Generalizing age effects on brain structure and cognition: A two‐study comparison approach

Abstract Normal aging is accompanied by an interindividually variable decline in cognitive abilities and brain structure. This variability, in combination with methodical differences and differences in sample characteristics across studies, pose a major challenge for generalizability of results from different studies. Therefore, the current study aimed at cross‐validating age‐related differences in cognitive abilities and brain structure (measured using cortical thickness [CT]) in two large independent samples, each consisting of 228 healthy older adults aged between 65 and 85 years: the Longitudinal Healthy Aging Brain (LHAB) database (University of Zurich, Switzerland) and the 1000BRAINS (Research Centre Jülich, Germany). Participants from LHAB showed significantly higher education, physical well‐being, and cognitive abilities (processing speed, concept shifting, reasoning, semantic verbal fluency, and vocabulary). In contrast, CT values were larger for participants of 1000BRAINS. Though, both samples showed highly similar age‐related differences in both, cognitive abilities and CT. These effects were in accordance with functional aging theories, for example, posterior to anterior shift in aging as was shown for the default mode network. Thus, the current two‐study approach provides evidence that independently on heterogeneous metrics of brain structure or cognition across studies, age‐related effects on cognitive ability and brain structure can be generalized over different samples, assuming the same methodology is used.


| INTRODUCTION
As we get older, our brain undergoes substantial structural changes that seem to be related to changes in behavior (i.e., cognitive decline in older adults). However, previous research has shown that it is far from simple to bring the two domains-namely brain structure and behavior-together (Fjell et al., 2006;Jockwitz et al., 2017;Liu et al., 2011;Raz & Rodrigue, 2006;Ziegler, Dahnke, Gaser, & Alzheimer's Disease Neuroimaging, 2012). One important reason for this is that age-related changes in both domains are complex and insufficiently understood. For example, large between-study heterogeneity of designs and methods, differences in sample characteristics and the generally larger interindividual variability in samples of older adults hamper the extraction of consistent findings regarding age-related changes in brain structure in the existing literature.
Although there is a more solid database when it comes to cognitive aging (Schaie (1993); Schaie and Willis (2010); Schaie, Willis, and Caskie (2004); for reviews, see Harada, Love, and Triebel (2013); Kaup, Mirzakhanian, Jeste, and Eyler (2011); Salthouse (2010) it has also been established that-in analogy to brain aging-age-related changes in cognitive abilities are complex. First, different cognitive abilities are differentially sensitive to age effects. Abilities such as processing speed, executive functions, episodic, and working memory have shown to be more vulnerable to age-related decline as compared to verbal memory and world knowledge (Habib, Nyberg, & Nilsson, 2007;Hedden & Gabrieli, 2004;Park & Reuter-Lorenz, 2009;Schaie et al., 2004;Schaie & Willis, 2010). And second, several studies suggest that cognitive performance follows nonlinear trends from early to late adulthood with a higher interindividual variability in older adults (Habib et al., 2007;Hartshorne & Germine, 2015;Hedden & Gabrieli, 2004). Hence, it is difficult to generalize results from one sample to another and, therefore, to draw reliable conclusions. Considering, for example, that lifespan trajectories of structural atrophy vary between brain regions Hogstrom et al., 2013;Sowell et al., 2003;Walhovd et al., 2011;Ziegler, Dahnke, Jancke, et al., 2012), age-related differences in brain atrophy might not be replicable across samples when they do not match with respect to age distributions or other sample characteristics.
At this time, there is a clear progress toward brain imaging consortia and multicenter studies, such as ENIGMA (Thompson et al., 2014), the German National Cohort study (Nationale Kohorte; NAKO (Bamberg et al., 2015;German National Cohort, 2014), ADNI (Alzheimer's Disease Neuroimaging Initiative ;Jack Jr. et al., 2008), U.K. Biobank (Miller et al., 2016;Sudlow et al., 2015), or Lifebrain (Walhovd et al., 2018). In the field of healthy aging, such projects use data pooling procedures (i.e., joint analysis of data from different independent samples) to fulfill the need for large sample sizes required to identify protective and risk factors that in combination might explain why some older adults develop neurodegenerative diseases, while others retain their cognitive integrity until very old. What comes along with this, however, is the necessity for a cross-validation of so far established results concerning the aging brain. Thus, the question that arises is whether independent samples of older adults that differ in demographics and lifestyle factors would still show similar association patterns between age, global and regional brain structure, and cognitive performance. While in the field of genetics, replication studies are already well established, it is not yet common practice in the field of neuroimaging. Therefore, the current study analyzed age-related differences in brain structure and cognitive ability in two large independent but closely matched cohorts of older adults-both situated in central Europe-to explore how similar results are when using the same state-of-the-art methodological protocols and what factors may explain potential between-study differences.
Regarding brain structure, we used mean CT for the two hemispheres as a rough outcome measure. In addition to that, we decided to focus on brain regions that constitute the default mode network (DMN), a network that recently received much attention in aging researchespecially with regard to functional connectivity (e.g., Hafkemeijer, van der Grond, & Rombouts, 2012). Because recent evidence from our group suggests a structural correlate for age differences in functional connectivity (Jockwitz et al., 2017), we were particularly interested to validate such first findings and assessed regional within-network differences of the age-brain structure relationships.

