A meta‐analysis of cognitive flexibility in aging: Perspective from functional network and lateralization

Abstract

Cognitive flexibility, the ability to switch between mental processes to generate appropriate behavioral responses, is reduced with typical aging. Previous studies have found that age‐related declines in cognitive flexibility are often accompanied by variations in the activation of multiple regions. However, no meta‐analyses have examined the relationship between cognitive flexibility in aging and age‐related variations in activation within large‐scale networks. Here, we conducted a meta‐analysis employing multilevel kernel density analysis to identify regions with different activity patterns between age groups, and determined how these regions fall into functional networks. We also employed lateralization analysis to explore the spatial distribution of regions exhibiting group differences in activation. The permutation tests based on Monte Carlo simulation were used to determine the significance of the activation and lateralization results. The results showed that cognitive flexibility in aging was associated with both decreased and increased activation in several functional networks. Compared to young adults, older adults exhibited increased activation in the default mode, dorsal attention, ventral attention, and somatomotor networks, while displayed decreased activation in the visual network. Moreover, we found a global‐level left lateralization for regions with decreased activation, but no lateralization for regions with higher activation in older adults. At the network level, the regions with decreased activation were left‐lateralized, while the regions with increased activation showed varying lateralization patterns within different networks. To sum up, we found that networks that support various mental functions contribute to age‐related variations in cognitive flexibility. Additionally, the aging brain exhibited network‐dependent activation and lateralization patterns in response to tasks involving cognitive flexibility. We highlighted that the comprehensive meta‐analysis in this study offered new insights into understanding cognitive flexibility in aging from a network perspective.

The magnetic resonance imaging meta‐analysis was employed to identify functional networks related to cognitive flexibility in aging. Our findings indicate that cognitive flexibility in aging is linked to alterations in the activation and lateralization patterns in several functional networks. CFA, cognitive flexibility in aging; MSC, Monte Carlo simulation.

Practitioner Points.

• Normal aging is associated with a wide range of changes in the activation patterns of functional networks associated with cognitive flexibility in aging.

• Over‐recruited regions exhibited network‐dependent lateralization patterns, whereas under‐recruited regions were left‐lateralized regardless of the network to which they belonged.

1. INTRODUCTION

The proportion of older adults in the general population has been rapidly increased over the past few decades and will continue to grow (Beard et al., 2016). Debilitating effects of brain aging, in particular cognitive declines, have come to the forefront. Notably, cognitive flexibility, a core component of cognitive control, shows a notable decline with age (Ferguson et al., 2021; Giller & Beste, 2019; Ging‐Jehli & Ratcliff, 2020; Grady, 2012; Li et al., 2023). Cognitive flexibility refers to the ability to ensure the goal‐directed behavior by coordinating and overseeing the flexible allocation of mental resources (Braem & Egner, 2018; Dajani & Uddin, 2015; Medaglia et al., 2018). Typically, this ability can be examined using task‐switching (e.g., letter‐number switch tasks) and set‐shifting paradigms (e.g., Wisconsin Card Sorting Task, see Diamond, 2013). Such flexibility is crucial for adapting to the ever‐changing situations in daily life. Therefore, declines in cognitive flexibility can profoundly affect the quality of life of older adults, with challenges in daily activities (e.g., work, family responsibilities) and a reduced capacity for self‐care (Davis et al., 2010). Knowledge of the neurological processes involved in age‐related variations in cognitive flexibility is necessary to facilitate the development of interventions for brain aging (Williams & Kemper, 2010).

2. REGIONAL PERSPECTIVES

The population aged 60 and above demonstrates significantly lower cognitive flexibility compared to young adults aged 18–35 (Craik & Bialystok, 2006; Ferguson et al., 2021), and this decline in cognitive flexibility involves functional changes in regions that support various mental functions. For example, compared to young adults, older adults exhibit higher activation in the dorsolateral prefrontal cortex (dlPFC) and superior parietal lobule, as well as lower activation in the insula and occipital regions (Nashiro et al., 2018; Spreng et al., 2017). These age‐related differences in regional activation have been interpreted by several theories in the field of neurocognitive aging. For example, the posterior–anterior shift in aging (PASA) theory suggests that older adults activate more anterior prefrontal regions to compensate for visual declines in posterior occipital regions (Davis et al., 2008; Festini et al., 2018). Furthermore, the compensation‐related utilization of neural circuits hypothesis (CRUNCH) proposes that older adults recruit more neural resources to meet task requirements at a lower cognitive load. This strategy, however, fails under high‐demand conditions, resulting in reduced regional activation (Reuter‐Lorenz & Cappell, 2008; Reuter‐Lorenz & Lustig, 2005). Overall, identifying age‐related differences in regional activation has contributed to our understanding of cognitive flexibility in aging.

3. NETWORK PERSPECTIVES

Accumulating evidence suggests that cognitive flexibility is not solely tied to specific regions, but is also implemented by interrelated regions nested in large‐scale networks (Chen et al., 2018; Dajani et al., 2020; Dajani & Uddin, 2015; Qiao et al., 2017; Qiao et al., 2020; Uddin, 2021; Xia et al., 2022). Although empirical studies have revealed a range of regions associated with cognitive flexibility in aging (Campbell et al., 2012; Ferreira & Busatto, 2013; Grady, 2012; Spreng et al., 2017), no meta‐analysis has examined how these regions are networked and what activation patterns they exhibit within their functional networks. To address this gap, a network‐based meta‐analysis is crucial. In order to make a refined distinction between networks, we followed the Yeo 7‐template to define functional networks (Yeo et al., 2011).

3.1. Task positive networks

The Yeo 7‐template defined three task positive networks associating with high‐order cognitive processes: the dorsal attention network, ventral attention network, and frontoparietal network (Cocchi et al., 2013; Di & Biswal, 2014; Menon & D'Esposito, 2022; Suo et al., 2021). Older adults display both decreased and increased activation in the hub regions of task positive networks. For example, compared to younger adults (aged 18–40), older adults (aged 63–83) exhibit lower activation in the anterior insula within the ventral attention network, while higher activation in the superior frontal gyrus (SFG) within the dorsal attention network and the dlPFC within the frontoparietal network (Nashiro et al., 2018; Zhu et al., 2014), indicating age‐related changes in attention and top‐down control. Furthermore, the CRUNCH suggests that older adults exhibit increased brain activity as a compensatory response when the cognitive load is low, but display decreased activation when the task is challenging. Given that tasks used in different empirical studies vary in difficulty, it is likely that older adults exhibit both decreased and increased activation in task positive networks.

