White Matter Integrity Supports BOLD Signal Variability and Cognitive Performance in the Aging Human Brain

Decline in cognitive performance in old age is linked to both suboptimal neural processing in grey matter (GM) and reduced integrity of white matter (WM), but the whole-brain structure-function-cognition associations remain poorly understood. Here we apply a novel measure of GM processing–moment-to-moment variability in the blood oxygenation level-dependent signal (SDBOLD)—to study the associations between GM function during resting state, performance on four main cognitive domains (i.e., fluid intelligence, perceptual speed, episodic memory, vocabulary), and WM microstructural integrity in 91 healthy older adults (aged 60-80 years). We modeled the relations between whole-GM SDBOLD with cognitive performance using multivariate partial least squares analysis. We found that greater SDBOLD was associated with better fluid abilities and memory. Most of regions showing behaviorally relevant SDBOLD (e.g., precuneus and insula) were localized to inter- or intra-network “hubs” that connect and integrate segregated functional domains in the brain. Our results suggest that optimal dynamic range of neural processing in hub regions may support cognitive operations that specifically rely on the most flexible neural processing and complex cross-talk between different brain networks. Finally, we demonstrated that older adults with greater WM integrity in all major WM tracts had also greater SDBOLD and better performance on tests of memory and fluid abilities. We conclude that SDBOLD is a promising functional neural correlate of individual differences in cognition in healthy older adults and is supported by overall WM integrity.


Introduction
Cognitive performance, such memory, reasoning, perceptual speed, and maintenance of semantic knowledge, relies on the neural processing in grey matter (GM) and the integrity of white matter (WM). Many neuroimaging studies attempt to link age-related differences in cognitive performance with either blood-oxygenation level dependent (BOLD) signal magnitude (TICS-M) questionnaire, (4) scored < 10 on the geriatric depression scale (GDS-15), (5) scored ! 75% right-handedness on the Edinburgh Handedness Questionnaire, (6) demonstrated normal or corrected-to-normal vision of at least 20/40 and no color blindness, (7) were cleared for suitability in the MRI environment, that is, no metallic implants that could interfere with the magnetic field or cause injury, no claustrophobia, and no history of head trauma. The participants were a pre-intervention cross-sectional subsample from an on-going randomized controlled exercise trial ("Influence of Fitness on Brain and Cognition II" at ClinicalTrials.gov, clinical study identifier NCT01472744), from whom good quality anatomical and resting state functional MRI (see section 2.4 and 2.6) was available.

Cognitive assessment and analysis
We administered a cognitive battery as described in the Virginia Cognitive Aging Project [17][18][19][20] to measure latent constructs of fluid intelligence, perceptual speed, episodic memory, and vocabulary (for more details on each task see Table 1). The computer-based tasks were Computer-based [77] .628 --.418 Shipley abstraction

Fluid intelligence
Determine the letters, words, or numbers that best complete a progressive sequence Paper-pencil [78] .525 --.564 Letter sets Fluid intelligence Identify which of five groups of letters is different from the others Computer-based [79] .346 .410 -.575 Spatial relations Spatial reasoning Determine which three dimensional object could be constructed by folding the two dimensional object Computer-based [80] .788 ---

Paper folding Spatial reasoning
Determine the pattern of holes that would result from a sequence of folds and a punch through folded paper Computer-based [79] .856 ---

Form boards Spatial reasoning
Determine shapes needed to fill in a space Computer-based [79] .725 --- To obtain components representing the four cognitive constructs and to confirm the validity of task structure as presented in [20], we performed principal component analysis (PCA) with varimax rotation. Individual scores on each of the 16 tasks were first screened for outliers and winsorized (maximum 3 cases out of 91 (<3%) were adjusted per variable). The resulting constructs are presented in Table 1 and the component scores were saved as variables.
Some participants did not complete all tasks in the cognitive battery, which resulted in a final sample of 91 participants (29 males, age range 60-78, M age = 65 ± 4 years, years of education 12-26, M edu = 17 ± 4 years).