| METHODS
Participants included in the current research project were recruited from two independent samples investigating brain-behavior relationships in older adults located in the larger Zurich area (Switzerland) and in the Ruhr district (Germany).
One sample comprised the ongoing Longitudinal Healthy Aging Brain (LHAB) database project at the University Research Priority Program "Dynamics of Healthy Aging" of the University of Zurich (Zollig et al., 2011). LHAB investigates age-related dynamics of brainbehavior relationships in healthy older adults. A particular focus is placed on assessing and explaining interindividual variability in the observed aging trajectories, thus a broad spectrum of factors that supposingly influence such trajectories (i.e., lifestyle, sleep, and nutrition) is collected. In LHAB, older adults from Zurich and surrounding areas aged 65 and older (at baseline) are observed longitudinally with between-measurement intervals of 1-2 years. Besides the eligibility requirements for the MR acquisition, further exclusion criteria were neurological and psychiatric diseases, a score on the Mini-Mental State Examination of 26 and below and left handedness. LHAB participants are German native speakers or at least as proficient in German as it would be their native language. The study protocol was approved by the local Ethics Committee (Kantonale Ethikkommission Zurich).
The initial sample of LHAB comprised 231 participants ranging from 64 to 87 years of age. Data acquisition in the LHAB project started in 2011. Currently, the data set covers an observation period of 4 years.
The second sample comprised 1000BRAINS at the Institute of Neuroscience and Medicine, Research Centre Jülich, a longitudinal population-based study that assesses variability in brain structure and function during aging (Caspers et al., 2014). The 1000BRAINS sample is drawn from the 10-year follow-up cohort of the Heinz Nixdorf Recall Study, an epidemiological population-based study of risk factors for atherosclerosis, cardiovascular disease, cardiac infarction, and death (Schmermund et al., 2002) and the affiliated MultiGeneration study. In 1000BRAINS, older adults aged 55 and older (at baseline) from the Heinz Nixdorf Recall study and their relatives (spouses and offspring; sampled from MultiGeneration study) are recruited, measured two times over a period of about 3-4 years. Exclusion from the study was dependent on the eligibility requirements for the MR acquisition based on the MR safety guidelines only (e.g., stents and heart pacemaker led to exclusion from the study). The study protocol was approved by the University of Duisburg-Essen. The initial sample of 1000BRAINS comprised 1,317 participants ranging from 18 to 87 years of age.
For the aim of the current study, we focused on the first time point in both samples. Participants with missing values for the whole neuropsychological and/or brain data were excluded. Furthermore, participants were matched with respect to the age ranges in the two samples. Therefore, we first excluded 666 participants from 1000BRAINS being younger than 64 years of age. Afterward, we matched the two samples for gender and group size by randomly selecting the same number of participants within each  Table 1. Both studies assessed years of formal education as part of a structured anamnestic interview. In addition, all participants filled in a questionnaire concerning their physical and mental well-being (LHAB: SF12; 1000BRAINS: SF36). In both samples, physical and mental health status scores (Ware, Keller, & Kosinski, 1995) were computed using only the SF12 items in order to assure comparability. Furthermore, global cognition was assessed in both samples. While participants from LHAB performed the Mini-Mental State Examination (Folstein, Robins, & Helzer, 1983), participants from 1000BRAINS performed the DemTect in order to estimate a global cognitive status for each participant (Kalbe et al., 2004).  Schmidt and Metzler (1992)). To extract comparable scores from the two vocabulary tests, we calculated the ratio between the total amount of words (MWT_B: 37 words; WST: 40 words) and the amount of correctly identified words. Since the selected neuropsychological tests were not normally distributed, all cognitive tests were first rank-transformed and mean-centered afterward before entering the statistical analysis. For a detailed test description, administration differences between samples and mean values per sample, see Table 2.

