performance in 2 cognitive domains (the exceptions being memory and crystallized intelligence) and with mean cognitive ability. As indicated by reliable GMV by age group interactions terms, 2 of these associations (with speed and mean cognitive ability) were stronger in older than in younger adults. Indeed, we could find no evidence in our younger sample of a reliable correlation between GMV and any cognitive score. This included fluency despite its interaction with age failing to attain significance in the regression model (see Supplementary Table S3). Moreover, the correlations in the younger age group were, in the main, trivially small. These findings are at odds with the conclusions of 2 recent meta-analyses (Pietschnig et al. 2015, 2022), neither of which identified an effect of age on the association between brain volume and intelligence. The findings are not, however, the first to describe a moderating effect of age on the association between brain volume and cognitive ability. Stronger associations in adults than children were reported in the meta-analysis of McDaniel (2005). And, it was recently reported that, when combined with other structural metrics (although not independently), GMV was more strongly associated with intelligence in older than middle-aged adults (Cox et al. 2019).

It is unclear why we were unable to detect an association between cognitive performance and GMV in our younger adult sample, given the prior evidence of a robust association in this age range. One possibility is that our inclusion and exclusion criteria resulted in range restriction at the lower end of the distributions of the cognitive performance scores in our younger age group. This possibility seems unlikely, however. Although the within-group variances of the speed and fluency scores were significantly greater in the older group (P = 0.027 and P = 0.05, respectively, according to Levene’s test), this was not the case for mean cognitive ability (P = 0.067). Moreover, statistical evidence for truncation is lacking: According to Cohen’s (1959) formula, the ratio of the estimated unrestricted variance of the younger adults’ mean cognitive ability scores to the observed variance is 1.01. A second, related, possibility arises from the fact that most of the present younger participants were sampled from a college population, upwardly biasing cognitive performance relative to the general population (for example, performance on the age-normed tests in our neuropsychological test battery averaged about 0.5 SDs greater than the norms). We are unaware of any evidence pertaining to whether associations between brain volumetric measures and cognitive ability vary according to the average ability of the sample, let alone whether any such effects interact with age. The present findings suggest, however, that examination of this question might prove worthwhile.

In contrast to the findings for cortical thickness, GMV remained a robust predictor of speed component scores after controlling for the variance shared with the other cognitive domains. Of importance, the relationship between GMV and speed also remained significant (P = 0.002), and was now unmoderated by age group (P = 0.196), after controlling for mean cognitive ability. Thus, GMV demonstrated a robust association with speed independently of the contribution of the speed component to mean ability. This finding is especially remarkable, given the high correlation between the speed and mean ability scores (r = 0.88, controlling for age group, sex, and years of education).

The reasons for and the functional significance of the association between speed and GMV are unclear, and we are unaware of any direct precedent in the literature (but see Brugulat-Serrat et al. 2020). In an effort to elucidate the association, we repeated the regression analysis described above using as independent variables the scores on the 3 tests—Trails A, Trails B, and Symbol Digit Modalities—that loaded most heavily on the speed component. The results were unilluminating; each score demonstrated a statistically significant but small correlation with GMV that did not exceed the correlation obtained for the component score. Given prior evidence of an association between white matter integrity and processing speed (e.g. Madden et al. 2008; Kerchner et al. 2012), we also examined whether GMV was acting as a proxy for WMV (the 2 measures correlated at r = 0.65). As described in the Supplemental Materials, unlike GMV, which continued to be a robust predictor of speed scores, WMV did not explain any variance in speed when the 2 variables were entered as predictors into the same regression model. Therefore, it seems reasonable to conclude that it is GMV, specifically, that is a predictor of processing speed. The functional and neurobiological significances of this association are unclear.

Cortical thickness and GMV each accounted for significant fractions of the variance in cognitive performance, but the effect sizes were small (R² ranging from around 0.03 to 0.04), as is typical in studies such as this (Salthouse et al. 2015; Pietschnig et al. 2022). The 2 metrics did, however, explain unique variance in 3 of the cognitive scores—speed, fluency, and mean ability—with which they demonstrated independent associations. Thus, cortical thickness and GMV appear to reflect distinct neurobiological correlates of cognitive performance. When thickness and GMV were employed as joint predictors of performance, the amount of explained variance remained modest—the largest fraction was for the speed component where, after controlling for age group, years of education, and sex, the metrics jointly accounted for an additional 5.3% of the sample-wide variance. Within the older sample, for which correlations between the brain metrics and cognition were strongest, the variance in the speed component jointly explained by cortical thickness and GMV still only amounted to some 7% of the total after controlling for sex, age, and years of education. Therefore, it appears that these neural measures have the potential to provide only limited insight into the determinants of cognitive ability in cognitively healthy older individuals.