MRI acquisition
We acquired all images during a single session on a 3T Siemens Trio Tim system with 45 mT/ m gradients and 200 T/m/sec slew rates (Siemens, Erlangen, Germany). T2 Ã -weighted resting state images were acquired with fast echo-planar imaging (EPI) sequence with Blood Oxygenation Level Dependent (BOLD) contrast (6min, TR = 2s, TE = 25ms, flip angle = 80 degrees, 3.4 x3.4 mm 2 in-plane resolution, 35 4mm-thick slices acquired in ascending order, Grappa acceleration factor = 2, 64 × 64 matrix). The participants were instructed to lay still with eyes closed. Additionally, gradient field maps were acquired to account for geometric distortions caused by magnetic field inhomogeneity [28]. The gradient field map was collected as 35, 4mm-thick slices, 3.4 x 3.4 mm 2 in-plane resolution, TR = 700ms, TE = 10ms, and flip angle = 35 degrees.
DTI images were acquired with a twice-refocused spin echo single-shot Echo Planar Imaging sequence [29] to minimize eddy current-induced image distortions. The protocol consisted of a set of 30 non-collinear diffusion-weighted acquisitions with b-value = 1000s/mm 2 and two T2-weighted b-value = 0 s/mm 2 acquisitions, repeated two times (TR/TE = 5500/98 ms, 128 x 128 matrix, 1.7x1.7 mm 2 in-plane resolution, FA = 90, GRAPPA acceleration factor 2, and bandwidth of 1698 Hz/Px, comprising 40 3-mm-thick slices). Resting state and DTI images were obtained parallel to the anterior-posterior commissure plane with no interslice gap.

BOLD variability (SD BOLD ) calculation
Data preprocessing was carried out using FSL v5.0.1 (FMRIB's Software Library, http://www. fmrib.ox.ac.uk/fsl; [30]). The preprocessing included high pass filtering (> 0.008Hz), slice timing correction, rigid body motion correction using MCFLRT [31], and removal of non-brain tissue with the Brain Extraction Tool [32]. Data from all subjects was screened for motion and all participants moved within a voxel dimension (< 4mm). Functional images of each participant were aligned to the standard stereotaxic space of the MNI 152 T1 2mm 3 template supplied in FSL in a three-step procedure. To improve the registration between the participant's functional and anatomical images we utilized the gradient field map data. First, the gradient field map was unwrapped via PRELUDE [33], then geometric distortions in the EPI-related images due to local magnetic inhomogeneity differences were compensated for with the use of gradient field map data via FUGUE within FSL [33]. Eleven out of 91 participants had missing field map images. Second, each participant's low-resolution functional images were aligned with their high-resolution T1-weighted anatomical images using the Boundary-Based Registration in FSL [34]. Third, the anatomical images were aligned to MNI 152T1 2mm 3 template using 12 degrees of freedom affine linear registration [31].
Next, as recommended by [13], we used Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC v3.10) tool in FSL [35] to decompose the 4D fMRI time series into spatial and temporal components. AZB together with Chanheng He and CNW identified artifact components for each subject using the criteria outlined in [36] based on the spatial pattern, time course, and power spectrum properties that were characteristic of physiological noise, motion, and scanner-related artifacts. The artifactual components were regressed out from the time series yielding the post-ICA 'cleaned' data. This post-ICA functional data as well as the six motion parameters outputted earlier by motion correction were bandpass filtered to restrict the frequencies in our data to. 008 < f <. 1 Hz [37]. Next, we extracted mean time series from two nuisance regions of interest (deep temporal white matter, bilateral lateral ventricles) in the post-ICA filtered data. The goal of including these two nuisance regressors is to remove residual cardiorespiratory physiological noise that would be captured by signal changes in the white matter and ventricles [38][39][40][41] and was not removed by the ICA cleanup. The two nuisance regressors (timeseries from white matter and ventricles) were regressed out using the general linear model with FEAT 6.00 (FMRI Expert Analysis Tool; http://www.fmrib.ox.ac.uk/analysis/research/feat/). Finally, we calculated the standard deviation (SD BOLD ) across the whole timeseries for each voxel and smoothed the images with a 6mm Gaussian kernel. The resulting SD BOLD maps were upsampled to MNI space using the registration steps described above. To restrict all multivariate analyses to the GM, we masked the SD BOLD maps with the grey matter tissue prior provided in FSL, thresholded at probability > 0.37. The intermediate outcomes of all the above procedures were carefully inspected by AZB and CNW.