| Preprocessing
Anatomical images from both samples were preprocessed using the same automated surface-based processing stream of the FreeSurfer Software package (version 6.0.0). For the LHAB sample, this was done via the FreeSurfer BIDS App (v6.0.0-2; Gorgolewski et al. (2017). A detailed description of this pipeline is provided by Dale, Fischl, and Sereno (1999) as well as on http://surfer.nmr.mgh.
harvard.edu. In short, the surface reconstruction pipeline includes (a) the segmentation of the structural brain images into gray matter, white matter, and cerebrospinal fluid, (b) motion correction, (c) intensity normalization, (d) transformation into Talairach space, (e) tessellation of gray/white matter boundary, and (f ) correction of topological defects. The gray/white matter interface was then (g) expanded to create the pial surface (boundary between gray matter and cerebrospinal fluid), which finally consists of about 150,000 vertices per hemisphere with an average surface area of 0.5 mm 2 .
Afterwards, (h) CT was calculated for each vertex as the shortest distance between the white matter surface and the corresponding vertex on the pial surface. No manual correction of the reconstructed surfaces (white matter, pial surface) was performed in the two studies.
For the purpose of the current study, mean measurements of CT per hemisphere were extracted from FreeSurfer (Fischl and Dale, 2000). Finally, the resulting probability map was thresholded at 95% (using fslmaths, FSL) and binarized.

| Statistical analysis
The purpose of the current study was to compare age-related differences in cognitive abilities and CT in two large independent samples of older adults. Therefore, we first assessed general differ- To test whether trajectories of age-related differences in the different dependent variables (cognitive abilities and CT) are comparable between the two samples, we calculated correlations between age and cognitive abilities and CT (while correcting for gender and education; MAIN) and compared them using Fisher's Z test (Eid, Gollwitzer, & Schmitt, 2011). Finally, in a supplementary analysis, we assessed age-related differences in terms of cognitive performance and CT in a joint analysis (pooled samples), with additionally including "sample" as covariate (for results, see Supporting Information). The reason for this was an additional validation whether the results obtained by the "individual analyses" versus the "joint analysis" would be comparable in the current study.

| RESULTS
When matching the two independent samples for age and gender, the two samples differed in both, demographic variables and cognitive performance. For raw scores and T statistics and Cohen's d, see T = −12.10; p < 0.001; d = 1.08; for detailed information, see Table 2).
When comparing structural brain metrics, we observed higher values for the participants from 1000BRAINS as compared to participants from LHAB, that is, total mean CT for right and left hemi- In the following analyses, the relation between age and cognitive performance and CT, respectively, was assessed using different models (BASE, MAIN, and SENS). With respect to BASE (covariate: gender), we found age-related differences for most of the cognitive tasks (i.e., lower cognitive performance in older adults). Effect sizes, measured using partial eta square were estimated as small to moderate (partial eta square is measured as the proportion of the total variance explained by the independent variable while correcting for the other independent variables, with partial eta square <0.01 is ranked as small; <0.06 as medium and >0.14 as large (Field, 2005;Richardson, 2011 exceptions: verbal fluency and concept shifting did not survive correction for multiple comparisons). Importantly, age-related differences were highly similar in the two samples (see Figure 1; results based on well-being), see Figure 2. Table S1 (see Supporting Information) contains the detailed statistics for the age differences in cognitive performance and for the effects of the covariates of no interest (gender, years of education, mental, and physical well-being).
In the second part of our analysis, we assessed age-related differences in mean CT within left and right hemisphere (Figure 3, for effects sizes, see Figure 4, for statistics, see Table S2, Supporting Information), as well as parts of the DMN (see Figure 5; left and right: anterior DMN, medial posterior DMN, and lateral posterior DMN, for effect sizes, see Figure 6, for statistics, see Figure S3, Supporting Information . With respect to regional differences in the association between CT and age, we found more pronounced age differences in CT for the posterior as compared to the anterior parts of the DMN (Table S3,   Taken together, participants from LHAB seem to show a general superiority in cognitive performance as compared to participants from 1000BRAINS. However, the analysis of age-related differences in cognitive performance and global and regional metrics of CT revealed similar results in both samples.

