Ageing is accompanied by a range of changes in cognitive function. While crystallised cognitive abilities such as general and vocabulary knowledge tend to remain stable or even slightly improve with age, other so-called fluid abilities show a gradual decline (1–3). For example, it is estimated that fluid cognitive function decreases by approximately 0.02 Standard Deviation (SD) per year throughout adulthood (4). These fluid abilities encompass domains such as memory, executive functioning and processing speed (5). Importantly, this age-related, non-pathological decline becomes clearer in middle age and tends to be even more pronounced in older adulthood (1, 6). Given the ongoing global demographic shift towards older populations, preserving optimal levels of cognitive function becomes increasingly important, as good cognitive health plays a critical role in maintaining quality of life and independence (7, 8). Therefore, the United Nations have declared the current decade as the "Decade of Healthy Ageing" underscoring the need to develop and implement effective strategies to support (cognitive) health in older adults (9).

Several factors are increasingly recognized for their role in age-related cognitive decline, including sociodemographic factors such as education level, as well as lifestyle factors including smoking habits, alcohol consumption and physical activity (PA) levels (2, 10). Recently, the focus has expanded from PA alone to incorporate all daily movement behaviours, including sedentary behaviour (SB) and sleep, due to their combined role in supporting cognitive health at older age. For instance, the first systematic review on the relationship between 24-hour movement behaviours (i.e. PA, SB and sleep, also referred to as ‘time use’) and cognition in healthy older adults (average age of 71.3 ± 5.4 years) concluded that spending more time in PA and reducing sitting time were most often associated with better general cognition (11). However, 21 out of the 23 included studies did not account for all three behaviours within a single statistical framework. Given that time in a day is finite, these behaviours are inherently co-dependent: spending more time in one behaviour (e.g. PA) requires reducing time in another (e.g. SB and/or sleep) (12). Therefore, to investigate their relationship with cognition, these behaviours should be studied in one model, addressing multicollinearity. A 24-hour compositional data analysis (CoDA) framework overcomes this problem of multicollinearity by expressing time use as isometric log-ratio coordinates. This mathematical transformation allows researchers to apply traditional statistical analyses such as regression analyses, as well as to hypothetically reallocate time between behaviours to investigate the associated differences in (cognitive) outcomes (13, 14).

To date, relatively few studies have examined the associations between 24-hour movement behaviours and cognitive functioning in middle-aged and older adults using CoDA (15–22). Moreover, the available studies differ in their measurement approaches: while earlier research combined self-reported data with accelerometery to construct the 24-hour activity profile (17, 18, 20), more recent studies have shifted towards solely device-based measures of movement behaviours (15, 16, 19, 21, 22). Overall, the existing evidence remains limited, cross-sectional and inconclusive. While two studies reported no significant associations between time-use patterns and cognition in this population (16, 20), five studies suggested a potential relationship and pointed out that reallocating time to moderate-to-vigorous physical activity (MVPA) may be positively associated with cognitive outcomes (15, 17, 19, 21, 22). Notably, the type of reallocation varied across these studies: one examined only proportional reallocations (i.e. proportionally from all behaviours) (15), another focused solely on specific substitutions (i.e. taking time from SB, light PA, or sleep) (19), while others considered both approaches (17, 21, 22). More importantly, the latter studies all investigated cross-sectional associations which do not allow for the examination of 24-hour patterns and cognitive function over time. This limits insight into their evolving relationship throughout ageing. Therefore, longitudinal studies are needed to more accurately assess the relationship between time use and cognitive function in older adults.

The current study will address this gap in the literature by investigating how device-measured 24-hour movement behaviours (i.e. light PA, moderate-to-vigorous PA, SB and sleep) are longitudinally associated with various cognitive domains, namely short- and long-term memory, executive function and processing speed, in healthy adults aged 55 years and above. More specifically, this study will examine if associations between 24-hour movement behaviours and cognition exist and if these remain stable or vary over a two-year period across three time points (baseline, one-year follow-up and two-year follow-up).

Participants were healthy, community-dwelling adults aged 55 years and older, recruited via convenience sampling. This age range was chosen based on research indicating that cognitive decline can be seen as early as midlife (6). Eligibility at baseline was determined through self-reported information. Specifically, individuals were excluded if they had: (a) a neurodegenerative disorder (e.g., Parkinson’s disease, Alzheimer’s disease, multiple sclerosis); (b) a psychiatric condition such as bipolar disorder or obsessive-compulsive disorder; (c) a serious brain injury within the past year or earlier with lasting effects; (d) a history of stroke; (e) an active depressive episode at the time of assessment; (f) a history of substance abuse or excessive alcohol abuse; or (g) a first-degree relative diagnosed with dementia. Additional exclusions included: (h) the use of a sleep device for diagnosed sleep apnoea; (i) a diagnosis of chronic insomnia; or (j) severe impairments in daily functioning. In sum, exclusion criteria were conditions known to affect cognitive functioning (a–g) and 24-hour movement patterns (h–j). Furthermore, cognitive screening was conducted using the Montreal Cognitive Assessment (MoCA), and individuals scoring 23 or below were excluded due to the potential presence of mild cognitive impairment (23, 24).