3.2. Networks involved in lower‐order cognitive processes

Cognitive flexibility in aging may also be partly due to alterations in networks involved in lower‐order cognitive processes, such as the visual and somatomotor network (Cocchi et al., 2013; Egner & Hirsch, 2005; Konstantinou et al., 2014; Levy & Wagner, 2011). For example, during tasks measure cognitive flexibility, compare to young adults (aged 18–35), older adults (aged 60–80) exhibit lower activation in the posterior occipital regions within the visual network, and higher activation in the precentral cortex within the somatomotor network (Berry et al., 2016; Eich et al., 2016; Nashiro et al., 2018), which corresponds to a decrease in visual search performance and age‐related alterations in motor control ability (Callaghan et al., 2017; Seidler et al., 2010). Furthermore, the PASA proposes that older adults display decreased activation in the hub regions (i.e., posterior occipital regions) within the visual network. Therefore, a network meta‐analysis may uncover age‐related differences in the activation of the visual and somatomotor networks during cognitive flexibility tasks.

3.3. Default mode network

Enhanced activation of the default mode network putatively engages in the aging of cognitive flexibility. The default mode network is thought to support internally directed cognition and shows decreased activation to reduced interference of endogenous processes during cognitive control tasks (Anticevic et al., 2012; Buckner & DiNicola, 2019; Raichle, 2015). However, compared to young adults, older adults exhibit higher activation in regions (medial prefrontal and posterior cingulate cortex) within the default mode network (Qin & Basak, 2020; Sambataro et al., 2010). Older adults typical show a decrease in the selectivity and specialization of regional activity for specific task demands (Koen et al., 2020; Koen & Rugg, 2019; Nashiro et al., 2018; Rakesh et al., 2020). Therefore, older adults may recruit regions not related to the task at hand, potentially disrupting the intended cognitive processes. A network meta‐analysis may further reveal this increased activity in the default mode network in older adults while performing cognitive flexibility tasks.

3.4. Network lateralization

Lateralization could be observed in regions associated with cognitive flexibility in aging. The hemispheric asymmetry reduction in older adults (HAROLD) model suggests that older adults display less activation asymmetry during cognitive performance compared to young adults (Cabeza, 2002; Festini et al., 2018). This theory is based on age‐related differences in activation patterns during memory, inhibition, and perceptual tasks. However, whether regions associated with the aging of cognitive flexibility display the asymmetry reduction lacks examination. In fact, age‐related asymmetry reduction proposed by HAROLD could be evident in cognitive flexibility. Successful cognitive switching is typically associated with functional domination in the left hemisphere (Capizzi et al., 2016; Serrien & Sovijarvi‐Spape, 2013). However, less activation in the left hemisphere was associated with age‐related declines in cognitive flexibility (Yeung et al., 2016). Moreover, a meta‐analytic study reported age‐related increases in brain activation in the right hemisphere during cognitive flexibility. Specifically, most of the regions with increased activation in older adults (five of six) were located in frontal regions of the right hemisphere, including the middle, inferior, and superior frontal regions (Spreng et al., 2017). Taken together, regions associated with cognitive flexibility in aging may partially exhibit lateralization, including less activation in the left hemisphere, as well as more activation in the right hemisphere. Moreover, age‐related differences in lateralization can be related to the activation pattern (e.g., increased or decreased activation) and restrict to specific networks.

4. THE PRESENT STUDY

Cognitive flexibility in aging encompasses a range of age‐related differences in cognitive processes. These differences include visual/motor processing, higher‐order cognitive processing (e.g., attention and top‐down control), and the suppression of internal thoughts (Kiesel et al., 2010; Koch et al., 2018; Uddin, 2021). Although previous magnetic resonance imaging (MRI) meta‐analyses have explored the regions associated with cognitive flexibility in aging (Spreng et al., 2017), critical questions remain unresolved. Specifically, it remains unclear whether and how these age‐related differences in cognitive processes manifest in the activation patterns of functional networks. Based on the general theories of neurocognitive aging (e.g., dedifferentiation, CRUNCH, and HAROLD), both age‐related changes in brain activation, as well as lateralization, should be examined within a network framework.

In the current study, we employed the multilevel kernel density analysis (MKDA) to identify regions exhibiting stable group differences during cognitive flexibility tasks. We first identified regions with decreased activation (Young > Older) or increased activation (Older > Young), and subsequently mapped them into the well‐defined Yeo‐7 network template (Yeo et al., 2011). Furthermore, we performed lateralization analyses on regions with group differences in activation, both before and after their projection into the Yeo 7‐template, which enabled us to determine the spatial lateralization at the global and network levels, respectively. We made the following predictions: (1) cognitive flexibility in aging may be associated with a wide range of functional networks, including the dorsal attention, ventral attention, frontoparietal, visual, somatomotor, and default mode networks; (2) functional networks involved in cognitive flexibility in aging may show different activation patterns (increased or decreased activation); and (3) lateralization could be observed in regions associated with cognitive flexibility in aging, but lateralization patterns may be varied in different networks.

5. METHODS

5.1. Literature search

An online search of EBSCOHost (PsycINFO, PsycARTICLES), PubMed, and Web of Science databases was performed to identify pertinent articles published before August 2024. The keywords related to older adults were “older adults, aging, age‐related, elderly, elders, ageing, adult life‐span, older people, older persons.” The keywords related to functional MRI (fMRI) included “Neuroimaging, fMRI, Functional magnetic resonance imaging, positron emission tomography, and PET.” The keywords related to cognitive flexibility were “multitasking, discrimination task, Wisconsin card sorting test, task‐switching, set‐shifting, cognitive flexibility.” We employed a Boolean logic approach to conduct the search on these keywords (see, supplementary material for details). To identify additional eligible studies, we also conducted a literature search using the reference lists of identified studies and several relevant reviews (Heckner et al., 2020; Li et al., 2015; Spreng et al., 2010; Spreng et al., 2017).

5.2. Eligibility

The inclusion criteria of studies were as follows: (1) Empirical studies should classify healthy participants into younger and older groups based on specific criteria. (2) The studies were task‐based. (3) Studies should report three‐dimensional Montreal Neurological Institute (MNI), or Talairach coordinates. (4) The results passed the statistical thresholds of p < .001 (uncorrected) or p < .05 (corrected). (5) Studies should perform the whole‐brain analysis. (6) Studies that focused on both global and local switch costs were included. (7) The results in studies were based on comparisons between young and older adults. The exclusion criteria were as follows: (1) Coordinates of peaks were not reported. (2) Articles were not written in English. (3) Study used imaging methods other than fMRI and positron emission tomography, such as single photon emission computed tomography or functional near‐infrared spectroscopy. (4) Results based on brain–behavior correlations. (5) Studies that did not set control conditions when performing the activation analysis. (6) Participants were not healthy young or older adults. The papers were evaluated after the keywords were agreed on two researchers, and the software Endnote was used to organize the data into Excel. We read each article and extracting the data to summarize the published results.