| DISCUSSION
The present study assessed age-related differences in cognitive abilities (processing speed, concept shifting, reasoning, verbal fluency, and vocabulary) and brain structure (measured by global and regional CT) in two closely matched samples of older adults. Despite significant differences in demographics between the two independent samples, we observed highly similar patterns of age-related differences in both, cognitive abilities and brain structure, when using the same methodological approach.

| Comparability of independent samples of older adults
In times of population aging, there is an increasing interest in assessing risk and protective factors that promote brain and cognitive health until old age. Especially in older adults, however, there is an enormous amount of variability between individuals regarding brain structure and cognitive abilities and the "biological age" does not  However, one has to keep in mind that data pooling across different study populations, might lead to an intermixture of samplespecific biological as well methodological variability which might result in an absence of effects, especially when assessing heterogeneous populations such as older adults. Differences in demographics, methods applied as well as scanner variability have been proposed to be main factors that lead to the heterogeneity of results in terms of brain structure and function in older adults in the field of neuroscience (Afonso et al., 2017;Han et al., 2006;Hanggi et al., 2015;Jancke, Merillat, Liem, & Hanggi, 2015;Kohncke et al., 2016;Liem et al., 2015;Lovden et al., 2017;Trachtenberg et al., 2012  In line with the predictions of the scaffolding theory of cognitive aging (Goh & Park, 2009;Park & Reuter-Lorenz, 2009;Reuter-Lorenz & Park, 2014), higher education as well as engagement in physical activities (which seems to be related to higher physical wellbeing as tested in the current studies; Bize, Johnson, and Plotnikoff (2007)) have repeatedly been shown to protectively influence the neurocognitive aging process. Both have been related to higher cognitive functioning and less brain atrophy during normal as well as pathological aging, such as mild cognitive impairment and Alzheimer's disease (Afonso et al., 2017;Amieva et al., 2014;Miller, Taler, Davidson, & Messier, 2012;Ritchie, Bates, Der, Starr, & Deary, 2013;Schneeweis, Skirbekk, & Winter-Ebmer, 2014;Sofi et al., 2011;Tucker-Drob, Johnson, & Jones, 2009;Zahodne et al., 2011). It is therefore plausible that participants from LHAB showed superior performances in all cognitive tests assessed (processing speed, concept shifting, reasoning, verbal fluency, and vocabulary).
The comparison of CT between the two samples, however, revealed higher global as well as regional CT values for participants from 1000BRAINS. This result seems counterintuitive at first sight.
Based on the sample differences in terms of demographics and cognitive abilities, one would have predicted participants from LHAB to show thicker cortices given that the pertinent literature tends to show positive associations between cognitive ability and the amount of gray matter as measured with CT and gray matter volume or density in the aging population (for an overview, see, e.g., Harada et al., 2013). From our view, the most likely explanation is that these sample differences in CT are due to the different MR scanners used. It has been shown before that even when assessing structural 3D brain images from one and the same person, CT values, but also other metrics, such as brain