PLS multivariate analysis of relations among SD BOLD , cognitive performance and fitness
First, we made sure that all behavioral variables were normally distributed and any outliers (> 2.5 SD) were accounted for by winsorizing, where not more than 2 cases were corrected per variable (2%).
The behavioral PLS analysis [42,43] begins with a correlation matrix (CORR) between our variables of interest (four cognitive constructs) and each voxel's signal (SD BOLD ); correlations are calculated across subjects. Then, this CORR matrix is decomposed via singular value decomposition (SVD): SVD CORR = USV'. This decomposition produces a left singular vector of behavioral weights (U), a right singular vector of SD BOLD weights (V), and a diagonal matrix of singular values (S). In other words, this analysis produces orthogonal latent variables (LVs) that optimally represent relations between behavior and SD BOLD in grey matter voxels. Each LV contains a spatial pattern depicting the brain regions where the SD BOLD shows the strongest relation to behavior. Each brain weight (in V) is proportional to the correlation between behavior and SD BOLD in all of the tracts. To obtain a summary measure of each participant's expression of a particular LV pattern, we calculated within-person "brain scores" by multiplying each voxel (i)'s weight (V) from each LV (j) produced from the SVD in equation (1) by the SD BOLD value in that voxel for person (k), and summing over all (n) brain voxels: in a single measure, a brain score indicates the degree to which a subject expresses the multivariate spatial pattern captured by a given behavior-driven latent variable. Significance of detected relations between multivariate spatial patterns and cognitive performance was assessed using 1000 permutation tests of the singular value corresponding to each LV. A subsequent bootstrapping procedure revealed the robustness of voxel saliences across 1000 bootstrapped resamples of our data [44]. By dividing each voxel's mean salience by its bootstrapped standard error, we obtained "bootstrap ratios" as normalized estimates of robustness. We thresholded bootstrap ratios at a value of ! 3.00, which approximates a 99% confidence interval and corresponds to p-value of <.001.

DTI analysis
DTI allows inferences about WM microstructure in vivo by quantifying the magnitude and directionality of diffusion of water within a tissue [45]. Visual checks were performed on every volume of the raw data of every participant by AZB. Sixty-six participants had good quality DTI data. In one dataset, one volume with the corresponding b-vectors and b-values was deleted from the dataset before processing due to artifact. Next, DTI data were processed using the FSL Diffusion Toolbox v.3.0 (FDT: http://www.fmrib.ox.ac.uk/fsl) in a standard multistep procedure, including: (a) motion and eddy current correction of the images and corresponding bvectors, (b) removal of the skull and non-brain tissue using the Brain Extraction Tool [32], and (c) voxel-by-voxel calculation of the diffusion tensors. Using the diffusion tensor information, FA maps were computed using DTIFit within the FDT. All motion-and eddy-current outputs, as well as FA images were visually inspected. We used TBSS [46,47], a toolbox within FSL v5.0.1, to create a representation of main WM tracts common to all subjects (WM "skeleton"). This included: (a) nonlinear alignment of each participant's FA volume to the 1 x 1 x 1 mm 3 standard Montreal Neurological Institute (MNI152) space via the FMRIB58_FA template using the FMRIB's Nonlinear Registration Tool (FNIRT, [48]; http://www.doc.ic.ac.uk/~dr/software), (b) calculation of the mean of all aligned FA images, (c) creation of the WM "skeleton" by perpendicular non-maximum-suppression of the mean FA image and setting the FA threshold to 0.25, and (d) perpendicular projection of the highest FA value (local center of the tract) onto the skeleton, separately for each subject. The outputs of all the above processing steps were carefully inspected by AZB. Given that SD BOLD is a relatively new way to asses brain function and its structural brain correlates are not yet understood, we did not make any regional predictions and used a global FA measure, obtained by averaging FA over the whole skeleton for each participant.

Post hoc statistical analyses
All statistical analyses were performed using SPSS (v.16, SPSS Inc., Chicago, IL, USA). We used multiple step-wise linear regressions (with chronological age and gender) to investigate the relationships between brains scores from SD BOLD -cognition and global FA. Two participants brain scores had outlier values > 2.5 SD and their values were winsorized, which did not change the results and was used for display purposes.
The demographic data, FA values, behavioral scores, and brain scores are available in S1 Dataset.