5.3. Quantitative meta‐analysis procedure

A modified MKDA was used to perform the meta‐analysis. MKDA was initially proposed to detect robust effects across studies by performing spatial convolution (Lindquist et al., 2016; Schurz et al., 2014; Wager et al., 2007; Wager et al., 2009). This method has two important features. First, MKDA can integrate information carried by spatially adjacent voxels. Second, MKDA exhibits relative stability when applied to small to medium sample sizes. Specifically, MKDA aggregates consistency of activation across studies rather than across peak coordinates, which avoid single study that reports a lot of nearby peaks excessively influence the final results (Wager et al., 2007; Wager et al., 2009). Therefore, the MKDA is suitable for meta‐analyses with smaller sample sizes. Notably, the choice of the kernel size for MKDA is typically habit‐based, and there is no quantitative method for MKDA to determine the kernel size with optimal convolution efficiency.

In the current study, the modified MKDA was proposed to determine the kernel size with optimal convolution efficiency automatically (see, Supplementary Materials, Figure S1). This method can search for the optimal kernel size based on the statistical power within a given range. A self‐written toolbox and functions in the CanlabCore Toolbox (https://github.com/canlab/Canlab_MKDA_MetaAnalysis) were used for the analysis. The Talairach coordinates were first transformed into the MNI coordinates (Lancaster et al., 2007). Subsequently, spherical kernels with a gradual change in size (r = 8–15 mm, step = 0.2 mm) were generated to convolve peak voxels. We performed the following analysis for each kernel size: (1) Spatial convolution. The peak coordinates were convolved using a spherical kernel. If there were active peak points in the convolution kernel, all voxels within the spherical kernel were assigned a value of 1 to obtain the indicator map for each study. (2) Weighting indicator map. Each indicator map was weighted by the sample size and statistical method (fixed vs. random effect). (3) Calculation of the density map. The density map was yielded by averaging over the weighted indicator maps. (4) Permutations test for family‐wise error correction (cluster corrected, p < .05). We employed the Monte Carlo simulation (MCS) to conduct 5000 permutations within an EPI mask. The MCS generated random distributions of activation probabilities for each voxel. By comparing the actual density maps to these random distributions, we identified clusters that showed significantly higher activation probabilities than random cases. Steps 1–4 were repeated to obtain the kernel size with the highest statistical power and regions with activation differences between age groups.

5.4. Mapping regions into functional networks

Regions with group differences in activation were identified and assigned to the widely used Yeo 7‐network parcellation (Yeo et al., 2011). The parcellation was identified using a data‐driven approach and encompassed seven networks with specialized functions, including the frontoparietal, dorsal attention, ventral attention, somatomotor, visual, default mode, and limbic networks. The limbic network primarily comprises subcortical structures that meet the cerebral cortex and modulate motivation, emotion, and instinctive behaviors. Given that the Yeo 7‐network parcellation template does not encompass subcortical areas, we also utilized the Tian template to examine age‐related changes in subcortical areas (Tian et al., 2020).

5.5. Lateralization analysis

5.5.1. Asymmetry index

The asymmetry index (AI) was calculated to determine whether there was a lateralization effect for regions with activation differences at the global and network levels. The AI values for regions with increased and decreased activation were calculated separately. A positive AI value indicates left lateralization, whereas a negative value suggests right lateralization. Generally, a challenge for the lateralization analysis in the meta‐analysis was the lack of approaches to test the statistical significance. Inspired by the MKDA (Wager et al., 2007; Wager et al., 2009), we adopted the MCS to randomly generate samples to obtain the natural distribution of the AI, which allowed the examination of the significance of the AI. Cluster‐level MCS with 5000 repetitions was used to generate the $\mathit{AI}\_p_i$ to determine the significance of each AI (see Supplementary Materials, Figure S1). Several equivalence principles were followed to ensure that the features other than the spatial location of the generated clusters matched the real situation (see Supplementary Materials).

$${\mathit{AI}}_i=\frac{v_{\mathit{il}}-v_{\mathit{ir}}}{v_i}$$

where v is the volume of the regions with group differences in activation. i indicates the spatial scope of regions, range 0–7; 0 represents regions throughout the whole brain, whereas i in the range of 1–7 corresponds to regions within each of the seven subnetworks. l and r indicate the spatial scopes of the left and right hemispheres, respectively. AI _i ∈ [−1, 1]. This formula has been used in several studies evaluating global asymmetry (Kong et al., 2018; Okada et al., 2016; Wyciszkiewicz & Pawlak, 2014), and we migrated it to the lateralization analysis at the network level.

$$I_{i,b}=\left\{\begin{array}{c}0,\left|{\mathit{AI}}_{i,b}\right|<\left|{\mathit{AI}}_i\right|\\ 1,\left|{\mathit{AI}}_{i,b}\right|\ge \left|{\mathit{AI}}_i\right|\end{array}\right.$$

where ${\mathit{AI}}_i$ is calculated from the real distribution of regions, ${\mathit{AI}}_{i,b}$ is calculated from the random distribution generated by the MCS in the iteration of b. The value of I _i,b is depends on the judgment formula in each iteration. b ∈ [1, 5000].

$$\mathit{AI}\_p_i=\frac{\sum _{b=1}^B{I}_{b,i}}{B}$$

B is 5000, and is the total number of repetitions in the MCS.

5.5.2. Delta AI

The delta AI (∆AI) is an indicator used to describe the lateralization effect of a specific network. Specifically, a positive value denotes that the current network contributes right lateralization at a global level. Without the contribution of a network with a positive ∆AI, the global lateralization would be less right‐oriented. Furthermore, the degree of lateralization of regions within a specific network is positively correlated with its contribution to the global lateralization. The MCS with 5000 repetitions was performed to determine the significance of ∆AI.

$$∆{\mathit{AI}}_i=\frac{\sum _{b=1}^B{\mathit{rAI}}_{i,b}}{B}-{\mathit{AI}}_0$$

where i is the label of the network and indicates that the regions within network i will be randomized. The ${\mathit{rAI}}_{i,b}$ is the random AI calculated by the spatial distribution with randomizing regions in network i in the iteration b. AI₀ is the true value of the global AI. $\Delta$ AI _i ∈ [−1, 1].

$$∆I_{i,b}=\left\{\begin{array}{c}0,\left|r{\mathit{AI}}_{i,b}\right|<\left|{\mathit{AI}}_0\right|\\ 1,\left|{\mathit{rAI}}_{i,b}\right|\ge \left|{\mathit{AI}}_0\right|\end{array}\right.$$

$∆I_{i,b}$ will be recorded as 1 if it meets the judgment formula in each iteration.

$$∆\mathit{AI}\_p_i=\frac{\sum _{b=1}^B∆I_{i,b}}{B}$$

B is the total number of repetitions in the MCS and is 5000 in the current study.