Participants wore at each time point a tri-axial accelerometer (ActiGraph wGT3X-BT) on their non-dominant wrist for seven consecutive days, removing it only for water-based activities. The device was set to sample at 100 Hz with idle sleep mode disabled. A diary was provided to daily log non-wear periods and sleep-related times, including time in bed, estimated sleep onset, final wake-up and time out of bed.

Raw data (.gt3x format) were downloaded using ActiLife (v6.13.4) and processed in R (v4.4.1) via RStudio (v2024.09.1) using the GGIR package (v3.1-4) (38, 39). Files were autocalibrated by comparing non-movement periods to gravitational acceleration (1g) (40). Non-wear detection followed GGIR’s 2023 algorithm, identifying 60-minute windows (shifted every 15 minutes) with low variability (SD < 13.0 mg and range < 50 mg in ≥ 2 axes) as non-wear. These segments were imputed using the average values from the same time on other days with valid data.

The ENMO (Euclidean Norm Minus One with negative values rounded to zero) metric was computed in 5-second epochs. To determine sleep periods, sustained inactivity bouts were identified as intervals where the z-angle remained within a 5-degree threshold for a duration of at least 5 minutes (41). Based on these bouts, the Sleep Period Time (i.e. the interval from sleep onset to final awakening) was identified by the van Hees-algorithm guided by information from participants’ sleep diaries, namely their reported time in bed and wake-up time (41). Quality control was performed to identify cases where the algorithm-derived sleep onset or wake-up times differed significantly (by one hour or more) from self-reported sleep onset and wake-up times. If so, alternative reference points were used as guiders for the algorithm. These included estimates from the HDCZA algorithm (Heuristic algorithm looking at Distribution of Change in Z-Angle) (42) or participants’ recorded times of getting out of bed – whichever provided Sleep Period Time windows that more closely aligned with self-reports (less than one hour difference) and more consistently aligned with the visual activity signal.

A day was defined from midnight-to-midnight and valid if it had ≥ 23 hours of recorded data (including non-wear), with ≥ 16 valid hours and at least two-thirds of valid waking hours. Inclusion of the measurement required at least four valid weekdays and one valid weekend day (43). Weighted averages (5:2 ratio) were applied to balance weekday and weekend data.

Activity intensities were classified using cut-points defined by Hildebrand et al. (2014) (44) and Hildebrand et al. (2016) (45): sedentary (< 44.8 mg), light (44.8–100.6 mg) and moderate-to-vigorous (> 100.6 mg). Full GGIR code specifications are available on GitHub: https://github.com/pjmarent/PASOCA-longitudinal-time-use-cognition.

To address missing data and participant attrition across measurements (see flowchart, Fig. 1), multiple imputation by Fully Conditional Specification was performed using SPSS (version 29.0.2.0). Constraints were applied to ensure the plausibility of imputed values, including bounded ranges for the cognitive test scores. The procedure generated 20 imputed datasets using automatic estimation methods. All the following statistical analyses were conducted across all imputations and final estimates were pooled using Rubin’s rule (50).

Statistical analyses were conducted using R version 4.4.2 (51). Given the inherently compositional nature of 24-hour movement behaviour data, a compositional data analysis (CoDA) framework was employed (52). Time-use variables were first normalized to a 24-hour total using the closure function from the compositions package (53). No zero values were recorded for any of the 24-hour movement behaviours. A sequential binary partitioning strategy was then applied to generate three isometric log-ratio (ILR) coordinates, reflecting: (1) sleep relative to waking behaviours (i.e., SB, LPA and MVPA); (2) SB relative to LPA and MVPA; and (3) LPA relative to MVPA (12). These ILRs served as predictors in subsequent modelling.

To evaluate longitudinal associations between the 24-hour movement behaviours and cognitive performance, linear mixed-effects models were fitted. Each model incorporated fixed effects for the time use variable (through the three ILRs previously defined), time (treated as a categorical variable) and their interaction (time use × time), allowing assessment of whether associations varied across the two-year follow-up period. Random intercepts were included to capture between-individual variability at baseline and to account for the nested structure of repeated measurements within individuals. A series of hierarchical linear models were developed for each of the four cognitive domain scores. The initial model included only time use, time and their interaction as predictors. Three subsequent models progressively included additional covariates: sociodemographic factors were added in the second model (age, sex, education, living situation), health-related variables (BMI category, total prescribed medications, hearing issues) in the third and behavioural health indicators (alcohol consumption and smoking status) completed the fourth model. Note that the models for long-term memory always included an additional covariate representing the time interval (in seconds) between the stimulus presentation and the recall phase (see Additional Table 1A for more information on this variable).