Correlations between cognitive performance and SD BOLD
To investigate the relationships between SD BOLD and performance on four main cognitive domains, we first performed principal component analysis (PCA) on 16 tasks from Table 1 to reduce their dimensionality. We replicated the findings of the Salthouse studies [17][18][19][20] by obtaining the four expected components of fluid intelligence, perceptual speed, episodic memory, and vocabulary (Table 1). Only speed (r = -.33 p = .002) and memory (r = -31, p = .003) components were negatively related to age, whereas fluid abilities (r = -.18 p = .098) and vocabulary (r = .11 p = .320) were not.
Next, to identify multivariate across-subject patterns of relations between SD BOLD at rest in the entire GM and the scores from the four cognitive components we performed behavioral PLS analysis. Importantly, previous studies related SD BOLD within spliced fixation periods in blocked fMRI series to performance on task [13,14], while the current study is the first application of resting state data in investigating behaviorally relevant SD BOLD . The behavioral PLS analysis begins with the correlation matrix between the individual scores on the four cognitive components and each voxel's SD BOLD ; correlations are calculated across subjects. Then, this matrix is decomposed via singular value decomposition. This decomposition produces orthogonal latent variables (LVs) that optimally represent relations between SD BOLD in GM voxels and cognitive performance. Each LV contains a spatial pattern depicting the brain regions where the activity shows the strongest relation to performance. In this analysis, because we examined the association with four cognitive components, four outcome latent variables (LV) were possible. We predicted that if there are domain-specific patterns of optimal SD BOLD , then multiple LVs may be significant, each representing an association between a different cognitive construct and BOLD variability. Alternatively, if SD BOLD is a more general feature common to different cognitive functions, there should be one LV representing the brainperformance relationship.
Our results supported the latter hypothesis: the PLS multivariate analysis yielded one significant latent variable (permuted p = 0.023, 59.46% cross-block covariance explained by this LV), suggesting that, overall, higher SD BOLD was related to better performance on fluid and memory constructs and lower performance on vocabulary. This relationship was reversed in only two small clusters (Fig. 1A). The same analysis with additional controlling for the global signal (i.e. centering the mean across the volumes) yielded the same spatial pattern, where higher SD BOLD was related to better performance on fluid and memory constructs (permuted p = 0.007, 60.41% cross-block covariance explained by this LV).
If the PLS model was run with vocabulary only (1 LV possible), only the clusters in temporal fusiform and cerebellum were above p <. 001 threshold, but the overall model was not significant. This suggests that the red-yellow cluster shown in Fig. 1A is attributable to the relationship with vocabulary. Similarly, a model with 4 cognitive constructs and additionally chronological age (5 LVs possible) explained only ca. 3% more of cross-block covariance than the four construct model from Fig. 1, and showed the same spatial pattern. This suggests that age is not driving the function-performance result from Fig. 1A. In this model age was positively related to vocabulary performance, but inversely to memory and fluid abilities, and SD BOLD.
Perceptual speed did not significantly contribute to the observed performance-SD BOLD correlation pattern, although there was a trend towards greater perceptual speed being related to lesser SD BOLD. Peak voxels' location and bootstrap ratios are reported in Table 2.

WM integrity predicts function-cognition relations independent of age
Next, we investigated whether the observed associations between memory and fluid performance and SD BOLD are related to the integrity of structural connections in the brain. To examine this hypothesis, we first obtained a summary measure of each participant's expression of the significant LV pattern by calculating "brain scores". This involved multiplying each voxel's weights from the significant LV by the SD BOLD in that voxel for each person, and summing it over all brain voxels. Thus, in a single measure, a brain score indicates the degree to which a subject expresses the multivariate spatial pattern of performance-SD BOLD associations reported in the LV depicted in Fig. 1 (see Methods for more details on brains score calculation).
Specifically, a person with a higher brain score showed better performance on memory and fluid abilities and greater SD BOLD in the voxels depicted in Fig. 1A.
Finally, we performed a multiple regression analysis with the brain scores as a dependent variable, age as the first independent variable and global FA (mean FA across the main WM tracts) as the second independent variable. Note that DTI data was available from only 66 out of 91. We included age in the model as both global FA (r = -.38 p = .002 n = 66) and brain scores (r = .21 p = .048, n = 91) were negatively related to age. In addition, memory was negatively related to age (see previous section). Therefore, it was important to test whether the SD BOLD -performance association is related to WM microstructure beyond the effects of chronological age. Indeed, we found that higher FA accounted for a significant amount of variance in brain scores, in addition to variance related to age (R 2 Δ age = 0.041, F c Δ age = 2.77, df = 64/ 1, p-value = .101; R 2 Δ globalFA = 0.12. F Δ globalFA = 8.7, df = 63/1, p-value = .004). We also note that global FA was not related to perceptual speed, memory and vocabulary components (p >. 50) and was related to fluid abilities only at a trend level (r = .23 p = .068, n = 66). Together, our results suggest that global WM integrity is associated with behaviorally relevant variability in the BOLD signal, beyond the effects of age.
We run an additional PLS model including age, four behavioral scores, and global FA (n = 66). It yielded one LV (p = .005, cross block covariance explained of 63%), where greater  FA and younger age was related to greater SD BOLD . Global FA contributed most to the relationship (r>0.4), and age to a lesser degree (r>0.2). Greater fluid intelligence and memory were also related to greater SD BOLD, but their contribution to the model was not significant (while vocabulary and processing speed showed a negative non-significant association). This result confirms that WM integrity is related to SD BOLD , that brain structure-function relationship may be stronger than brain-performance associations, and this issue should be further investigated (see Discussion). We highlight, however, that the purpose of this article was to investigate the structural WM correlates of behaviorally relevant SD BOLD only.