5.5.3. Comparison of AI and delta AI

Two cross‐corroborating parameters, the AI and delta AI, were used to examine the reliability of the lateralization results. The computation of the AI was restricted within a specific network and the spatial locations of all brain regions within that network were randomized in each MCS iteration; when calculating the delta AI, the randomization process was limited to a specific network, but the calculation of parameters was conducted incorporating all age‐related regions. These two parameters examined the pattern of lateralization within a particular network (e.g., AI) and the contribution of the particular network to global laterality (e.g., ∆AI), providing a comprehensive depiction of the lateralization pattern.

5.6. Quality control

5.6.1. Quality control for eligible empirical studies

We employed a checklist developed by Kmet et al. (2004), specifically designed for assessing the quality of empirical studies. This checklist encompasses various aspects of a study, including the research question, study design, sampling strategy, data analysis, reliability of results, and credibility of conclusions. Two independent reviewers (XHS and LHY) evaluated the quality of each included study. Each item in the checklist was scored based on the degree to which specific criteria were met (“yes” = 2, “partial” = 1, “no” = 0). Prior to the evaluation, the two reviewers discussed the criteria for each item. For instance, for the item “Controlled for confounding?,” a maximum score (“yes” = 2) would only be given if quality control was implemented in both the MRI and behavioral analysis. Based on the total score, each study was classified as “low” (0–13), “moderate” (14–17), or “high” (18–22) quality.

5.6.2. Quality control for quantitative analysis

To verify the reliability of the results, we performed the following analyses. First, we conducted a series of additional meta‐analyses using the leave‐one‐study‐out validation (LOOCV) approach to assess whether our findings were independent of the exclusion of any particular study (Etkin & Wager, 2007; Sha et al., 2019). We then compared the similarity and difference between the initial results and LOOCV‐generated results. Second, this study primarily included two types of cognitive flexibility tasks: the task‐switching and set‐shifting tasks. To evaluate whether the task type influenced our main findings, we carried out subgroup analyses for each type of tasks. Third, to evaluate the impact of kernel size variations on results, we generated activation maps under kernel sizes ranging from 8 to 15 mm (step = 0.2 mm) for both the young > older and older > young contrasts. Then, we compared the similarity (using Pearson correlation analysis) and differences (using paired t test, p < .05, FDR‐corrected) between the activation maps with optimal kernel sizes and the activation maps with other kernel sizes.

6. RESULTS

6.1. Search results

The results of the reference search and exclusion are shown in Figure 1. Then, 13 articles were included in the current meta‐analysis, with a total of 681 participants (316 young adults and 365 older adults) and 231 peak voxels. In each of the included studies, the mean age of the young group was below 35, with the minimum age of individual subjects being over 18 years old. The mean age of the older group was above 60 in the included studies, except for the study by Kuptsova et al. (2016), which reported an older age group range of 50–65 years. For the experimental tasks, more than half of the studies (N = 7) used the task‐switch paradigm. The other tasks were set‐shifting paradigm, including the card‐sorting, decision‐making, multitasking, and attention shift tasks. All tasks required subjects to adopt switch strategies according to changing rules during task performance. A study recruited two subtypes (e.g., monolingual, and bilingual subjects) of young and older participants separately (Gold et al., 2013). However, Gold et al. (2013) combined these two subtypes when identifying brain regions with differing activations between age groups. Consequently, the findings from this study were treated as a contrast in the subsequent meta‐analysis. All studies conducted between‐group comparisons of activation during the performance of cognitive flexibility tasks. The other information (e.g., year of publication) of the included studies is summarized in Table 1.

FIGURE 1.

Flowchart of the study selection process.

TABLE 1.

Characteristics of studies included in the meta‐analysis.

Article	Task	Group comparison	Task contrast	Number of foci	YA		OA
					N	Age (M ± SD)	N	Age (M ± SD)
Eich et al. (2016)	Task switch	YA vs. OA	Switch vs. nonswitch	28	62	25.8	75	64.8
Gold et al. (2013)	Task switch	YA vs. OA	Switch vs. nonswitch	5	20	31.6 ± 4.3	20	63.9 ± 4.0
					20	32.3 ± 3.3	20	64.4 ± 5.1
Townsend et al. (2006)	Attention shift task	YA vs. OA	Shift attention vs. baseline	23	10	27.9 ± 8	10	70.7 ± 7
Kunimi et al. (2016)	Task switch	YA vs. OA	Switch vs. repeat	26	20	23.9 ± 5.4	20	67.4 ± 4.3
Kuptsova et al. (2016)	Task switch	YA vs. OA	Attention switch vs. control	29	39	20–30	40	51–65
Madden et al. (2010)	Task switch	YA vs. OA	Switch vs. repeat	35	20	22.4 ± 2.5	20	69.6 ± 6.1
Martins et al. (2012)	WCST	YA vs. OA	Set‐shift vs. non‐shift	7	14	26 ± 5	10	62 ± 8
Nagahama et al. (1997)	Card sorting test	YA vs. OA	Task vs. control	22	6	22.2	6	67.7
Nashiro et al. (2018)	Multitasking task	YA vs. OA	Task vs. fixation	12	27	25.1 ± 3.1	40	65.9 ± 6.3
Van Impe et al. (2011)	Multitasking task	YA vs. OA	Dual vs. single	19	20	25.2 ± 3.0	20	68.9 ± 4.2
Worthy et al. (2016)	Decision‐making task	YA vs. OA	State‐change vs. baseline	3	18	23.6	18	67
Zhu et al. (2014)	Task switch	YA vs. OA	Switch vs. nonswitch	2	32	32.1 ± 3.6	33	68.4 ± 5.4
Zhu et al. (2015)	Task switch	YA vs. OA	Switch vs. nonswitch	18	28	32 ± 3.8	33	68.4 ± 5.4

Abbreviations: OA, older adults; YA, young adults.

6.2. Network‐based results of the modified MKDA

The MKDA revealed that older adults exhibited both increases and decreases in brain activation compared to young adults (Table 2, Figure 2). The optimal kernel sizes were 13.6 mm for the regions with increased activation and 10.0 mm for the regions with decreased activation, indicating that the spatial distribution of the over‐recruited regions was relatively discrete and that of the regions exhibiting lower activation in older adults was relatively clustered. Regions with increased activation exhibited higher cross‐study robustness (weight range: 0.13–0.53) relative to the regions with decreased activation (.05–0.33). The proportions of regions with increased activation (55.11%) and decreased activation (44.89%) to the total were approximately equal. Compared to young adults, older adults showed higher activation in the dlPFC, SFG, inferior frontal gyrus, insula, and precentral cortex in the right hemisphere. Moreover, older adults showed decreased activation in posterior cingulate cortex, middle temporal gyrus, angular gyrus, precuneus, insula, and inferior parietal lobe in the left hemisphere. Most of the regions with decreased activation were distributed in the occipital lobe, involving the bilateral lingual, right cuneus, right declive, and left inferior occipital gyrus.