A major obstacle to understanding the significance of the associations between cognitive ability and cortical thickness and GMV identified here and in prior research is the paucity of knowledge about their neurobiological determinants. T1-weighted MRI provides no information about cortical microstructure, let alone the microstructural bases of individual differences in cortical thickness at different stages in the lifespan (the reversal in the direction of the relationship between thickness and cognitive ability in early relative to late adulthood strongly suggests that the factors driving individual differences in thickness differ with age). A potentially important insight into these factors comes from a recent study (Heyer et al. 2022) of a sample of epilepsy patients (mean age = 39 years) who underwent left temporal lobectomy. Histological analysis of resected cortical tissue from the middle temporal gyrus revealed that individual differences in the thickness of this region were driven solely by variation in the thickness of cortical layers II and III. Moreover, increased thickness was associated with fewer but larger and more highly arborized pyramidal neurons in these layers. Greater thickness was also associated with higher preoperative intelligence scores. These findings offer intriguing clues about the neurobiological underpinnings of the association between cortical thickness and adult cognitive ability. It remains to be seen whether the findings generalize to younger or older populations or to other cortical regions.

Even less is known about the neurobiological determinants of individual differences in GMV, or of brain size more generally. The notion that larger brains confer a cognitive advantage is frequently advanced in an evolutionary context (e.g. Gonda et al. 2013; Lee et al. 2019) but is subject to numerous caveats (e.g. Striedter 2005; Nave et al. 2019). Not least, much of the variance in human brain volume is associated not with cognitive ability but with sex (see Deary et al. 2010 for discussion of this apparent paradox). In the present study, for example, despite GMV’s that were on average some 10% lower than those of their male counterparts, memory component scores were robustly higher in females (consistent with prior findings, Asperholm et al. 2019; Hirnstein et al. 2022), and no sex difference was evident in mean cognitive ability. Clearly, the factors responsible for sex differences in brain volume are independent of those that mediate its much weaker association with cognitive ability.

In addition to the samples of younger and older adults, we also included a small sample of middle-aged individuals. Unsurprisingly, given the modest effect sizes evident even in the older sample for associations between either neural metric and cognitive performance, no significant associations were detected in the middle-aged group. We note, however, that in regression analyses restricted to the middle-aged and older samples—in which cognitive performance was predicted by sex, years of education, age group, and either cortical thickness or GMV and their interaction with age group—we found no evidence that the strength of the associations between cognitive performance and the neural measures differed significantly between the age groups. Moreover, partial correlations between each measure and cognitive performance differed little from those for the older group alone—for example, for mean cognitive ability, the correlations (controlling for age, sex, and years of education) in the combined group were r = 0.23 (P < 0.002) and r = 0.20 (P < 0.008) for thickness and GMV, respectively, as opposed to r = 0.23 and r = 0.22 for the older group alone. These findings are consistent with the possibility that the associations with cognitive performance we identified in our older age group extend to middle age. Confirmation of this possibility will of course require the employment of a much larger middle-aged sample than was employed here.

The present study suffers from several limitations. First, the sample sizes are low by contemporary standards, most especially in respect of the middle-aged sample, limiting sensitivity to small effects. Second, the neuropsychological test battery was heavily weighted toward verbal cognition. Notably, there were no tests of spatial or visual memory and only a single test of reasoning (Progressive Matrices). Third, our study samples were unrepresentative of the general population. Mean cognitive performance was relatively high, and most participants were white, and college educated. Lastly, the study design was crosssectional, and hence, does not allow any inferences to be made about the effects of aging (rather than age) on cognitive or neural metrics and their associations.

These limitations notwithstanding, the present findings add to the existing literature in several ways. Of most significance, the findings indicate that cortical thickness and GMV are nonredundant predictors of cognitive performance, suggesting that the 2 metrics are neurobiologically distinct correlates of cognitive ability. Furthermore, while confirming a general lack of domain-specificity in associations between cortical thickness and cognitive performance, the findings point to a possible exception in the case of GMV and its association with processing speed. If confirmed, this association deserves further scrutiny. More tentatively, given the post hoc nature of the analyses, the findings converge with prior research to suggest that the negative associations between cortical thickness and cognitive ability that have consistently been reported in adolescence extend into adulthood but not beyond about 23–25 years of age.

Funding

This work was supported by the National Institute on Aging (grant numbers R21AG054197 and RF1AG039103) and the National Science Foundation (grant number 1633873).

Conflict of interest statement: None declared.

Data availability

The data generated by this study are undergoing additional analyses. The data that support the findings of this study are available from the authors on request subsequent to a formal data sharing agreement.