| Generalizability of age-related differences in cognitive abilities and CT
Within the scope of the current study, we decided to separately analyze the associations between age and brain structure and cognitive abilities in the two samples and compared the resulting associations using Fisher's Z. Although the two samples differed regarding both, cognitive performance and CT, we revealed highly similar slopes for age-related differences in global as well as regional CT. In line with preceding studies examining global CT, higher age was associated with lower mean CT in both hemispheres for the two samples (Lemaitre et al., 2012;Long et al., 2012;Salat et al., 2004). Similarly, the ageeffect patterns found for cognitive ability did not differ across samples. Higher age was associated with lower cognitive functioning in all cognitive tasks assessed, except in the vocabulary test, where no significant relationship was revealed between age and ability scores.
The similarity of the cross-sectional age-effect patterns that we observe across LHAB and 1000BRAINS indicates that the lower level of education or physical well-being evident in 1000BRAINS does not considerably enhance age differences (i.e., steeper slope in 1000BRAINS sample). Put into the context of cognitive reserve, the between-sample differences in cognitive ability together with the similarity of age slopes, may suggest that participants from LHAB (with a higher education and higher physical well-being) reach the criterion for cognitive impairment later as compared to participants from 1000BRAINS, primarily because they started off at higher levels of cognitive ability. However, by means of the presently used cross-sectional data sets, this proposition cannot To explore in more detail whether the relationship between age and cognitive performance/CT would be differentially influenced by the different covariates (education, physical, and mental well-being) in the two samples, we set up different statistical models (BASE, MAIN, and SENS). Although the different covariates seemed to explain different amounts of variance in the cognitive abilities/CT in the two samples, age-related differences in cognition/CT remained highly similar across samples. For example, mental well-being had a significant influence on processing speed for the sample of 1000BRAINS, but not LHAB. Nevertheless, this difference obviously did not have a considerable impact on the age-related differences in processing speed.
Thus, while education and physical well-being might influence the general level of cognitive performance, it seems that these age-related differences seem to be robust against the possible influences tested in the current samples.

| Regional differences in CT
Beyond assessing mean CT for the two hemispheres, we also analyzed age-related differences in regional CT (different parts of the DMN).
The choice of regions of interest was based on an earlier study of Jockwitz et al. (2017). Herein, the authors aimed at assessing structural correlates for functionally established theories of the aging brain.
In detail, it has been shown that during performance of a memory task (but also in the resting state), older in comparison to younger adults, show stronger activation/connectivity patterns in the more anterior parts of the DMN. At the same time, activation patterns in the more posterior parts of the DMN were reported to be stronger in younger compared to older participants. Thus, with increasing age, there seems to be a shift in brain activation patterns from more posterior to more anterior brain regions (posterior to anterior shift in aging [PASA]) that helps to maintain cognitive performance as stable as possible (Davis, Dennis, Daselaar, Fleck, & Cabeza, 2008;Jones et al., 2011).
In the current study, we exemplarily used the parts of the DMN to assess regional generalizability of age-related differences in brain structure and found age-related decreases in CT for all the posterior parts of the DMN in both samples. In contrast to that, the anterior parts of the DMN did not show age-related differences in any of the two samples. This finding supports a previous study by Jockwitz et al. (2017), in which the authors presented a structural correlate for the posterior to anterior shift in activation patterns, namely a more pro-

| Brain-behavior associations
In the current study, the associations between cognitive abilities and CT were weak and did not survive correction for multiple comparisons. showing only weak associations between brain structure and cognitive performance, especially when examining older adults (e.g., Gunning-Dixon & Raz, 2000;de Mooij, Henson, Waldorp, & Kievit, 2017). This, in turn accords with the scaffolding theory of aging stating that intraindividual regulatory processes (e.g., changes in functional connectivity) within older adults might compensate for structural brain decline thereby keeping cognitive abilities relatively stable (Reuter-Lorenz & Park, 2014). Thus, in the current study, the relation between CT and cognitive abilities were expected to be rather weak. To explore this in more detail, further longitudinal studies are warranted that assess both, structural as well as functional changes in the course of aging in relation to intraindividual changes in cognitive abilities.
Another reason, especially when comparing the current results to the results reported in Jockwitz et al. (2017) for an absence of significant relationships between cognitive abilities and CT could be due differences in structural brain metrics used. The aforementioned study of Jockwitz et al. (2017) used the local gyrification index as measure for cortical atrophy in the regions of interest, measuring the complexity of the brain composed of gray matter and structural connectivity.
The current study used CT as measure for cortical atrophy, since this is one of the most often used brain metrics to study the effects of age were not available for the two samples, might be interesting to investigate in this context for example, episodic memory function.
And finally, larger sample sizes might be necessary to obtain small but significant results, as it has been the case in the aforementioned study of Jockwitz et al. (2017); n = 749.