TABLE 2.

Comparisons between older and young adults.

Cluster	Voxel numbers	MNI coordinate			Weight	BA	Anatomical label
		x	y	z
Activation: OA > YA
1	1354	−37	−76	37	0.39	39	L AG
		−8	−60	21	0.33	23	L PCC
		−7	−62	24	0.33	32	L precuneus
		−41	−73	30	0.32	39	L MTG
2	1082	27	−5	63	0.47	6	R dlPFC
		28	−5	68	0.47	6	R precentral
		26	−2	68	0.47	6	R SFG
3	712	−60	−34	21	0.22	13	L insula
		−61	−29	22	0.22	40	L IPL
4	639	40	24	−8	0.53	47	R IFG
		40	18	−3	0.42	/	R insula
Activation: YA > OA
1	2299	−2	−94	−8	0.33	18	L lingual
		−12	−95	−5	0.32	17	L IOG
2	1088	8	−75	25	0.26	18	R cuneus
		11	−89	−14	0.26	/	R declive
		22	−67	6	0.22	19	R lingual

Abbreviations: AG, angular gyrus; dlPFC, dorsolateral prefrontal cortex; IFG, inferior frontal gyrus; IOG, inferior occipital gyrus; IPL, inferior parietal lobule; L, left; MNI, Montreal Neurological Institute; MTG, middle temporal gyrus; PCC, posterior cingulate cortex; R, right; SFG, superior frontal gyrus.

FIGURE 2.

Regions show lower (cold) and higher activations (hot) in older adults. Clusters were displayed using a threshold of p < .05 (cluster‐level, false discovery rate correction). L, left hemisphere; R, right hemisphere.

We calculated the distribution of the identified regions within each network to determine networks that were correlated with cognitive flexibility in aging. The results showed that these identified regions were distributed in six functional networks, including the somatomotor, visual, frontoparietal, dorsal attention, ventral attention, and default mode networks. No identified regions fell into the limbic network and subcortex areas (see, Figure S2). The default mode network (28.80%), dorsal attention network (23.25%), somatomotor network (19.64%), and ventral attention network (14.87%) accounted for 86.56% of regions with increased activation (Figure 3, red bars), whereas the visual network accounted for 82.8% of regions with decreased activation (Figure 3, blue bars). The frontoparietal network contained only a small portion of regions with group differences in activation. Thereafter, we calculated the percentage of regions within each network (Figure 3, gray line). The results showed that the visual network predominantly exhibited reduced activation, whereas all regions within the somatomotor network displayed increased activation in older adults. The dorsal attention network mainly had regions with increased activation, and all regions within the ventral attention network displayed increased activation. The frontoparietal network contained regions with both increased and reduced activation, but was not dominated by either type of activation. The default mode network showed increased activation, with only a small number of regions (9.43%) exhibiting reduced activation.

FIGURE 3.

Mapping regions to the Yeo 7‐network template. The bars (left y‐axis) represent the proportion of regions with increased (Older > Young) and decreased activation (Young > Older) within each network. For example, the red bars represent the relative volume of regions with increased activation within a specific network relative to the total volume of over‐recruited regions. The gray line (right y‐axis) represents the proportion of over‐recruited region within each network. The higher the value of the gray line, the more evident the higher activation of the corresponding network. In calculating the values of the gray line, the numerator consists of the volume of over‐recruited regions, while the denominator contains the combined volumes of both types of regions within an individual network. DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; SMN, somatomotor network; VAN, ventral attention network; VN, visual network.

In summary, age‐related differences in activation patterns were complex and network‐dependent, aligning with the neurocognitive theories outlined in the introduction. Specifically, we observed both increased and decreased activation in task positive networks, a mixed activation pattern that aligns with the CRUNCH. Moreover, decreased activation in the visual network suggested a decrease in visual ability in older adults as proposed by the PASA. Finally, increased activation in the default mode network supported our hypothesis, suggesting that older adults may inappropriately recruit task‐irrelevant regions during cognitive flexibility tasks.

6.3. Lateralization results

The global AI was significant for regions showing reduced activation (AI = 0.23, p < .001), but not for regions exhibiting increased activation in older adults (Table 3, Figure S3). However, the network‐level results showed that the over‐recruited regions within the dorsal attention and somatomotor networks were significantly right‐lateralized. Furthermore, the over‐recruited regions within the default mode, ventral attention, and visual networks were left‐lateralized (Table 3, Figure S4). Within the visual network, regions exhibiting lower activation in older adults were left‐lateralized (AI = 0.26, p < .001). A few regions within other three networks (e.g., dorsal attention, frontoparietal, and default mode networks) that showed reduced activation were also all left‐lateralized (Table 3, Figure S5). Taken together, the lateralization patterns for regions with increased activation were varied within different networks but were left for regions with decreased activation.

TABLE 3.

The lateralization at global and network levels.

Network	Increased activation		Decreased activation
	AI	∆AI	AI	∆AI
Global	0.04	/	0.23***	/
VN	0.94***	−0.03***	0.26***	−0.12***
SMN	−0.54***	0.06***	/	/
VAN	0.25**	−0.02**	/	/
DAN	−0.30**	0.03**	0.64***	−0.02***
FPN	−0.21	−0.00	0.58***	−0.02***
DMN	0.67***	−0.07***	0.52***	−0.01

Abbreviations: ∆AI, the contribution of the particular network to global laterality; AI, asymmetry index; DAN, dorsal attention network; DMN, default mode network; FPN, frontoparietal network; SMN, somatomotor network; VAN, ventral attention network; VN, visual network.

^*** p < .001.

^** p < .01.

Furthermore, the ∆AI was calculated to measure how regions within a specific network contribute to the global‐level lateralization. In other words, the ∆AI measures the changes in global AI after randomizing the spatial location of the regions within a target network. For regions with increased activation (Figure 4 and Table 3), the ∆AI of the somatomotor and dorsal attention networks were positive and significant, which indicated that regions with increased activation within the two networks contributed to rightward lateralization at the global level. However, the ∆AI of the default mode, ventral attention, and visual networks were negative, indicating that over‐recruited regions within these networks contribute to left lateralization at the global level. The default mode, ventral attention, and visual networks contribute to left lateralization, while the dorsal attention and somatomotor networks play roles in the maintenance of right lateralization. The mixed lateralization patterns of these networks masked the detection of global‐level lateralization for regions with increased activation. For regions with decreased activation (Figure 4 and Table 3), the ∆AI of the visual, frontoparietal, and dorsal attention networks were negative, which indicated that the three networks have a left contribution at the global level. Notably, the absolute value of the ∆AI in the visual network was much larger than that of the dorsal attention and frontoparietal networks, indicating that the visual network was the most important network in the maintenance of left lateralization.

FIGURE 4.