Discussion
We investigated the associations between resting SD BOLD and performance on four distinct cognitive constructs in healthy older adults with a whole-brain, multivariate approach. We demonstrated that 1) better fluid abilities and memory was linked to greater SD BOLD in multiple regions including precuneus, insula, temporal, parietal, and prefrontal regions, and cingulate, and 2) behaviorally relevant SD BOLD pattern was shared by fluid abilities and memory. Moreover, inter-individual differences in these SD BOLD -cognition relationships were related to the global WM integrity, above and beyond the effects of chronological age.

Association of SD BOLD with performance differs by cognitive domain
A previous study reported that greater SD BOLD in healthy adults was associated with younger age, faster, and more consistent response times (RT) across three levels of a perceptual matchto-sample task (immediate comparison, cued short-delay comparison, and delayed comparison; [14]). Our results provide further evidence for greater SD BOLD being related to better performance in aging. Specifically, we showed that the cognitive constructs requiring adaptive and flexible processing-fluid abilities and memory-were driving this positive SD BOLD -performance association. For example, tasks defining the fluid abilities require abstract reasoning and problem solving that enable optimal adaptation to a changing and complex environment [49]. Similarly, episodic memory involves association formation and binding, as well as flexible and context-dependent retrieval. As a result, both fluid abilities and memory should benefit from greater dynamic range and the ability to explore different network states at the neuronal level [4,12,50]. On the contrary, the vocabulary construct representing semantic knowledge requires robust retrieval of information from long-term memory that was acquired, stored, and reinforced over years. Thus, vocabulary knowledge operates on "hard-wired", automatic and repetitive responses and therefore may benefit from less SD BOLD at the neural level. As an additional behavioral PLS analysis with only vocabulary construct did not yield a significant LV, this result relating lower SD BOLD to better vocabulary performance should be treated as preliminary and further investigated with more cognitive tasks defining this domain.
The dissociation of SD BOLD -performance relationship between the cognitive domains parallels their differential sensitivity to age. Namely, advanced age is related to decline in fluid abilities, memory and speed, with relative sparing of vocabulary knowledge [51,52]. The regions where we observed an association of SD BOLD with fluid abilities and memory (visual cortex, temporal pole, insula, cingulate, parietal cortex, lateral frontal regions) overlap with regions showing decreased SD BOLD in older compared to younger adults [13]. Therefore, we speculate that SD BOLD might be one of the neural correlates underlying the discrepancy of age-related effects on the four main cognitive domains. Further exploration of this claim should be done by extending analyses to samples with broader age range.

Behaviorally relevant SD BOLD may support integration of brain networks
Many regions where we observed a positive association of SD BOLD with fluid abilities and memory have been defined as degree-based hubs, "rich club" regions, or connector hubs in structural and functional network analyses: posterior cingulate cortex, superior frontal, parietal and insular cortex, as well as inferior temporal and fusiform cortex [22,[53][54][55]. Brain "hubs" are regions with high connectivity degree in a given neural community [23,55,56], while "richclub" regions are the high-degree hubs that tend to connect to each other [57]. Of particular relevance to our findings are the connector hubs: regions highly connected primarily to distinct brain networks [58][59][60]. Such connector hubs are localized to the insula, parietal, premotor, lateral occipital, and dorsal superior frontal cortex [60], where we also observed higher SD BOLD in better performing older adults. Connector hubs integrate functionally segregated domains with possibly very distinct processing or oscillatory properties. We therefore suggest that the hub's high connectivity with multiple brain functional networks requires or results in the higher moment-to-moment variability in neural function, which should be reflected by greater SD BOLD . Importantly, we predict that such SD BOLD related to a region's cross-talk between different neural networks should be driven by high variability in signal frequency and not only by variability as a result of high amplitude signal with a constant frequency. Clearly, our results need to be followed by a direct comparison of SD BOLD patterns with functional connectivity network properties, time-frequency analyses to tease apart time-constant SD BOLD from time-varying SD BOLD , as well as changes in SD BOLD and power-law exponents in fMRI signal between rest and task states [61,62], and their significance for cognitive performance in aging.
Despite careful removal of physiological noise with ICA, we acknowledge that some of SD BOLD regions, such as posterior cingulate, occipital cortex and regions near large vessels such as temporal pole and regions along the brain midline, may partly overlap with respiratory or cardiac-related fluctuations [63,64]. High static cerebral blood flow (CBF) and high amplitude of low-frequency fluctuations in CBF at rest in regions such as posterior cingulate cortex and insula, however, suggest that spontaneous fluctuations of fMRI signal in these regions are neuronally-driven rather than of vasomotor origin [65].
Finally, we note that our analysis yielded one model for memory and fluid abilities instead of two LVs specific for each cognitive construct. This further supports the possibility that the hub-related pattern of greater SD BOLD represents a common rather than a domain-specific neural feature. In other words, our findings suggest that preserving high SD BOLD in regions associated with intra-and inter-network communication is linked with better performance on a set of cognitive tasks requiring flexible neural processing. We speculate that hub regions that show greater SD BOLD during spontaneous brain activity at rest would also have the capacity for increased neural processing complexity during cognitive tasks (e.g. memory and reasoning) [5,62]. Longitudinal designs and broader age ranges should help to tease apart age-related from individual differences in SD BOLD .

White matter as a scaffold for behaviorally relevant SD BOLD
Our study provided the first evidence for an association between whole-brain behaviorally relevant variability in the BOLD signal and WM integrity. Therefore, our study further extends previous reports on a positive relationship between WM integrity and task-related changes in BOLD signal [27,66], and structure-function brain network properties [54,67,68]. We propose that poor WM integrity, most likely due to age-related changes in myelination, precludes fast and reliable signal transduction. Consequently, optimal interaction between brain hubs within or between brain networks becomes impaired [69,70]. For instance, some signals may be "lost" in between the GM regions, others may not arrive in a timely fashion to be optimally integrated in the neural processes [27,71], or the resting kinetic energy of the system may not be sufficient to adjust to externally driven cognitive challenges [5,62]. This may result in reduced processing complexity that could be detected as reduced SD BOLD at rest and during task, for instance, during the creation of mnemonic representations or updating information during mental rotation.
Our result that older adults with greater FA in all major WM tracts had greater SD BOLD and better performance on memory and fluid abilities converges with previous reports on relationships between diffusivity properties and fluid intelligence defined by reasoning abilities, cognitive flexibility, episodic memory, and processing speed in older adults [25,[72][73][74][75]; for a review see also [76]. Although we observed only a trend relationship between WM integrity and fluid abilities, this lack of strong diffusion-cognition association may be because our participants represented a relatively narrow age range and being relatively high functioning, healthy older adults (all qualified for the MRI, aerobic capacity test and an exercise intervention), which may limit the variability in the FA and behavioral measures.
Together, our data suggests that magnitude and spatial pattern of SD BOLD that is linked to high cognitive performance-and therefore represents optimal complexity of neural processing-relies on the integrity of structural brain connectivity via WM in the healthy aging brain. Our findings lay foundation for future investigations addressing more specific questions about structural correlates of SD BOLD . One direction will be to define the regional (both GM and WM) specificity of WM-SD BOLD associations in aging and across lifespan. Another important issue that needs to be addressed is the role of cortical atrophy and the related partial volume effect in estimating SD BOLD in aging population, and the possible mediating role of GM volume on the SD BOLD -cognition associations.

Conclusions
We found that greater SD BOLD in multiple brain regions, most of which have been identified as inter-or intra-network connecting hubs, was linked to better fluid abilities and memory. This suggests that optimal dynamic range of neural processing in hub regions may support cognitive operations that specifically rely on moment-to-moment processing adaptability and flexibility. Moreover, we showed that this behaviorally relevant SD BOLD is supported by global WM integrity. We conclude that SD BOLD is a promising functional neural correlate of individual differences in cognition in healthy older adults.
Supporting Information S1 Dataset. Demographic, DTI, cognitive, and brain score data for the 91 participants. "Win" in the variable name indicates this variable was winsorized. (XLSX)