To compare the nested models for each cognitive outcome, the D1 multivariate Wald test was employed (54). This test evaluates whether the inclusion of additional covariates significantly improves model fit relative to a simpler, nested model (e.g., assessing whether the sociodemographic model provides a better fit than the time-use-only model). Type II ANOVA tests were used to assess the significance of individual predictors in the final models. The ILR coordinates are not interpreted in isolation; rather, they collectively capture the compositional structure of time-use data. Consequently, a single overall F-statistic is reported for time use, representing the combined contribution of all three ILR coordinates. The significance level was set at α = 0.05 level for all analyses.

Post-hoc compositional isotemporal substitution analysis was performed if time use was significantly associated with a cognitive outcome. These hypothetical time reallocations were simulated using a reference profile with characteristics reflecting the average participant over the three time points as presented in Table 3. Results must be interpreted relative to this profile. Time was reallocated in 5-minute increments, ranging from 0 (no reallocation) to 30 minutes, by transferring time from one behaviour to another while holding all other behaviours constant (one-for-one reallocations). This approach estimates the hypothetical change in cognitive composite scores resulting from reallocating time between pairs of 24-hour movement behaviours.

Previous research has identified potential associations between the 24-hour movement behaviours (i.e. physical activity, sedentary behaviour and sleep) and cognitive function in older adults. However, their cross-sectional designs have limited insight into how these associations evolve over time. Therefore, this longitudinal study measured healthy adults aged 55 years and older at three time points over a two-year period, with approximately one year between assessments, to examine associations between device-measured time use and multiple cognitive domains over time. The results showed a consistent association between time use and executive function throughout the follow-up period, whereas no such associations were observed for short-term memory, long-term memory or processing speed. Time spent in moderate-to-vigorous PA was the primary driver of this association: hypothetically reallocating time from light PA or sleep towards MVPA seemed most beneficial for EF. Conversely, reallocating time away from MVPA resulted in lower EF, reinforcing the cognitive benefits of maintaining or increasing MVPA levels. Reallocation results remained directionally consistent across all time points, supporting the presence of a stable association between time use and executive function over time.

In line with previous research, these findings underscore the importance of examining the effect of time-use composition on cognitive health, particularly highlighting the role of MVPA. The systematic review of Maddison et al. (2022) (11) examining the combined role of PA, SB and sleep on cognition in older adults concluded that higher proportions of MVPA were most consistently associated with better general cognition. However, most studies included in that review assessed only two of the three time-use behaviours, limiting their ability to capture the full interdependent nature of the 24-hour movement behaviours. More recent studies have addressed this limitation by applying compositional data analysis, which allowed for the simultaneous modelling of all three behaviours (15–21). For example, Dumuid et al. (2022) (15) demonstrated that reallocating time to MVPA – proportionally from LPA, SB and sleep – was positively associated with global cognition and executive function, with effects being more pronounced in individuals carrying the APOE ε4 allele, a genetic risk factor for Alzheimer’s disease. Similarly, Collins et al. (2025) (21) found that time-use composition of healthy older adults (n = 648, 69.8 ± 3.7 years) was significantly associated with processing speed, working memory and executive function/attentional control, but not with episodic memory or visuospatial function. Their post-hoc isotemporal substitution analyses indicated that MVPA was the primary driver of these associations, with hypothetical time reallocations towards MVPA being associated with better cognitive performance. In contrast, reallocations away from MVPA were associated with poorer performance, irrespective of which behaviour time was taken from (21). The present study is the first to demonstrate similar associations using longitudinal data, offering stronger evidence for associations between time use and cognitive outcomes over time. Nevertheless, uncertainty remains regarding which cognitive domains are most sensitive to time-use compositions. While this study did not identify any relationships with memory and processing speed, similar to the cross-sectional study of time use and cognition in healthy older adults by Mellow et al. (2022) (16), others studies have reported such associations (17, 21). This variability likely reflects methodological differences in both the assessment of cognitive domains (e.g. different cognitive tests) and the measurement of 24-hour movement behaviours (e.g. accelerometer wear location (wrist vs. waist), cut-points to classify activity intensities), underscoring the need for harmonized protocols in future research (11).

Furthermore, several studies highlighted the potential benefits of reallocating time from other behaviours to MVPA (15, 17, 19, 21, 22). However, this should not be interpreted as diminishing the established importance of behaviours such as sleep for cognitive function and healthy ageing (55, 56). Because MVPA constitutes a relatively small proportion of the 24-hour day in this cohort (around 4–5% on average), modest reallocations of only a few minutes represent a comparatively large proportional increase, which were hypothetically associated with positive estimated changes in cognitive performance. While small reductions in sleep are unlikely to meaningfully affect cognition, reallocating time from LPA or SB rather than sleep may be more realistic and feasible (57). Future interventional studies will have to investigate how people can reallocate time to MVPA and to clarify the contextual circumstances under which such reallocations are most effective.

Interestingly, in the post-hoc compositional isotemporal substitution analysis, reallocating time from LPA to SB was associated with positive estimated changes in EF, suggesting potential cognitive benefits of certain sedentary activities. This underscores the need for future studies to consider the context of 24-hour movement behaviours, as cognitively engaging sedentary behaviour (e.g. reading, social interaction, computer-based tasks…) may partly explain these findings (58, 59). Nevertheless, reallocations to MVPA from LPA generally predicted better EF z-scores estimates, reinforcing the importance of MVPA for cognitive health.

The observed cognitive benefits of MVPA may be explained by several biological mechanisms (60). At the neurobiological level (higher intensity) PA not only increases the release of growth factors playing an important role in angiogenesis and neurogenesis (i.e. brain-derived neurotrophic factor (BDNF), insulin-like growth factor 1 (IGF-1), vascular endothelial growth factor (VEGF)), but also stimulates the production of lactate which can serve as an additional energy source for the brain (61). Additionally, reductions of pro-inflammatory cytokine (e.g. tumour necrosis factor-α (TNF)) and oxidative stress are reported (61). At the brain health level, structural changes have been observed in response to physical activity (60). For example, Erickson et al. (2011) (62) demonstrated that a 12-month aerobic exercise at moderate intensity led to significant increases in hippocampal volume in older adults between 55 and 80 years old.

Interestingly, an improvement was seen in cognitive scores over time across the domains short-term memory and executive function. While the nature of this study remains purely observational and the found associations between time use and cognitive outcomes remain robust (63), these results warrant further contextualization. First of all, CANTAB tests are sensitive to age-related decline (64–67) and showed adequate 3-month test-retest reliability similar to other cognitive tests (68, 69). However, the observed improvements in cognitive scores are likely attributable to practice effects, a well-documented phenomenon in longitudinal cognitive research (70). These effects can persist even across long intervals and multiple follow-up assessments, particularly in domains like memory and EF (71). The observed improvements or potential practice effects in this study ranged from 0.01 to 0.20 SD, which can be considered as ‘optimal’, meaning they fall within an acceptable range for repeated cognitive testing (72). Despite efforts to mitigate these effects – such as using alternate test versions, including a training phase before each test and ensuring participant familiarity with the researcher and testing environment – some degree of improvement due to repeated exposure is inevitable (73). This is especially true for EF tasks, which rely on novelty, problem-solving and strategy, making them particularly sensitive to prior exposure (74). Overall, these factors might explain the improved cognitive scores.

A key strength of this study is its longitudinal design, which allowed the examination of within-person changes in cognitive function over time in relation to daily time-use composition. This approach provides stronger evidence for temporal associations than cross-sectional analyses and enhances the ecological validity of our findings. Additionally, the use of device-based measures for PA, SB and sleep, combined with a compositional data analysis framework, enabled a more accurate assessment of 24-hour time use. This study also employed a comprehensive neurological test battery assessing multiple cognitive domains, rather than relying on a single global measure or brief screening tool. However, several limitations should be acknowledged. First, the relatively short follow-up interval (one year between measurements) and study period (two years in total) may have been insufficient to detect meaningful cognitive decline, particularly given that the sample consisted of healthy older adults with relatively preserved cognitive function at baseline. Second, although practice effects may partly account for the observed improvements in cognitive scores, reliable change indices for all test outcomes were not available to adjust for this. Future research should develop and apply such indices to better distinguish true cognitive change from retest effects. Third, while the analyses focused on the composition of time-use behaviours, the context or quality of PA, SB and sleep was not accounted for (e.g. social vs. solitary activity, cognitively engaging vs. passive activities, sleep quality…), which may differentially influence cognitive outcomes (75). Fourth, GGIR classifies “inactivity” which may not precisely represent sedentary behaviour as it cannot distinguish standing from sitting or lying. Fifth, the study sample was not fully representative of the general population, with a higher proportion of highly educated participants than typically seen among Belgian middle-aged and older adults. Additionally, exclusion of individuals with health conditions related to cognition or movement behaviours may have introduced a healthy volunteer bias, limiting generalizability to broader populations.

Future research should replicate the analysis of this study in larger sample size and over longer follow-up periods and in diverse populations. In addition, future work should aim to explore the contextual characteristics of daily behaviours and their cognitive relevance and consider intervention studies that actively promote realistic shifts towards more beneficial time-use compositions. Such work could help clarify not only whether but also how reallocating time to MVPA can support cognitive health in ageing populations.