The $∆\mathit{AI}$ and distribution of random $\mathit{AI}$ generated by the MCS. The left and right panels indicate the increased (Older > Young) and decreased (Young > Older) conditions, respectively. The violin plot shows the distribution of global $\mathit{AI}$ after randomizing the spatial locations of regions within a network. The box plot shows the median (centerline in box), interquartile ranges (top and bottom of the box, 25–75%), and total ranges (min–max). The horizontal lines indicate the true global AI. The $∆\mathit{AI}$ for each network is the value of the horizontal line minus centerline in the box. ***p < .001, **p < .01, *p < .05. AI = Asymmetry Index. ∆AI = the contribution of the particular network to global laterality (see Figure 3 for network abbreviations).

The regions with decreased activation within the dorsal attention network were left‐lateralized, whereas regions with increased activation were right‐lateralized. Thus, the left–right shift index (LRS) was proposed to reflect the extent to which the network exhibits a low‐left and high‐right activation pattern simultaneously (Supplementary Materials). The significance of the LRS was examined using MCS with 5000 iterations. The result (Figure S6) showed that the LRS of the dorsal attention network was negative (LRS = −0.47, p < .001) and significantly lower than the random cases generated by MCS, indicating a LRS of activation patterns in the dorsal attention network. Taken together, in accordance with our prediction and the HAROLD model, older adults exhibited network‐dependent lateralization in regions associated with cognitive flexibility in aging.

6.4. Quality control

6.4.1. Quality ratings of eligible empirical studies

Overall, the quality of included studies was high, with 12 studies rated as “high” quality, and 1 study rated as “moderate” quality (Table S1). The study conducted by Nagahama et al. (1997), assigned as “moderate” quality, principally encountered deficiencies concerning sample size (under 10), the analytical quality control, and the sufficiency of result reporting. Nonetheless, the LOOCV results in the subsequent section suggested that the presence of a single study with moderate quality did not significantly impact the overall results.

6.4.2. Impact of a single study on overall results

The impact of a single study on the overall results was investigated. No study had a disproportionate effect on the original results (see Supplementary Materials, Table S2). All density maps created using LOOCV showed high positive correlations with the original results (range of correlation coefficients: 0.89–0.99), and the overlap ratios between the original and LOOCV maps ranged from 0.88 to 0.99. A one‐sample t test revealed no significant difference between the original and LOOCV‐generated maps.

6.4.3. Subgroup analysis

We conducted subgroup analyses to explore whether task types influence age‐related differences in brain activation. A two‐sample t test was performed to examine the between‐group differences (task‐switching vs. set‐shifting) in density maps. We found no differences in activation patterns under either relatively strict (FDR corrected, p < .05) or lenient (uncorrected, p < .001) statistical thresholds. Furthermore, a followed correlation analysis on the group‐level activation maps revealed similar activation patterns in both types of tasks (r = .46, p < .001).

6.4.4. Results of the sensitivity analysis on kernel size

As the kernel size changes, the activation maps change smoothly (Figure S7). Thus, activation maps corresponding to similarly sized kernels exhibited high similarity. Given the optimal kernel sizes for increased activation (10 mm) and decreased activation (13.6 mm) differed, we calculated the activation maps with a middle kernel size of 11.8 mm. The optimal and middle activation maps were highly similar in both the regions with decreased (r = .80, p < .001) and increased activation (r = .88, p < .001). Furthermore, there was no significant difference in activation patterns between the optimal and middle activation maps in either condition. This result suggested that the optimal activation maps in both decreased and increased activation conditions exhibited substantial comparability.

7. DISCUSSION

The results were summarized in Table 4 and verified our predictions. First, the current meta‐analysis revealed age‐related differences in the activation of the dorsal attention, ventral attention, frontoparietal, somatomotor, visual, and default mode networks. This indicates that cognitive flexibility declines in older adults are linked to functional changes across several cognitive processes, encompassing both high‐ and lower‐order functions. Second, age‐related differences in activation patterns were network‐dependent. For example, networks can exhibit predominantly decreased activation (i.e., the visual network) or increased activation (i.e., the default mode, ventral attention, dorsal attention, and somatomotor networks). Third, network‐dependent lateralization was found in regions associated with cognitive flexibility in aging.

TABLE 4.

Summarized findings in the current study.

Network	Activation type	Lateralization patterns
		Regions with increased activation	Regions with decreased activation
Global	Mixed activation pattern	No lateralization	Left
VN	Primarily decreased activation a	Left	Left
SMN	Pure increased activation b	Right	/
VAN	Pure increased activation	Left	/
DAN c	Primarily increased activation a	Right	Left
FPN	Mixed activation pattern	No lateralization	Left
DMN	Primarily increased activation	Left	Left

^a Primarily increased or decreased activation indicates that this type of regions within the network accounts for more than 80% of the total volume.

^b Pure increased activation indicates that regions within the networks were all over‐recruitment in older adults. The SMN and VAN do not contain regions with decreased activation, and the corresponding AI was not calculated.

^c The left–right shift index (LRS) has confirmed the low‐left and high‐right activation pattern within the DAN (see Table 3 for abbreviations).

Our study yielded results similar to published findings, and offering new insights into cognitive flexibility in aging. A previous meta‐analysis on cognitive flexibility in aging reported both increased and decreased activation in the superior, middle, and inferior frontal gyrus (Spreng et al., 2017), which aligns with our findings in the frontoparietal network. Furthermore, the previous study found that five of the seven regions with decreased activation were located in the left hemisphere. Our study confirmed significant left lateralization for regions with decreased activation. To our knowledge, this study was the first to provide meta‐analytic evidence from a network perspective about cognitive flexibility in aging.

It is worth noting that the inclusion criteria for this study slightly differed from Spreng et al. (2017) (detailed comparisons can be found in the Supplementary Material). For example, Müller et al. (2018) proposed guidelines restricting meta‐analyses to only include studies reporting results at certain statistical thresholds. Therefore, this study required empirical studies to meet a specific significance threshold (p _corrected  < .05 or p _uncorrected  < .001), a criterion not imposed by the previous meta‐analysis, leading to the exclusion of some studies (DiGirolamo et al., 2001; Esposito et al., 1999; Gazes et al., 2012; Steffener et al., 2014) that were included in the previous meta‐analysis.

7.1. Task positive networks

Cognitive flexibility in aging was associated with age‐related differences in the activation of the frontoparietal network. Cognitive flexibility tasks typically involve trial‐by‐trial control, including the suppression of irrelevant task settings and the configuration of new settings (Kiesel et al., 2010; Koch et al., 2018). The frontoparietal network is believed to support this trial‐by‐trial control through a top‐down manner. Furthermore, when implementing cognitive flexibility tasks, the frontoparietal regions has been found to be positively correlated with dopamine synthesis capacity (Berry et al., 2016; Samanez‐Larkin et al., 2013). The dopamine synthesis is thought to enhance the neural efficiency of the frontoparietal network during top‐down control processing. Studies have revealed age‐related differences in dopaminergic modulation of frontoparietal network (Berry et al., 2016; Berry et al., 2018). Therefore, the evidence of activation and dopaminergic modulation in the frontoparietal network highlights age‐related differences in top‐down control processing during cognitive flexibility tasks.

Moreover, our meta‐analysis revealed both increased and decreased activation in regions within the frontoparietal network, echoing findings in previous studies. For example, compared to young adults, older adults displayed lower activation in hub regions within the frontoparietal network during task performance (Eich et al., 2016; Nashiro et al., 2018). This decrease in activation was correlated with poorer behavioral performance (Nagahama et al., 1997; Nashiro et al., 2018). However, increased activation of regions within the frontoparietal network (e.g., the right dlPFC and IFG) has also been reported in other studies (Worthy et al., 2016; Zhu et al., 2014; Zhu et al., 2015). Reduced activity in the elderly can reasonably be assumed to reflect a reduced level of neural functioning (Grady, 2012; Spreng et al., 2010), while the increased activation posed a challenge of interpretation.

Two possible explanations may account for increased activation in the aging brain: compensation or dedifferentiation. The CRUNCH proposed that the aging brain may try to recruit more neural resources at a low cognitive load, but this over‐recruitment can be invalidated and lead to decreased activation in difficult conditions (Festini et al., 2018; Schneider‐Garces et al., 2010). The CRUNCH is an appealing hypothesis because it addresses the coexistence of both increased and decreased activation by considering the task difficulty. Therefore, the mixed activation patterns in frontoparietal regions could be a result of varying task difficulties in different studies.

On the other hand, empirical studies suggested that increased activation in older adults may indicate neural dedifferentiation. For example, studies included in the current meta‐analysis found that older adults with over‐recruitment in frontoparietal and attention networks had higher switch costs (Gold et al., 2013; Nashiro et al., 2018; Zhu et al., 2015). However, older adults with less over‐recruitment demonstrated better task‐switching performance (Gold et al., 2013). Furthermore, when dividing older adults into younger‐old and older‐old groups, both groups showed that over‐recruitment in frontoparietal regions correlated with larger global switch costs (Nashiro et al., 2018). This finding aligns with the concept of neural dedifferentiation. Specifically, the brain hinges on the fine‐tuned orchestration of regions within task positive networks to achieve task goals (Menon & D'Esposito, 2022). In older adults, it is likely that neural dedifferentiation leads to imprecise recruitment of target regions, which prevents older adults from achieving better cognitive flexibility performance (Grady, 2012; Koen & Rugg, 2019; Rieck et al., 2021).

In addition to the frontoparietal network, the dorsal attention network was associated with the cognitive flexibility in aging. During tasks involving cognitive flexibility, subjects typically need to proactively switch their attention between stimuli in accordance with the task rules. The dorsal attention network is typically involved in top‐down attentional processes (Tamber‐Rosenau et al., 2018; Vossel et al., 2014). Therefore, older adults may have difficulties in flexibly and proactively directing their attention to the target stimuli when task rules are changing. Notably, when analyzing lateralization in the dorsal attention network, we found that older adults exhibited lower activation in the left hemisphere, but higher activation in the right hemisphere compared to younger adults. Cognitive flexibility is functionally dominant in the left hemisphere (Capizzi et al., 2016; Serrien & Sovijarvi‐Spape, 2013). However, older adults displayed a LRS in the activation pattern within the dorsal attention network, indicating reduced functionally dominant during cognitive flexibility. These findings align with the HAROLD model, which proposes that age‐related reductions in hemispheric asymmetry are evident in cognitive flexibility in aging. Typically, decreased brain activity is interpreted as a reflection of a reduced level of neural functioning, while increased activity is seen as greater recruitment of neural resources (Festini et al., 2018; Grady, 2012; Reuter‐Lorenz & Park, 2014; Stuss & Knight, 2013). Therefore, the aging brain may recruit more neural resources in the right hemisphere to compensate for the decreased top‐down attentional control, which is associated with diminished functional dominations in the left hemisphere.

Finally, compared to young adults, older adults predominantly exhibited increased activation in the ventral attention network. The volume of over‐recruited regions within the ventral attention network was smaller than that of the dorsal attention network. The dorsal attention network mediates the voluntary attentional control (i.e., top‐down attention control), whereas the ventral attention network is responsible for detecting salient stimuli in a bottom‐up manner (Suo et al., 2021; Tamber‐Rosenau et al., 2018; Vossel et al., 2014). Functional alterations in the top‐down attentional control may be more closely related to age‐related declines in cognitive flexibility than to salience attention. Notably, the regions with increased activation within the ventral attention network were left‐lateralized. Abundant anatomical and neuroimaging observations suggest that the ventral attention network is a naturally right‐lateralized network (Farrant & Uddin, 2015; Geng & Vossel, 2013; Igelstrom et al., 2015). This raises the possibility that it is difficult for the aging brain to recruit more neural resources in the fully utilized hemisphere and that the non‐dominant hemisphere may act as a “backup resource pool” to respond when the attentional control is required.

7.2. Age‐related changes in the visual and somatomotor networks

In the current study, most of the regions (82.8%) showing decreased activation in older adults fell into the visual network. Moreover, the activation strength of these visual regions was negatively correlated with behavioral performance (Nagahama et al., 1997), supporting the PASA model's view on age‐related declines in perceptual abilities. Extensive studies have confirmed age‐related differences in the activation of the occipital cortex across various cognitive control subdomains, such as attention control, inhibitory control, and working memory (Dennis et al., 2014; Madden et al., 2002; Payer et al., 2006; Rieck et al., 2015). Therefore, it is likely that decreased activation in the visual network is not exclusive to cognitive flexibility in aging, but rather represents an age‐related difference affecting multiple cognitive processes. Additionally, age‐related decrease in the activation of the visual network was primarily observed in the left hemisphere. Prior research has shown that the left hemisphere is primarily responsible for processing local features visually (Brederoo et al., 2020). Switch‐based tasks often require subjects to attend and respond to local features of stimuli, such as responding color or shape of a stimulus. This left lateralization for visual regions with decreased activation potentially reflected an association between decreased cognitive flexibility and age‐related differences in perceiving the local features of switch stimuli.

Unlike the decreased activation observed in the visual network, the somatomotor network exhibited increased activation in older adults. Cognitive flexibility tasks require participants to suppress goal‐unrelated motor responses, and instead generate responses that align with the current goal. This necessitates the flexible updating of response patterns supported by regions within the somatomotor network (Cocuzza et al., 2020; Remington et al., 2018; Schultz et al., 2022; Zhang et al., 2015). The increased activation in the somatomotor network suggested that the aging brain may attempt to increase the recruitment of neural resources to ensure flexible motor control when task rules change. Moreover, it should be noted that these regions with increased activation were right‐lateralized. The skill for performing hand movements in right‐handers generally depends on the activation of left hubs within the somatomotor network (e.g., the left supplementary motor area) and the inhibition of contralateral regions (Fuente‐Fernandez et al., 2000; Kilincer et al., 2019). Similar to the ventral attention network, when the network is naturally lateralized, the aging brain may try to compensate for the non‐dominant hemisphere, but not the side that is already utilized. Thus, the non‐dominant hemisphere of the somatomotor network is able to compensate for aging, but the dominant hemisphere is more susceptible to aging due to years of use.

7.3. Increased activation in the default mode network

Our results showed that older adults exhibited higher activation in the default mode network compared to young adults. Moreover, among the six identified networks, the default mode network contained the most regions with increased activation. This finding is in accordance with that of previous studies focusing on older adults that reported increased activation in hubs within the default mode network during control tasks (Qin & Basak, 2020; Sambataro et al., 2010). Additionally, an empirical study included in the current meta‐analysis confirmed a negative relationship between the activation strength of default mode network and cognitive flexibility (Gold et al., 2013). These findings align with the default‐executive coupling hypothesis of aging (DECHA). The default mode network is involved in self‐reflective and internally directed cognition, such as autobiographical memory and semantic knowledge (Buckner & DiNicola, 2019; Raichle, 2015). The DECHA proposes that the default mode network increasingly couples with prefrontal regions (Spreng & Turner, 2019). When endogenous thoughts are irrelevant to task goals, the default mode network can interfere with the control processes, leading to poorer task performance (Turner & Spreng, 2015). Cognitive flexibility mainly involves generating appropriate responses to external stimuli. Therefore, over‐recruitment of the default mode network suggests that cognitive flexibility in aging is partly due to increased interference of task‐irrelevant information (Smallwood et al., 2021).

7.4. Methodological considerations

From a methodological perspective, the current study has the following considerations: First, we adopt the MKDA rather than the ALE to perform our analysis. A recommendation was proposed that meta‐analysis based on ALE should include at least 17 experiments to reduce the influence of a single experiment on the general results (Eickhoff et al., 2016; Müller et al., 2018). However, the MKDA nests peak coordinates within study contrast maps to ensure that no single study excessively influence the final results (Wager et al., 2007; Wager et al., 2009). Therefore, the MKDA was adopted given that our meta‐analysis included a modest number of studies.

Second, unlike the habit‐based choice of kernel size in MKDA (Wager et al., 2007; Wager et al., 2009), this study used an adaptive approach to determine the kernel size, which allowed the maximum of statistical sensitivity for regions with different activation patterns in older adults. The results showed that regions with decreased activation were contiguous and mainly concentrated in the occipital lobe (Figure 2, upper panel); therefore, it is not surprising that a smaller convolution kernel (r = 10 mm) can result in statistical optimization. Meanwhile, the regions with increased activation were distributed across networks and were spatially dispersed (Figure 2, lower panel); thus, a larger convolutional kernel (r = 13.6 mm) was more appropriate.

Third, the current study selected the Yeo‐7 network template to map regions associated with cognitive flexibility in aging. The Yeo‐7 template is appropriate because of its boundary‐based partition of functional networks, which ensures that the activated regions can be adequately associated to functional networks. Other templates (e.g., Power 264 and Dosenbach 160) using representative regions to define functional networks, and cover only a small portion of the whole brain (Dosenbach et al., 2010; Power et al., 2011). As a result, regions involved in cognitive flexibility in aging might be located outside these templates. Additionally, other templates define regions' partitions without incorporating network divisions (Glasser et al., 2016; Gordon et al., 2016), making them unsuitable for our study.

Fourth, in line with the common practice for studies that employ MKDA (Li et al., 2015; Wager et al., 2009), we excluded one study with non‐statistically significant unreported effects (Nashiro et al., 2013). Excluding one study with a nonsignificant group difference is unlikely to have a significant impact the overall results. Methodologically, the MKDA can be performed when only the spatial coordinate positions are known, but it cannot handle non‐statistically significant unreported effects. Recently developed methods, such as seed‐based d mapping, however, can estimate these unreported effects based on the reported statistical parameters (e.g., z‐ or t‐value) from other studies (Albajes‐Eizagirre, Solanes, & Radua, 2019; Albajes‐Eizagirre, Solanes, Vieta, & Radua, 2019; Radua et al., 2015). Several studies on cognitive flexibility in aging only report voxel coordinate positions without corresponding precise statistical parameters (Eich et al., 2016; Worthy et al., 2016), making it difficult to use seed‐based d mapping to evaluate these unreported effects. We recommend re‐evaluating these results in the future when more research accumulates.

7.5. Limitations

Few empirical studies have directly compared activation difference between older and young adults during cognitive flexibility performance. However, a reduction in cognitive flexibility is a major manifestation of brain aging, which needs to be fully examined. Although relatively few studies were included in the current analysis, smaller sample sizes may be sufficient for reliable meta‐analysis when a strong effect is expected (Müller et al., 2018). It is widely known that there are significant activation differences between older and younger adults during cognitive control performance. Thus, as with many previous studies (Dijkstra et al., 2020; Duda & Sweet, 2020; Picó‐Pérez et al., 2017), we performed meta‐analysis based on relatively few empirical studies but with large effect size. We further recommend that a meta‐analysis should be performed again to validate the results after more studies have been accumulated in this filed. Finally, based on the framework proposed by Diamond (2013), we have classified cognitive flexibility tasks into two categories: task switching and set shifting. As the body of research grows, adopting a more refined approach to task classification will be crucial for exploring the common and distinct mechanisms of cognitive flexibility in aging (Kim et al., 2011; Kim et al., 2012).

8. CONCLUSION

Cognitive flexibility in aging is associated with age‐related differences in the activation patterns of several functional networks, including the dorsal attention, ventral attention, frontoparietal, somatomotor, visual, and default mode networks. Regions with increased activation in older adults exhibited network‐dependent lateralization patterns, whereas regions with decreased activation were left‐lateralized regardless of the network to which they belonged. We highlight that MCS‐based lateralization analysis offers new metrics for measuring the spatial distribution of regions in MRI meta‐analyses.

FUNDING INFORMATION

This work was supported by the grant from the National Natural Science Foundation of China (32171040).

CONFLICT OF INTEREST STATEMENT

The authors have no conflicts of interest to declare.

Supporting information

DATA S1: Supplementary Information.

Xia, H. , Hou, Y. , Li, Q. , & Chen, A. (2024). A meta‐analysis of cognitive flexibility in aging: Perspective from functional network and lateralization. Human Brain Mapping, 45(14), e70031. 10.1002/hbm.70031

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the corresponding author upon reasonable request.