| Pooled versus individual analyses
In the current study, we decided not to pool data of the two samples but to analyze the samples individually with respect to age-related differences in cognitive performance and brain structure. While the results were highly similar for the cross-sectional age trajectories in terms of CT and cognitive performance, differences were found for the relation between the other covariates included in the models (i.e., SENS) and cognitive performance and CT, respectively. For example, when looking at 1000BRAINS, we found a moderate effect of mental well-being on processing speed. On the other hand, for LHAB and for the pooled sample, there was no effect of mental well-being on processing speed. These distinct outcomes might be the result of differences in sample characteristics. The sample of 1000BRAINS is a population-based sample. In contrast to that, the LHAB study only included participants without any neurological and psychiatric diseases and a score on the Mini-Mental State Examination of at least 26. These sample characteristics might be one explanation why mental well-being plays a significant role in terms of cognitive performance differences in 1000BRAINS but not in LHAB.
Previous studies often assessed age as independent factor in pooled data analyses consisting of older adults (e.g., Fjell et al. (2009)). In the current study, we could show that age revealed the strongest effects on both cognitive performance and CT, and this seems to be highly similar even in independent samples of older adults. Thus, for such robust effects data pooling might be a good option to increase sample sizes and statistical power (Button et al., 2013). However, other risk and protective factors on the aging brain (such as mental well-being) might be study specific, depending on the sample characteristics. Following, when samples are highly heterogeneous, a pooled analysis might underestimate such influences. A combination of both, pooled and individual analyses seem to be an optimal solution to explore influencing factors on the aging brain.

| Limitations and future directions
The study has several advantages as well as limitations which should be addressed. First, the current study investigated CT as one metric of brain structure. CT is a popular and sensitive metric in the frame of age-related differences or changes in gray matter, for example, see Fjell et al. (2009Fjell et al. ( , 2013Fjell et al. ( , 2015; Hogstrom et al., 2013. Given the upcoming trend in data pooling procedures, we thought that CT would therefore be of interest in the current cross-validation study. Nevertheless, in future research, other estimated of gray and white matter as well as functional connectivity should be validated between independent studies, to further evaluate the generalizability of results and advantages and disadvantages of data pooling procedures. Second, with respect to the current study, we decided to match the two samples with respect to age and gender distributions and compare the correlations using Fisher's Z. For the future, we suggest to further evaluate different methodological approaches when cross-validating independent samples with regard to brain metrics and or cognitive functions. First, different matching procedures should be investigated and compared. For example, future studies could not only match samples with regard to age and gender, but also with respect to cognitive functioning using propensity score matching. Furthermore, it would be useful to evaluate other statistical methods to cross validate age-related differences in brain structure and cognitive performance, especially when examining more than two samples. Finally, future studies should explore the importance of covariates. Since the choice of covariates to include into statistical models is highly variable across studies (see Silberzahn et al., 2017), future research should investigate this topic more intensively. For example, the current study assessed education as one indicator for socioeconomic status. Since socioeconomic status includes more than education, for example, occupation and income, future research should also assess other indicators and investigate the influence of these factors on cognition and brain structure.
Moreover, we are aware of the fact that scanner differences might contribute to the differences in sample means in terms of CT in the current study. One way to systematically explore this would be a traveling phantom that can be used to assess scanner differences. The current analyses investigated two independent samples of already completed measurements. Therefore, a retrospective methodical validation was not feasible. However, we would suggest such quality control measurement for future studies with planned study comparisons.
Finally, we have to mention that PASA is just one explanation for the results found in the current study. However, differences in image quality between anterior and posterior parts of the brain might be also responsible for the findings on age-related differences in CT. Future studies should be designed to systematically investigate betweensubject variability across the different regions of the brain, its sources (i.e., measurement quality) and implications for analysis of data resulting from regions with differing variability.

| CONCLUSIONS
Taken together, the current results show that when comparing agerelated differences in cognitive abilities and CT in two different and independent samples within the same age range and composed of the same gender distribution, age-related differences in cognitive performance as well as global and regional CT can be generalized over different samples, assuming the same methodology is used. While data pooling has the advantage to increase statistical power to uncover small effects in the aging population, the current results show the usefulness of conducting separate analyses across samples consisting of distinct study populations, with comparison of the overall trends obtained in each analysis. Future multicenter studies and imaging consortia might at least use a combination of the two approaches to unravel the complexity of the aging brain in its entirety.

ACKNOWLEDGMENTS
The current analysis incorporates data from the Longitudinal Healthy Aging Brain (LHAB) database project carried out at the University of Zurich (UZH). The following researchers at the UZH were involved in the design, set up, maintenance, and support of the LHAB database: