How are age-related differences in sleep quality associated with health outcomes? An epidemiological investigation in a UK cohort of 2406 adults

Objectives To examine age-related differences in self-reported sleep quality and their associations with health outcomes across four domains: physical health, cognitive health, mental health and neural health. Setting Cambridge Centre for Ageing and Neuroscience (Cam-CAN) is a cohort study in East Anglia/England, which collected self-reported health and lifestyle questions as well as a range of objective measures from healthy adults. Participants 2406 healthy adults (age 18–98) answered questions about their sleep quality (Pittsburgh Sleep Quality Index (PSQI)) and measures of physical, cognitive, mental and neural health. A subset of 641 individuals provided measures of brain structure. Main outcome measures PSQI scores of sleep and scores across tests within the four domains of health. Latent class analysis (LCA) is used to identify sleep types across the lifespan. Bayesian regressions quantify the presence, and absence, of relationships between sleep quality and health measures. Results Better self-reported sleep is generally associated with better health outcomes, strongly so for mental health, moderately for cognitive and physical health, but not for sleep quality and neural health. LCA identified four sleep types: ‘good sleepers’ (68.1%, most frequent in middle age), ‘inefficient sleepers’ (14.01%, most frequent in old age), ‘delayed sleepers’ (9.28%, most frequent in young adults) and ‘poor sleepers’ (8.5%, most frequent in old age). There is little evidence for interactions between sleep quality and age on health outcomes. Finally, we observe U-shaped associations between sleep duration and mental health (depression and anxiety) as well as self-reported general health, such that both short and long sleep were associated with poorer outcomes. Conclusions Lifespan changes in sleep quality are multifaceted and not captured well by summary measures, but instead should be viewed as as partially independent symptoms that vary in prevalence across the lifespan. Better self-reported sleep is associated with better health outcomes, and the strength of these associations differs across health domains. Notably, we do not observe associations between self-reported sleep quality and white matter.


STATISTICAL ANALYSES
We examine whether self-reported sleep patterns change across the lifespan, both for the PSQI sum score and for each of the seven PSQI components. We then examine the relationships between the sleep quality and the four health domains in three ways: First, simple regression of the health outcome on sleep variables, to determine evidence for association between poor sleep quality and poor health outcomes. Second, we include age as a covariate. Finally, we include a (standard normal rescaled) continuous interaction term to examine whether there is evidence for a changing relationship between sleep and outcomes across the lifespan.
For all regressions we will use a default Bayesian approach advocated by Liang, Paulo, Molina, Clyde, & Berger, (2008); Rouder & Morey, (2012) ;Wagenmakers, (2007); Wei et al., (2012); Wetzels et al., (2011), which avoids several well-documented issues with p-values (57), allows for quantification of null effects, and decreases the risk of multiple comparison problems (e.g. Gelman, Hill, & Yajima, 2012). Bayesian regressions allows us to symmetrically quantify evidence in favour of, or against, some substantive model as compared to a baseline (e.g. null) model. This evidentiary strength is expressed as a Bayes Factor (see Jeffreys (61), which can be interpreted as the relative likelihood of one model versus another given the data and a certain prior expectation. A Bayes Factor of, e.g., 7, in favour of a regression model suggests that the data are seven times more likely under that model than an intercept only model (for an empirical comparison of p-values and Bayes factors, see Wetzels et al., 2011). A heuristic summary of evidentiary interpretation can be seen in Figure 1.
[insert Figure 1 here] We report log Bayes Factors for large effects and regular Bayes Factors for smaller effects.
To compute Bayes Factors we will use Default Bayes Factor approach for model selection (55,56) 12 the true effect will lie between -.707 and .707. Prior to further analysis, scores on all outcomes were transformed to a standard normal distribution, and any scores exceeding a z-score of 4 or -4 were recoded as missing (aggregate percentage outliers across the four health domains: Cognitive, 0.41%, Mental, 0.16%, Neural, 0.37% Physical, 0.031%).
To better elucidate individual differences in sleep quality we next use Latent Class Analysis (64). This technique will allow us examine individual differences in sleep quality across the lifespan in more detail than afforded by simple linear regressions: Rather than examining continuous variation in sleep components, LCA classifies individuals into different sleep types, each associated with a distinct profile of 'sleep symptoms'. If there are specific constellations of sleep problems across individuals, we can quantify and visualize such sleep types. Moreover, by using Latent Class Regression, we can examine whether the likelihood of belonging to any sleep 'type' changes as a function of age. To analyse the data in this manner, we binarized the responses on each component into 'good' (0 or 1) or 'poor' (2 or 3).

Age-related differences in sleep quality
First, we examined sleep changes across the lifespan by examining age-related differences in the PSQI sum score (N= 2178, M=5.16, SD=3.35, Range=0-19). Regressing the PSQI global score on age, (see Supplementary Figure 1) showed evidence for a positive relationship across the lifespan (logBF 10 = 10.45). This suggests that on the whole, sleep quality decreases across the lifespan (note that higher PSQI scores correspond to worse sleep). Although we observe strong statistical evidence for an age-related difference ('Extreme' according to ), age explained only 1.23 % of the variance in the PSQI Total score. Next, we examined each of the seven components on age in the same manner. In Supplementary Figure 2 we see that that age has varying and specific effects on different aspects of sleep quality, and did not worsen uniformly across the lifespan. For example, we observed moderate evidence that sleep latency did not change across the lifespan (Sleep Latency, Finally, we entered all seven components into a Bayesian multiple regression simultaneously, to examine to what extent they could, together, predict age. The best model included every component except Sleep Latency (logBF 10 = 142.71). Interestingly, this model explained 13.41% of the variance in age, compared to 1.23% for the PSQI Total score, and 6.4% for the strongest single component. This shows that lifespan changes in self-reported sleep are heterogeneous and partially independent, and that specific patterns and components need to be taken into account simultaneously to fully understand age-related differences in sleep quality. These finding shows that neither the PSQI sum score nor the sleep components in isolation fully capture differences in sleep quality across the lifespan.
Next we examined evidence for distinct sleep types using Latent Class Analysis (64). We fit a set of possible models (varying from 2 to 6 sleep types) We found that the four class solution gives the best solution, according to the Bayesian Information Criterion (65) Figure 2A. Their responses to any of the seven sleep components are likely to be 'poor' or 'very poor', almost universally so for 'sleep quality' (94%) and 'Sleep Efficiency' (97.7%).
Next, we including age as a covariate (simultaneously including a covariate is known as latent class regression or concomitant-variable latent class models (66). This analysis, visualised in versus class 1: beta/SE 0.01269/0.00478, t=2.655, for more details on generalized logit coefficients , see Linzer & Lewis, 2011, p. 21). The frequency of Class 1 (Good sleepers) peaks in middle to late adulthood, dropping increasingly quickly after age 50. Class 2 (Inefficient sleepers) are relatively rare in younger individuals, but the prevalence increases rapidly in individuals over age 50. On the other hand, Class 3 (Delayed sleepers) shows a steady decrease in the probability of an individual showing this profile across the lifespan, suggesting that this specific pattern of poor sleep is more commonly associated with younger adults. Finally, the proportion of Class 4 (poor sleepers) members increases only slightly across the lifespan. Together, the latent class analysis provides additional evidence that the PSQI sum score as an indicator of sleep quality does not fully capture the subtleties of agerelated differences. Age-related changes in sleep patterns are characterized by specific, clustered patterns of sleep problems that cannot be adequately characterized by summation of the component scores. The above analyses show how both a summary measure and individual measures of sleep quality change across the lifespan. Next, we examined the relationships between sleep quality measures (seven components and the global PSQI score) and health variables (specific variables across four domains, as shown in Table 1).

Cognitive health
First, we examined the relationships between sleep quality and seven measures of cognitive health (see Table 1 for details). As can be seen in Figure 3, several relationships exist between measures of cognitive health and measures of sleep quality. We visualise these results using a tile plot (68), as shown in Figure 3.
[Insert Figure 3 here] Each cell shows the numeric effect size (R-squared, 0-100) of the bivariate association between a sleep component and a health outcome, colour coded by the statistical evidence for a relationship using the Bayes Factor. If the parameter estimate is positive, the r-squared value has the symbol '+' added (note the interpretation depends on the nature of the variable, cf. Table 1). The strongest associations were found for poorer Total Sleep, poorer sleep Efficiency and use of Sleep Medication, all associated with poorer performance on cognitive tests. The cognitive abilities most strongly associated with poor sleep are immediate and delayed memory, fluid reasoning and a measure of general cognitive health, ACE-R. Two patterns emerged: First, the strongest predictor across the simple and multiple regressions was for the PSQI Total score. Tentatively this suggests that a cumulative index of sleep problems, rather than any specific pattern of poor sleep, is the biggest risk factor for poorer cognitive performance. Secondly, after controlling for age, the most strongly affected cognitive measure is phonemic fluency, the ability to generate name as many different words as possible starting with a given letter within a minute. Verbal fluency is commonly used as a neuropsychological test (e.g. . Previous work suggests it depends on both the ability to cluster (generating words within a semantic cluster) and to switch (switching between categories), and is especially vulnerable to frontal lobe damage Although modest in size, our findings suggests this task, dependent on multiple executive processes, is particularly affected by poor sleep quality (70). The second strongest association was with the ACE-R, a general cognitive test battery When an interaction term with age was included, no evidence for interactions with age were observed (mean logBF 10 =-2.08, see Supplementary Figure 4), suggesting that the negative associations between sleep and cognitive performance are a constant feature across the lifespan, rather than specifically in elderly individuals. Together this suggests that poor sleep quality is modestly and consistently associated with poorer general cognitive performance across the lifespan, most strongly with semantic fluency.

Neural Health
Using Diffusion Tensor Imaging, we estimated a general index of white matter integrity in 10 tracts (52) (shown in Supplementary Figure 5), by taking the average Fractional Anisotropy in each white matter ROI (see (71) for more information). We use the data from a subsample of 641 individuals (age M=54.87, range 18.48-88.96) who were scanned in a 3T MRI scanner (for more details regarding the pipeline, sequence and processing steps, see (71)). Regressing neural WM ROI's on sleep quality, we find several small effects, with the strongest associations between sleep efficiency and neural health (see Supplementary Figure 6). All effects are such that poorer sleep is associated with poorer neural health, apart from a small effect in the opposite direction for Uncinate and Daytime Dysfunction (BF 10 = 6.20). However, when age is included as a covariate, the negative associations between sleep quality and white matter health are attenuated virtually to zero ( Figure 4, mean/median BF 10 = 0.18/.10), with Bayes Factors providing strong evidence for the lack of associations between sleep quality and white matter integrity. One exception was observed: The use of Sleep Medication is associated with better neural health in the corticospinal tract, a region previously found to be affected by pathological sleep problems such as sleep apnoea (28). However, this effect is very small (BF 10 =3. 24) given the magnitude of the sample and the range of comparisons, so should be interpreted with caution.  Figure 4 here] Finally, we tested for any interactions by including a mean-scaled interaction term (sleep*age, Supplementary Figure 7). This analysis found evidence for a significant interaction, between the Superior Longitudinal Fasciculus (SLF) and Sleep Medication (BF 10 = 13.77), such that better neural health in the SLF was associated with the use of Sleep Medication more strongly in older adults.
Together, these findings suggest that in general, once age is taken into account, self-reported sleep problems in a non-clinical sample are not associated with poorer neural health, although there is some evidence for a modest associations between better neural health in specific tracts and the use of sleep medication in the elderly.

Physical health
Next we examined whether sleep quality is associated with physical health. Figure 5 shows the simple regressions between sleep quality and physical health. Strong associations were found between poor overall sleep (PSQI sum score) and poor self-reported health, both in general (logBF 10 =77.51) and even more strongly for health in the past 12 months (logBF 10 =91.25). This may be because poorer sleep, across all components, directly affects general physical health  or because people subjectively experience sleep quality as a fundamental part of overall general health. A second association was between BMI and poor sleep quality, most strongly poor Duration (logBF 10 =4.69).
[Insert Figure 5 here] This not only replicates previous findings but is in line with an increasing body of evidence that suggests that shorted sleep duration causes metabolic changes, which in turn increases the risk of both diabetes mellitus and obesity (17,73,74). Next, we examined whether these effects were attenuated once age was included. We show that although the relationships are slightly weaker, the overall pattern remains (Supplementary Figure 8), suggesting these associations are not merely co- Duration, is related to differences in physical health outcomes in a healthy sample.
Finally, there was evidence of a single interaction with age (Supplementary Figure 9): Although poor sleep Duration was associated with higher diastolic blood pressure in younger adults, it was associated with lower diastolic blood pressure in older individuals (BF 10 = 8.53). This may reflect the fact that diastolic blood pressure is related to cardiovascular health in a different way across the lifespan, although given the small effect size it should be interpreted with caution.

Mental health
Finally, we examined the relationship between sleep quality and mental health, as measured by the Hospital Anxiety and Depression Scale (48). One benefit of the HADS in this context is that, unlike some other definitions (e.g. the DSM-V), sleep quality is not an integral (scored) symptom of these dimensions. As shown in Supplementary Figure 10, there are very strong relationships between all aspects of sleep quality and measures of both anxiety and depression. The strongest predictors of Depression are Daytime Dysfunction (logBF 10 = 245.9, R^2=20.9%), followed by the overall sleep score (logBF 10 = 170.5, R^2=14.6%) and sleep quality (logBF 10 = 106.8, R^2=9.7%). The effects size for Anxiety was comparable but slightly smaller in magnitude. When age is included as a covariate the relationships remained virtually unchanged (Supplementary Figure 11), suggesting these relationships are present throughout across the lifespan. These findings replicate and extend previous work, suggesting that sleep quality is strongly associated with both anxiety and depression across the lifespan.
Finally we examined a model with an interaction term (Supplementary Figure 12). Most prominently we found interactions with age in the relationship between HADS depression and the PSQI Total, and in the relationship between HADS depression and Sleep Duration, such that for the relationship between anxiety and overall sleep quality is stronger in younger adults (BF 10 =9.91, see  Figure 6). Together our findings show that poor sleep quality is consistently, strongly and stably associated with poorer mental health across the adult lifespan.

DISCUSSION
In this study, we report on the associations between age-related differences in sleep quality and health outcomes in a large, age-heterogeneous sample of community dwelling adults of the  Figure 10) and there is previous evidence from this cohort linking white matter health and cognition, the absence of the relationship between poor sleep and neural health cannot be (fully) explained away by the possible noisiness of self-report measures or white matter measures. For this reason, our study provides a potentially reassuring message that for typically-ageing, healthy individuals, poorer self-reported sleep quality is not associated with poorer brain health.
While there are limitations of self-report measures including in older cohorts (15) In our paper we focus on a healthy, age-heterogeneous community dwelling sample. This allows us to study the associations between healthy aging and self-reported sleep quality, but comes with two key limitations of the interpretations of our findings. First and foremost, our findings are cross-sectional, not longitudinal. This means we can make inferences about age-related differences, but not necessarily age-related changes (Raz & Lindenberger, 2011;Schaie, 1994). One reason why cross-sectional and longitudinal estimates may diverge is that older adults can be thought of as cohorts that differ from the younger adults in more ways than age alone. For example, our age range includes individuals born in the twenties and thirties of the 20th century. Compared to someone born in the 21 st century, these individuals will likely have experience various differences during early life development (e.g. less broadly accessible education, lower quality of healthcare, poorer nutrition and similar patterns). For some of our measures, these are inherent limitations -truly longitudinal study of neural aging is inherently impossible as scanner technology has not been around sufficiently long. This means our findings likely reflect a combination of effects attributable to age-related changes as well as baseline differences between subpopulations that may affect both mean differences as well as developmental trajectories.
Second, our sample reflects an atypical population in the sense that they are willing and able to visit the laboratory on multiple occasions for testing sessions. This subsample is likely a more healthy subset of the full population, which will mean the range of (poor) sleep quality as well as (poorer) health outcomes will likely be less extreme that in the full population. However, this challenge is not specific to our sample. In fact, as the Cam-CAN cohort was developed using stratified sampling based on primary healthcare providers, our sample is likely as population-representative as is feasible for a cohort of this magnitude and phenotypic breadth (see Shafto et al., 2014b) 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59                Results Better sleep is generally associated with better health outcomes, strongly so for 25 mental health, moderately for cognitive and physical health, but not for sleep quality and neural 26 health. Latent Class Analysis identified four sleep types: 'Good sleepers' (68.1%, most frequent in 27 middle age), 'inefficient sleepers' (14.01%, most frequent in old age), 'Delayed sleepers' (9.28%, 28 most frequent in young adults) and 'poor sleepers' (8.5%, most frequent in old age). There is little 29 evidence for interactions between sleep quality and age on health outcomes. Finally, we observe u-30 shaped associations between sleep duration and mental health (depression and anxiety) as well as 31 self-reported general health, such that both short and long sleep were associated with poorer 32

outcomes. 33
Conclusions Lifespan changes in sleep quality are multifaceted and not captured well by 34 summary measures, but instead as partially independent symptoms that vary in prevalence across 35 the lifespan. Better self-reported sleep is associated with better health outcomes, and the strength 36  Sleep is a fundamental human behaviour, with humans spending almost a third of their lives asleep. 56 Regular and sufficient sleep has been shown to benefit human physiology through a number of 57 different routes, ranging from consolidation of memories (1) to removal of free radicals (2) and 58 neurotoxic waste (3). Sleep patterns are known to change across the lifespan in various ways. 59 including decreases in quantity and quality of sleep (4), with up to 50% of older adults report 60 difficulties initiating and/or maintaining sleep (5). A meta-analysis of over 65 studies reflecting 3577 61 subjects across the lifespan reported a complex pattern of changes, including an increase of stage 1 62 but a decrease of stage 2 sleep in old age, as well as a decrease in REM sleep (6). An epidemiological 63 investigation of self-reported sleep in older adults observed marker sex differences in age-related 64 sleep changes, with females more likely to report disturbed sleep onset but men reporting night-65 time awakenings (7). Other findings age-related physiological changes in the alignment of 66 homeostatic and circadian rhythms (8), decreases in sleep efficiency (9) the amount of slow-wave 67 sleep, and an increase in daytime napping (10). Importantly, interruption and loss of sleep has been 68 shown to have wide ranging adverse effects on health (11), leaving open the possibility that age-69 related changes in sleep patterns and quality may contribute to well-documented age-related 70 declines in various health domains. 71 In the current study, we examine self-reported sleep habits in a large, population-based 72 cohort Cambridge Centre for Ageing and Neuroscience (Cam-CAN (12)). We relate sleep measures to 73 measures of health across four health domains: cognitive, brain health, physical and mental health. 74 Our goal is to quantify and compare the associations between typical age-related changes in sleep 75 quality and a range of measures of health measures that commonly decline in later life. We assess 76 sleep using a self-reported measure of sleep quality, the Pittsburgh Sleep Quality Index (PSQI) (13). 77 The PSQI has good psychometric properties (14) and has been shown to correlate reliably with 78 diseases of aging and mortality (15)(16)(17). Although polysomnography (18) is commonly considered 79 the gold standard of sleep quality measurement, it is often prohibitively challenging to employ in 80 collecting self-report sleep quality data in a large, deeply phenotyped cohort offers several 84 additional benefits. 85 By utilising a population cohort of healthy adults, and studying a range of health outcomes in 86 the same population, we can circumvent challenges associated with studying clinical populations 87 and provide new insights. First and foremost, by investigating associations between sleep and 88 outcomes across multiple health domains in the same sample, we can make direct comparisons of 89 the relative magnitude of these effects. Second, larger samples allow us to can generate precise 90 effect size estimates, as well as adduce in favour of the null hypothesis. Third, we investigate the 91 associations between sleep quality and neural health in a uniquely large healthy population. 92 Previous investigations of the consequences of poor sleep on especially neural health have generally 93 focuses on clinical populations such as those suffering from insomnia (20,21). Although such studies 94 are crucial for understanding pathology, the demographic idiosyncrasies and often modest sample 95 sizes of these approaches make it hard to generalize to healthy, community dwelling lifespan 96 populations. Moreover, most studies that study age-related changes or differences focus on (very) 97 old age, while far less is known about young and middle aged adults (6). For these reasons, our focus 98 on a healthy, multimodal lifespan cohort is likely to yield novel insights into the subtle changes in 99 sleep quality across the lifespan. 100 We will focus on three questions within each health domain: First, is there a relationship 101 between sleep quality and health? Second, does the strength and nature of this relationship change 102 when age is included as a covariate? Third, does the strength and nature of the relationship change 103 across the lifespan? We will examine these questions across each of the four health domains. 104 105 106  in their home, with questions on health, lifestyle demographics and core cognitive assessments. 113 Sample size was chosen to allow for 100 participants per decile in further acquisition stages, giving 114 sufficient power to separate age-related change from other sources of individual variation. For 115 additional details of the project protocol see (12,22) and for further details of the Cam-CAN dataset 116 visit http://www.mrc-cbu.cam.ac.uk/datasets/camcan/. A further subset of participants who were 117 MRI compatible with no serious cognitive impairment participated in a neuroimaging session (22)  118 between the 2011 and 2013. Participants included were native English speakers, had normal or 119 corrected to normal vision and hearing, scored 25 or higher on the mini mental state (23). Note that 120 other, more stringent cut-offs are sometimes employed to screen for premorbid dementia, such as a 121 score of 88 or higher in the Addenbrookes Cognitive Examination -Revised (24). For the sake of 122 comprehensiveness we repeated our analyses using this more stringent cut off (ACE-R>88), but 123 observed no noteworthy differences in our findings, so we only report the findings based on the 124 MMSE. Ethical approval for the study was obtained from the Cambridgeshire 2 (now East of England-125 Sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI), a well-validated 132 self-report questionnaire (13,19) designed to assist in the diagnosis of sleep disorders. The questions 133 concern sleep patterns, habits, and lifestyle questions, grouped into seven components, each 134 yielding a score ranging from 0 (good sleep/no problems) to 3 (poor sleep/severe problems), that 135 are commonly summed to a PSQI Total score ranging between 0 and 21, with higher scores 136 reflecting poorer sleep quality. 137

Health Measures 138
Cognitive health. A number of studies have found associations between poor sleep and 139 cognitive decline, including in elderly populations. Poor sleep affects cognitive abilities such as 140 executive functions (25) and learning and memory processes (26), whereas short term 141 pharmaceutical interventions such as administration of melatonin improve both sleep quality and 142 cognitive performance (27,28). Recent work (29) concluded that "maintaining good sleep quality, at 143 least in young adulthood and middle age, promotes better cognitive functioning and serves to 144 protect against age-related cognitive declines". As sleep may affect various aspects of cognition 145 differently (30), we include measures that cover a range of cognitive domains including memory, 146 reasoning, response speed, and verbal fluency, as well as including a measure of general cognition 147 (See Table 1 and (12)  showing low-level neural pathology (although some study grey matter, e.g. (36,37). White matter 155 hyperintensities are often used as a clinical marker, as longitudinal increases in WMHs are 156 associated with increased risk of stroke, dementia and death (38) and are more prevalent in patients 157  clinical, healthy populations. To address this question, we use a more general indicator of white 160 matter neural health; Fractional Anisotropy (FA). FA is associated with white matter integrity and 161 myelination (39,40). We use FA as recent evidence suggests that WMHs represent the extremes 162 (foci) of white matter damage, and that FA is able to capture the full continuum of white matter 163 integrity (41). For more information regarding the precise white matter pipeline, see (12,22,42). 164 Physical health. Sleep quality is also an important marker for physical health, with poorer 165 sleep being associated with conditions such as obesity, diabetes mellitus (43), overall health (11,44)  166 and increased all-cause mortality (45,46). We focus on a set of variables that capture three types of 167 health domains commonly associated with poor sleep: Cardiovascular health measured by pulse, 168 systolic and diastolic blood pressure (47) sleep and depression, such that individuals regularly sleeping less than 6, or more than 8, hours were 175 more likely to be depressed. Both depression (53) and anxiety (54,55)

STATISTICAL ANALYSES 184
We examine whether self-reported sleep patterns change across the lifespan, both for the PSQI sum 185 score and for each of the seven PSQI components. We then examine the relationships between the 186 sleep quality and the four health domains in three ways: First, simple regression of the health 187 outcome on sleep variables, to determine evidence for association between poor sleep quality and 188 poor health outcomes. Second, we include age as a covariate. Finally, we include a (standard normal 189 rescaled) continuous interaction term to examine whether there is evidence for a changing 190 relationship between sleep and outcomes across the lifespan. 191 For all regressions we will use a default Bayesian approach advocated by (62-65) which  192 avoids several well-documented issues with p-values (64), allows for quantification of null effects, 193 and decreases the risk of multiple comparison problems (66). Bayesian regressions allows us to 194 symmetrically quantify evidence in favour of, or against, some substantive model as compared to a 195 baseline (e.g. null) model. This evidentiary strength is expressed as a Bayes Factor (67) We report log Bayes Factors for (very) large effects and regular Bayes Factors for smaller 203 effects. To compute Bayes Factors we will use Default Bayes Factor approach for model selection 204 (62,63) in the package BayesFactor (68) using the open source software package R (69). As previous 205 papers report associations between sleep and outcomes ranging from absent to considerable in size 206 we utilize the default, symmetric Cauchy prior with width √ଶ ଶ which translates to a 50% confidence 207 that the true effect will lie between -.707 and .707. Prior to further analysis, scores on all outcomes 208 were transformed to a standard normal distribution, and any scores exceeding a z-score of 4 or -4 209 Age-related differences in sleep quality 214 First, we examined sleep changes across the lifespan by examining age-related differences in the 215 PSQI sum score (N= 2178, M=5.16, SD=3.35, Range=0-19). Regressing the PSQI global score on age, 216 (see Supplementary Figure 1) showed evidence for a positive relationship across the lifespan 217 (logBF 10 = 10.45). This suggests that on the whole, sleep quality decreases across the lifespan (note 218 that higher PSQI scores correspond to worse sleep). Although we observe strong statistical evidence 219 for an age-related difference ('Extreme' according to (70)) age explained only 1.23 % of the variance 220 in the PSQI Total score. Next, we examined each of the seven components on age in the same 221 manner. In Supplementary Figure 2 we see that that age has varying and specific effects on different 222 aspects of sleep quality, and did not worsen uniformly across the lifespan. For example, we observed 223 moderate evidence that sleep latency did not change across the lifespan (Sleep Latency, BF 01 = 9.25, 224 in favour of the null), Sleep Quality showed no evidence for either change or stasis (BF 10 = 1.63) and 225 one sleep component, Daytime Dysfunction, improved slightly across the lifespan (BF 10 = 7.03). 226 Medication). The strongest age-related decline is that of Efficiency, showing an R-squared of 6.6%. 227 Finally, we entered all seven components into a Bayesian multiple regression 228 simultaneously, to examine to what extent they could, together, predict age. The best model 229 included every component except Sleep Latency (logBF 10 = 142.71). Interestingly, this model 230 explained 13.66% of the variance in age, compared to 1.23% for the PSQI Total score, and 6.6% for 231 the strongest single component (efficiency). This shows that lifespan changes in self-reported sleep 232 are heterogeneous and partially independent, and that specific patterns and components need to be 233 taken into account simultaneously to fully understand age-related differences in sleep quality. These 234 'fairly bad' and 'very bad' on the other. As analytical work in psychometrics (72) suggests that likert-251 like graded scales can be treated as continuous only from five ordinal categories upwards, by fitting 252 an LCA we are erring on the side of caution (although a latent profile analysis would likely give 253 similar results). Note that although our analysis divides individuals into discrete classes with specific 254 profiles, it is still possible to examine the conditional response likelihood of responding 'yes' to each 255 symptom as a continuous metric (between 0 and 1) that reflects the nature of the association 256 between the class and the outcome. By modelling sleep 'types' we hope to illustrate the complex 257 patterns in a more intelligible manner -notably, doing so allows us to examine whether the 258 likelihood of belonging to any sleep 'type' changes as a function of age. 259  (5 classes see (71). The frequency of Class 1 (Good sleepers) peaks in middle to late adulthood, dropping 285 individuals, but the prevalence increases rapidly in individuals over age 50. On the other hand, Class 287 3 (Delayed sleepers) shows a steady decrease in the probability of an individual showing this profile 288 across the lifespan, suggesting that this specific pattern of poor sleep is more commonly associated 289 with younger adults. Finally, the proportion of Class 4 (poor sleepers) members increases only 290 slightly across the lifespan. Together, the latent class analysis provides additional evidence that the 291 PSQI sum score as an indicator of sleep quality does not fully capture the subtleties of age-related 292 differences. Age-related changes in sleep patterns are characterized by specific, clustered patterns 293 of sleep problems that cannot be adequately characterized by summation of the component scores. 294 The above analyses show how both a summary measure and individual measures of sleep quality 295 change across the lifespan. Next, we examined the relationships between sleep quality measures 296 (seven components and the global PSQI score) and health variables (specific variables across four 297 domains, as shown in Table 1). 298 299

Sleep, health domains and age 300
Cognitive health 301 First, we examined the relationships between sleep quality and seven measures of cognitive health 302 (see Table 1 for details). We visualize our findings using tileplots (75). Each cell shows the numeric 303 effect size (R-squared, 0-100) of the bivariate association between a sleep component and a health 304 outcome, colour coded by the statistical evidence for a relationship using the Bayes Factor. If the 305 parameter estimate is positive, the r-squared value has the symbol '+' added (note the 306 interpretation depends on the nature of the variable, cf. Table 1). 307 As can be seen in Supplementary Figure 3, several relationships exist between measures of cognitive 308 health and measures of sleep quality. However, these results attenuate in a multiple regression 309 model including age as shown in Figure 3.   16 The cognitive abilities most strongly associated with poor sleep are a measure of general cognitive 312 health, ACE-R, and a test of verbal phonemic fluency. Two patterns emerged: First, the strongest 313 predictor across the simple and multiple regressions was for the PSQI Total score. Tentatively this 314 suggests that a cumulative index of sleep problems, rather than any specific pattern of poor sleep, is 315 the biggest risk factor for poorer cognitive performance. Secondly, after controlling for age, the most 316 strongly affected cognitive measure is phonemic fluency, the ability to generate name as many 317 different words as possible starting with a given letter within a minute. Verbal fluency is commonly 318 used as a neuropsychological test (76). Previous work suggests it depends on both the ability to 319 cluster (generating words within a semantic cluster) and to switch (switching between categories), 320 and is especially vulnerable to frontal and temporal lobe damage (with specific regions dependant 321 on either a semantic or phonemic task (77)). Although modest in size, our findings suggests this task, 322 dependent on multiple executive processes, is particularly affected by poor sleep quality (78). The 323 second strongest association was with the ACE-R, a general cognitive test battery similar in style and 324 content to the MMSE. When an interaction term with age was included, little evidence for 325 interactions with age (mean logBF 10 =-2.08, see Supplementary Figure 4), suggesting that the 326 negative associations between sleep and cognitive performance are a constant feature across the 327 lifespan, rather than specifically in elderly individuals. Together this suggests that poor sleep quality 328 is modestly but consistently associated with poorer general cognitive performance across the 329 lifespan, most strongly with semantic fluency. 330 331 Neural Health 332 Using Diffusion Tensor Imaging, we estimated a general index of white matter integrity in 10 tracts 333 (59) (shown in Supplementary Figure 5), by taking the average Fractional Anisotropy in each white 334 matter ROI (see (79) for more information). We use the data from a subsample of 641 individuals 335 (age M=54.87, range 18.48-88.96) who were scanned in a 3T MRI scanner (for more details regarding 336 the pipeline, sequence and processing steps, see (22,79  Together, these findings suggest that in general, once age is taken into account, self-reported sleep 354 problems in a non-clinical sample are not associated with poorer neural health, although there is 355 some evidence for a modest associations between better neural health in specific tracts and the use 356 of sleep medication in the elderly. 357

358
Physical health 359 Next we examined whether sleep quality is associated with physical health. Figure 5 shows 360 the simple regressions between sleep quality and physical health. Strong associations were found 361 between poor overall sleep (PSQI sum score) and poor self-reported health, both in general 362 (logBF 10 =77.51) and even more strongly for health in the past 12 months (logBF 10 =91.25). This may 363 This not only replicates previous findings but is in line with an increasing body of evidence 369 that suggests that shorted sleep duration causes metabolic changes, which in turn increases the risk 370 of both diabetes mellitus and obesity (43,81,82). Next, we examined whether these effects were 371 attenuated once age was included. We show that although the relationships are slightly weaker, the 372 overall pattern remains (Supplementary Figure 8), suggesting these associations are not merely co-373 occurences across the lifespan. Our findings suggest self-reported sleep quality, especially sleep 374 Duration, is related to differences in physical health outcomes in a healthy sample. 375 Finally, there was evidence of a single interaction with age (Supplementary Figure 9): 376 Although poor sleep Duration was associated with higher diastolic blood pressure in younger adults, 377 it was associated with lower diastolic blood pressure in older individuals (BF 10 = 8.53). This may 378 reflect the fact that diastolic blood pressure is related to cardiovascular health in a different way 379 across the lifespan, although given the small effect size it should be interpreted with caution. 380 381

Mental health 382
Finally, we examined the relationship between sleep quality and mental health, as measured by the 383 Hospital Anxiety and Depression Scale (56). One benefit of the HADS in this context is that, unlike 384 some other definitions (e.g. the DSM-V), sleep quality is not an integral (scored) symptom of these 385 dimensions. As shown in Supplementary Figure 10 Depression are Daytime Dysfunction (logBF 10 = 245.9, R^2=20.9%), followed by the overall sleep 388 score (logBF 10 = 170.5, R^2=14.6%) and sleep quality (logBF 10  Anxiety was comparable but slightly smaller in magnitude. When age is included as a covariate the 390 relationships remained virtually unchanged (Supplementary Figure 11), suggesting these 391 relationships are present throughout across the lifespan. These findings replicate and extend 392 previous work, suggesting that sleep quality is strongly associated with both anxiety and depression 393 across the lifespan. 394 Finally we examined a model with an interaction term (Supplementary Figure 12). Most 395 prominently we found interactions with age in the relationship between HADS depression and the 396 PSQI Total, and in the relationship between HADS depression and Sleep Duration, such that for the 397 relationship between anxiety and overall sleep quality is stronger in younger adults (BF 10 =9.91, see 398 Figure 6). Together our findings show that poor sleep quality is consistently, strongly and stably 399 associated with poorer mental health across the adult lifespan. 400 [Insert Figure 6 here] 401 402

Non-linear associations between sleep and health outcomes 403
In the above analyses, we focused on linear associations between symptoms and health outcomes. 404 However, for one aspect of sleep, namely sleep duration (in hours), evidence exists that these 405 associations are likely to be non-linear, such that both shorter and longer than average sleep are 406 associated with poorer health outcomes (e.g. (83)(84)(85). This is echoed in clinical criteria for 407 depression, which commonly include that include both hyper-and hypo-somnia as 'sleep disruption' 408 symptoms -In other words, both too much or too little sleep are suboptimal. To examine whether 409 we observe evidence for non-linearities we examined the relationship between raw scores on sleep 410 duration (in hours, not transformed to PSQI norms) and health outcomes across the four domains. If 411 the association between sleep and outcomes is indeed u-shaped (or inverted U, depending on the 412 scale) then a Bayesian regression would prefer the less parsimonious model that includes the 413 quadratic term. We observed no non-linear associations between any neural or cognitive health 414 variables. We find strong evidence for a quadratic (subscript q) over a linear (subscript l) associations 415  Figure 7A shows the strongest curvilinear association, namely with 417 depression). We find a similar u-shaped curve with general health (BF ql = 277.81) and self-reported 418 health over the last 12 months (BF ql =887.59), the latter shown in Figure 7b. Together, these analyses 419 support previous conclusions that some (although not all) poorer health outcomes can be associated 420 with both too much and too little sleep. 421 [Insert Figure 7 here] 422

DISCUSSION 423
In this study, we report on the associations between age-related differences in sleep quality and 424 health outcomes in a large, age-heterogeneous sample of community dwelling adults of the 425 Cambridge Neuroscience and Aging (Cam-CAN) cohort. We find that sleep quality generally 426 decreases across the lifespan, most strongly for sleep Efficiency. However age-related changes in 427 sleep patterns are complex and multifaceted, so we used Latent Class Analysis to identify 'sleep 428 types' associated with specific sleep quality profiles. We found that Younger adults are more likely 429 than older adults to display a pattern of sleep problems characterised by poor sleep quality and 430 longer sleep latency, whereas older adults are more likely to display inefficient sleeping, 431 characterised by long periods spent in bed whilst not asleep. Moreover, the probability of being a 432 'good' sleeper, unaffected by any adverse sleep symptoms, decreases considerably after age fifty. 433 Notably, closer investigation of the sleep classes reveals likely further complexities of age-434 related differences. The category 'poor sleepers', most prevalent in older adults, shows high 435 conditional likelihood of 'poor sleep' across all symptoms except 'daytime dysfunction'. One possible 436 explanation is that almost all individuals in this group are beyond retirement age. For this reason, 437 they likely have greater flexibility in tailoring their day to day activities to their energy levels (as 438 opposed to individuals working fulltime), and are therefore less likely to consider themselves 439 'disrupted' even in the presence of suboptimal sleep. Although more detailed, interview-based 440 investigations would be necessary to examine the precise nature of these findings, it stands to 441 reason that certain symptoms change not just in prevalence but also in meaning across the lifespan. 442 One key strength of our broad phenotypic assessment allows for direct comparison of the 443 different measures of sleep quality and four key health domains. We find strongest associations 444 between sleep quality and mental health, moderate relations between sleep quality and physical 445 health and cognitive health and sleep, virtually all such that poorer sleep is associated with poorer 446 health outcomes. We did not find evidence for associations between self-reported sleep and neural 447 health. Notably, the relationships we observe are mostly stable across the lifespan, affecting 448 younger and older individuals alike. A notable exception to these effects is the absence of any strong 449 relation (after controlling for age) between sleep quality and neural health as indexed by tract-based 450 average fractional anisotropy. Perhaps surprisingly, given we found strong relationships in the same 451 sample between sleep and other outcomes (e.g. mental health, Figure 10) we find that self-reported 452 sleep problems in a non-clinical sample are not associated with fractional anisotropy above and 453 beyond old age. This is despite the fact that previous work within the same cohort observed 454 moderate to strong associations between white matter and various cognitive outcomes (42,86,87). 455 However, although notable, our finding does not rule out that such associations do exist with other 456 white matter metrics, that they would be observed with objective measures of sleep such as 457 polysomnography, or that the co-occurrence of age-related declines in sleep quality and white 458 matter share an underlying causal association that cannot be teased apart in a cross-sectional 459 sample. 460 One strength of our study is the assessment of neuroimaging metrics, namely fractional 461 anisotropy, in a large, community-dwelling healthy population. Fractional anisotropy is often used in 462 studies of aging (e.g. Madden, is relatively reliable (88)) and is sensitive to clinical anomalies such as 463 white matter hyperintensities. However, the relationship between FA and white-matter health is 464 indirect (40,89) and drawbacks include its inability to distinguish crossing fibers (e.g. (40,89) and 465 vulnerability to movement and the fact that it likely reflects a combination of underlying 466  (93). Although our findings are in keeping with previous findings, our cross-sectional 487 sample cannot tease apart the causal direction of the observed associations, more work remains to 488 be done to disentangle these complex causal pathways. 489 In our paper we focus on a healthy, age-heterogeneous community dwelling sample. This 490 allows us to study the associations between healthy aging and self-reported sleep quality, but comes 491 with two key limitations of the interpretations of our findings. First and foremost, our findings are 492 estimates may diverge is that older adults can be thought of as cohorts that differ from the younger 495 adults in more ways than age alone. For example, our age range includes individuals born in the 496 twenties and thirties of the 20th century. Compared to someone born in the 21 st century, these 497 individuals will likely have experience various differences during early life development (e.g. less 498 broadly accessible education, lower quality of healthcare, poorer nutrition and similar patterns). For 499 some of our measures, these are inherent limitations -truly longitudinal study of neural aging is 500 inherently impossible as scanner technology has not been around sufficiently long. This means our 501 findings likely reflect a combination of effects attributable to age-related changes as well as baseline 502 differences between subpopulations that may affect both mean differences as well as 503 developmental trajectories. 504 Second, our sample reflects an atypical population in the sense that they are willing and able 505 to visit the laboratory on multiple occasions for testing sessions. This subsample is likely a more 506 healthy subset of the full population, which will mean the range of (poor) sleep quality as well as 507 (poorer) health outcomes will likely be less extreme that in the full population. However, this 508 challenge is not specific to our sample. In fact, as the Cam-CAN cohort was developed using stratified 509 sampling based on primary healthcare providers, our sample is likely as population-representative as 510 is feasible for a cohort of this magnitude and phenotypic breadth (see (12) for further details). 511 Nonetheless, a healthier subsample may lead to restriction of range (96)  Taken together, our study allows several conclusions. First, although we replicate the age-520 related deterioration in some aspects of sleep quality, other aspects remain stable or even improve. 521 Second, we show that the profile of sleep quality changes across the lifespan. This is important 522 methodologically, as it suggests that PSQI sum scores do not capture the full picture, especially in 523 age-heterogeneous samples. Moreover, it is important from a psychological standpoint: We show 524 that 'sleep quality' is a multidimensional construct and should be treated as such if we wish to 525 understand the complex effects and consequences of sleep quality across the lifespan. Third, 526 moderate to strong relations exist between sleep quality and cognitive, physical and mental health, 527 and these relations largely remain stable across the lifespan. In contrast, we show evidence that in 528 non-clinical populations, poorer self-reported sleep is not reliably associated with poorer neural 529 health. Finally, we find that for absolute sleep duration, we replicate previous findings that both 530 longer and shorter than average amounts of sleep are association with poorer self-reported general 531 health and higher levels of depression and anxiety. 532 Together with previous experimental and longitudinal evidence, our findings suggest that at 533 least some age-related decreases in health outcomes may be due to poorer sleep quality. We show 534 that self-reported sleep quality can be an important indicator of other aspects of healthy functioning 535 throughout the lifespan, especially for mental and general physical health. Our findings suggest 536 accurate understanding of sleep quality is essential in understanding and supporting healthy aging 537 across the lifespan.                         Results Better sleep is generally associated with better health outcomes, strongly so for 25 mental health, moderately for cognitive and physical health, but not for sleep quality and neural 26 health. Latent Class Analysis identified four sleep types: 'Good sleepers' (68.6%, most frequent in 27 middle age), 'inefficient sleepers' (13.05%, most frequent in old age), 'Delayed sleepers' (9.76%, 28 most frequent in young adults) and 'poor sleepers' (8.6%, most frequent in old age). There is little 29 evidence for interactions between sleep quality and age on health outcomes. Finally, we observe u-30 shaped associations between sleep duration and mental health (depression and anxiety) as well as 31 self-reported general health, such that both short and long sleep were associated with poorer 32

outcomes. 33
Conclusions Lifespan changes in sleep quality are multifaceted and not captured well by 34 summary measures, but instead as partially independent symptoms that vary in prevalence across 35 the lifespan. Better self-reported sleep is associated with better health outcomes, and the strength 36  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  Sleep is a fundamental human behaviour, with humans spending almost a third of their lives asleep. 56 Regular and sufficient sleep has been shown to benefit human physiology through a number of 57 different routes, ranging from consolidation of memories (1) to removal of free radicals (2) and 58 neurotoxic waste (3). Sleep patterns are known to change across the lifespan in various ways. 59 including decreases in quantity and quality of sleep (4), with up to 50% of older adults report 60 difficulties initiating and/or maintaining sleep (5). A meta-analysis of over 65 studies reflecting 3577 61 subjects across the lifespan reported a complex pattern of changes, including an increase of stage 1 62 but a decrease of stage 2 sleep in old age, as well as a decrease in REM sleep (6). An epidemiological 63 investigation of self-reported sleep in older adults observed marker sex differences in age-related 64 sleep changes, with females more likely to report disturbed sleep onset but men reporting night-65 time awakenings (7). Other findings age-related physiological changes in the alignment of 66 homeostatic and circadian rhythms (8), decreases in sleep efficiency (9) the amount of slow-wave 67 sleep, and an increase in daytime napping (10). Importantly, interruption and loss of sleep has been 68 shown to have wide ranging adverse effects on health (11), leaving open the possibility that age-69 related changes in sleep patterns and quality may contribute to well-documented age-related 70 declines in various health domains. 71 In the current study, we examine self-reported sleep habits in a large, population-based 72 cohort Cambridge Centre for Ageing and Neuroscience (Cam-CAN (12)). We relate sleep measures to 73 measures of health across four health domains: cognitive, brain health, physical and mental health. 74 Our goal is to quantify and compare the associations between typical age-related changes in sleep 75 quality and a range of measures of health measures that commonly decline in later life. We assess 76 sleep using a self-reported measure of sleep quality, the Pittsburgh Sleep Quality Index (PSQI) (13). 77 The PSQI has good psychometric properties (14) and has been shown to correlate reliably with 78 diseases of aging and mortality (15)(16)(17). Although polysomnography (18)  collecting self-report sleep quality data in a large, deeply phenotyped cohort offers several 84 additional benefits. 85 By utilising a population cohort of healthy adults, and studying a range of health outcomes in 86 the same population, we can circumvent challenges associated with studying clinical populations 87 and provide new insights. First and foremost, by investigating associations between sleep and 88 outcomes across multiple health domains in the same sample, we can make direct comparisons of 89 the relative magnitude of these effects. Second, larger samples allow us to can generate precise 90 effect size estimates, as well as adduce in favour of the null hypothesis. Third, we investigate the 91 associations between sleep quality and neural health in a uniquely large healthy population. 92 Previous investigations of the consequences of poor sleep on especially neural health have generally 93 focuses on clinical populations such as those suffering from insomnia (20,21). Although such studies 94 are crucial for understanding pathology, the demographic idiosyncrasies and often modest sample 95 sizes of these approaches make it hard to generalize to healthy, community dwelling lifespan 96 populations. Moreover, most studies that study age-related changes or differences focus on (very) 97 old age, while far less is known about young and middle aged adults (6). For these reasons, our focus 98 on a healthy, multimodal lifespan cohort is likely to yield novel insights into the subtle changes in 99 sleep quality across the lifespan. 100 We will focus on three questions within each health domain: First, is there a relationship 101 between sleep quality and health? Second, does the strength and nature of this relationship change 102 when age is included as a covariate? Third, does the strength and nature of the relationship change 103 across the lifespan? We will examine these questions across each of the four health domains. in their home, with questions on health, lifestyle demographics and core cognitive assessments. 113 Sample size was chosen to allow for 100 participants per decile in further acquisition stages, giving 114 sufficient power to separate age-related change from other sources of individual variation. For 115 additional details of the project protocol see (12,22) and for further details of the Cam-CAN dataset 116 visit http://www.mrc-cbu.cam.ac.uk/datasets/camcan/. A further subset of participants who were 117 MRI compatible with no serious cognitive impairment participated in a neuroimaging session (22)  118 between the 2011 and 2013. Participants included were native English speakers, had normal or 119 corrected to normal vision and hearing, scored 25 or higher on the mini mental state (23). Note that 120 other, more stringent cut-offs are sometimes employed to screen for premorbid dementia, such as a 121 score of 88 or higher in the Addenbrookes Cognitive Examination -Revised (24). For the sake of 122 comprehensiveness we repeated our analyses using this more stringent cut off (ACE-R>88), but 123 observed no noteworthy differences in our findings, so we only report the findings based on the 124 MMSE. Ethical approval for the study was obtained from the Cambridgeshire 2 (now East of England-125 Cambridge Central) Research Ethics Committee (reference: 10/H0308/50). Participants gave written 126 informed consent. The raw data and analysis code are available upon signing a data sharing request 127 form (see http://www.mrc-cbu.cam.ac.uk/datasets/camcan/ for more detail).  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  Sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI), a well-validated 132 self-report questionnaire (13,19) designed to assist in the diagnosis of sleep disorders. The questions 133 concern sleep patterns, habits, and lifestyle questions, grouped into seven components, each 134 yielding a score ranging from 0 (good sleep/no problems) to 3 (poor sleep/severe problems), that 135 are commonly summed to a PSQI Total score ranging between 0 and 21, with higher scores 136 reflecting poorer sleep quality. 137

Health Measures 138
Cognitive health. A number of studies have found associations between poor sleep and 139 cognitive decline, including in elderly populations. Poor sleep affects cognitive abilities such as 140 executive functions (25) and learning and memory processes (26), whereas short term 141 pharmaceutical interventions such as administration of melatonin improve both sleep quality and 142 cognitive performance (27,28). Recent work (29) concluded that "maintaining good sleep quality, at 143 least in young adulthood and middle age, promotes better cognitive functioning and serves to 144 protect against age-related cognitive declines". As sleep may affect various aspects of cognition 145 differently (30), we include measures that cover a range of cognitive domains including memory, 146 reasoning, response speed, and verbal fluency, as well as including a measure of general cognition 147 (See Table 1 and (12)   clinical, healthy populations. To address this question, we use a more general indicator of white 160 matter neural health; Fractional Anisotropy (FA). FA is associated with white matter integrity and 161 myelination (39,40). We use FA as recent evidence suggests that WMHs represent the extremes 162 (foci) of white matter damage, and that FA is able to capture the full continuum of white matter 163 integrity (41). For more information regarding the precise white matter pipeline, see (12,22,42). 164 Physical health. Sleep quality is also an important marker for physical health, with poorer 165 sleep being associated with conditions such as obesity, diabetes mellitus (43), overall health (11,44) 166 and increased all-cause mortality (45,46). We focus on a set of variables that capture three types of 167 health domains commonly associated with poor sleep: Cardiovascular health measured by pulse, 168 systolic and diastolic blood pressure (47)

STATISTICAL ANALYSES 184
We examine whether self-reported sleep patterns change across the lifespan, both for the PSQI sum 185 score and for each of the seven PSQI components. We then examine the relationships between the 186 sleep quality and the four health domains in three ways: First, simple regression of the health 187 outcome on sleep variables, to determine evidence for association between poor sleep quality and 188 poor health outcomes. Second, we include age as a covariate. Finally, we include a (standard normal 189 rescaled) continuous interaction term to examine whether there is evidence for a changing 190 relationship between sleep and outcomes across the lifespan. 191 For all regressions we will use a default Bayesian approach advocated by (62-65)  We report log Bayes Factors for (very) large effects and regular Bayes Factors for smaller 203 effects. To compute Bayes Factors we will use Default Bayes Factor approach for model selection 204 (62,63) in the package BayesFactor (68) using the open source software package R (69). As previous 205 papers report associations between sleep and outcomes ranging from absent to considerable in size 206 we utilize the default, symmetric Cauchy prior with width √ଶ ଶ which translates to a 50% confidence 207 that the true effect will lie between -.707 and .707. Prior to further analysis, scores on all outcomes 208 were transformed to a standard normal distribution, and any scores exceeding a z-score of 4 or -4 209 Age-related differences in sleep quality 214 First, we examined sleep changes across the lifespan by examining age-related differences in the 215 PSQI sum score (N= 2178, M=5.16, SD=3.35, Range=0-19). Regressing the PSQI global score on age, 216 (see Supplementary Figure 1) showed evidence for a positive relationship across the lifespan 217 (logBF 10 = 10.45). This suggests that on the whole, sleep quality decreases across the lifespan (note 218 that higher PSQI scores correspond to worse sleep). Although we observe strong statistical evidence 219 for an age-related difference ('Extreme' according to (70)) age explained only 1.11 % of the variance 220 in the PSQI Total score. Next, we examined each of the seven components on age in the same 221 manner. In Supplementary Figure 2 we see that that age has varying and specific effects on different 222 aspects of sleep quality, and did not worsen uniformly across the lifespan. For example, we observed 223 moderate evidence that sleep latency did not change across the lifespan (Sleep Latency, BF 01 = 9. 66,224 in favour of the null), Sleep Quality showed no evidence for either change or stasis (BF 10 = 1.64) and 225 one sleep component, Daytime Dysfunction, improved slightly across the lifespan (BF 10 = 7.04). 226 Medication). The strongest age-related decline is that of Efficiency, showing an R-squared of 6.6%. 227 Finally, we entered all seven components into a Bayesian multiple regression 228 simultaneously, to examine to what extent they could, together, predict age. The best model 229 included every component except Sleep Duration (logBF 10 = 142.98). Interestingly, this model 230 explained 13.66% of the variance in age, compared to 1.12% for the PSQI Total score, and 6.6% for 231 the strongest single component (efficiency). This shows that lifespan changes in self-reported sleep 232 are heterogeneous and partially independent, and that specific patterns and components need to be 233 taken into account simultaneously to fully understand age-related differences in sleep quality. These 234 'fairly bad' and 'very bad' on the other. As analytical work in psychometrics (72) suggests that likert-251 like graded scales can be treated as continuous only from five ordinal categories upwards, by fitting 252 an LCA we are erring on the side of caution (although a latent profile analysis would likely give 253 similar results). Note that although our analysis divides individuals into discrete classes with specific 254 profiles, it is still possible to examine the conditional response likelihood of responding 'yes' to each 255 symptom as a continuous metric (between 0 and 1) that reflects the nature of the association 256 between the class and the outcome. By modelling sleep 'types' we hope to illustrate the complex 257 patterns in a more intelligible manner -notably, doing so allows us to examine whether the 258 likelihood of belonging to any sleep 'type' changes as a function of age. 259  individuals, but the prevalence increases rapidly in individuals over age 50. On the other hand, Class 287 3 (Delayed sleepers) shows a steady decrease in the probability of an individual showing this profile 288 across the lifespan, suggesting that this specific pattern of poor sleep is more commonly associated 289 with younger adults. Finally, the proportion of Class 4 (poor sleepers) members increases only 290 slightly across the lifespan. Together, the latent class analysis provides additional evidence that the 291 PSQI sum score as an indicator of sleep quality does not fully capture the subtleties of age-related 292 differences. Age-related changes in sleep patterns are characterized by specific, clustered patterns 293 of sleep problems that cannot be adequately characterized by summation of the component scores. 294 The above analyses show how both a summary measure and individual measures of sleep quality 295 change across the lifespan. Next, we examined the relationships between sleep quality measures 296 (seven components and the global PSQI score) and health variables (specific variables across four 297 domains, as shown in Table 1). 298 299

Sleep, health domains and age 300
Cognitive health 301 First, we examined the relationships between sleep quality and seven measures of cognitive health 302 (see Table 1 for details). We visualize our findings using tileplots (75). Each cell shows the numeric 303 effect size (R-squared, 0-100) of the bivariate association between a sleep component and a health 304 outcome, colour coded by the statistical evidence for a relationship using the Bayes Factor. If the 305 parameter estimate is positive, the r-squared value has the symbol '+' added (note the 306 interpretation depends on the nature of the variable, cf. Table 1). 307 As can be seen in Supplementary Figure 3, several relationships exist between measures of cognitive 308 health and measures of sleep quality. However, these results attenuate in a multiple regression 309 model including age as shown in Figure 3.   16 The cognitive abilities most strongly associated with poor sleep are a measure of general cognitive 312 health, ACE-R, and a test of verbal phonemic fluency. Two patterns emerged: First, the strongest 313 predictor across the simple and multiple regressions was for the PSQI Total score. Tentatively this 314 suggests that a cumulative index of sleep problems, rather than any specific pattern of poor sleep, is 315 the biggest risk factor for poorer cognitive performance. Secondly, after controlling for age, the most 316 strongly affected cognitive measure is phonemic fluency, the ability to generate name as many 317 different words as possible starting with a given letter within a minute. Verbal fluency is commonly 318 used as a neuropsychological test (76). Previous work suggests it depends on both the ability to 319 cluster (generating words within a semantic cluster) and to switch (switching between categories), 320 and is especially vulnerable to frontal and temporal lobe damage (with specific regions dependant 321 on either a semantic or phonemic task (77)). Although modest in size, our findings suggests this task, 322 dependent on multiple executive processes, is particularly affected by poor sleep quality (78). The 323 second strongest association was with the ACE-R, a general cognitive test battery similar in style and 324 content to the MMSE. When an interaction term with age was included, little evidence for 325 interactions with age (mean logBF 10  matter ROI (see (79) for more information). We use the data from a subsample of 641 individuals 335 (age M=54.87, range 18.48-88.96) who were scanned in a 3T MRI scanner (for more details regarding 336 the pipeline, sequence and processing steps, see (22,79  Together, these findings suggest that in general, once age is taken into account, self-reported sleep 354 problems in a non-clinical sample are not associated with poorer neural health, although there is 355 some evidence for a modest associations between better neural health in specific tracts and the use 356 of sleep medication in the elderly. 357

358
Physical health 359 Next we examined whether sleep quality is associated with physical health. Figure 5 shows 360 the simple regressions between sleep quality and physical health. Strong associations were found 361 between poor overall sleep (PSQI sum score) and poor self-reported health, both in general 362 (logBF 10 =77.51) and even more strongly for health in the past 12 months (logBF 10 =91.25). This may 363 This not only replicates previous findings but is in line with an increasing body of evidence 368 that suggests that shorted sleep duration causes metabolic changes, which in turn increases the risk 369 of both diabetes mellitus and obesity (43,81,82). Next, we examined whether these effects were 370 attenuated once age was included. We show that although the relationships are slightly weaker, the 371 overall pattern remains (Supplementary Figure 8), suggesting these associations are not merely co-372 occurences across the lifespan. Our findings suggest self-reported sleep quality, especially sleep 373 Duration, is related to differences in physical health outcomes in a healthy sample. 374 Finally, there was evidence of a single interaction with age (Supplementary Figure 9): 375 Although poor sleep Duration was associated with higher diastolic blood pressure in younger adults, 376 it was associated with lower diastolic blood pressure in older individuals (BF 10 = 8.43). This may 377 reflect the fact that diastolic blood pressure is related to cardiovascular health in a different way 378 across the lifespan, although given the small effect size it should be interpreted with caution. 379 380

Mental health 381
Finally, we examined the relationship between sleep quality and mental health, as measured by the 382 Hospital Anxiety and Depression Scale (56). One benefit of the HADS in this context is that, unlike 383 some other definitions (e.g. the DSM-V), sleep quality is not an integral (scored) symptom of these 384 dimensions. As shown in Supplementary Figure 10, there are very strong relationships between all 385 aspects of sleep quality and measures of both anxiety and depression. The strongest predictors of 386 Depression are Daytime Dysfunction (logBF 10 = 245.9, R^2=19.26%), followed by the overall sleep 387 score (logBF 10 = 170.5, R^2=14.92%) and sleep quality (logBF 10 = 106.8, R^2=8.9%). The effects size for 388 Anxiety was comparable but slightly smaller in magnitude. When age is included as a covariate the 389  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  Finally we examined a model with an interaction term (Supplementary Figure 12). Most 394 prominently we found interactions with age in the relationship between HADS depression and the 395 PSQI Total, and in the relationship between HADS depression and Sleep Duration, such that for the 396 relationship between anxiety and overall sleep quality is stronger in younger adults (BF 10 =9.91, see 397 Figure 6). Together our findings show that poor sleep quality is consistently, strongly and stably 398 associated with poorer mental health across the adult lifespan. 399 [Insert Figure 6 here] 400 401

Non-linear associations between sleep and health outcomes 402
In the above analyses, we focused on linear associations between symptoms and health outcomes. 403 However, for one aspect of sleep, namely sleep duration (in hours), evidence exists that these 404 associations are likely to be non-linear, such that both shorter and longer than average sleep are 405 associated with poorer health outcomes (e.g. (83)(84)(85). This is echoed in clinical criteria for 406 depression, which commonly include that include both hyper-and hypo-somnia as 'sleep disruption' 407 symptoms -In other words, both too much or too little sleep are suboptimal. To examine whether 408 we observe evidence for non-linearities we examined the relationship between raw scores on sleep 409 duration (in hours, not transformed to PSQI norms) and health outcomes across the four domains. If 410 the association between sleep and outcomes is indeed u-shaped (or inverted U, depending on the 411 scale) then a Bayesian regression would prefer the less parsimonious model that includes the 412 quadratic term. We observed no non-linear associations between any neural or cognitive health 413 variables. We find strong evidence for a quadratic (subscript q) over a linear (subscript l) associations 414 between sleep duration and HADS anxiety (logBF ql = 19.98), even more strongly so with HADS 415  Figure 7A shows the strongest curvilinear association, namely with 416 depression). We find a similar u-shaped curve with general health (BF ql = 277.81) and self-reported 417 health over the last 12 months (BF ql =887.59), the latter shown in Figure 7b. Together, these analyses 418 support previous conclusions that some (although not all) poorer health outcomes can be associated 419 with both too much and too little sleep. 420 [Insert Figure 7 here] 421

DISCUSSION 422
In this study, we report on the associations between age-related differences in sleep quality and 423 health outcomes in a large, age-heterogeneous sample of community dwelling adults of the 424 Cambridge Neuroscience and Aging (Cam-CAN) cohort. We find that sleep quality generally 425 decreases across the lifespan, most strongly for sleep Efficiency. However age-related changes in 426 sleep patterns are complex and multifaceted, so we used Latent Class Analysis to identify 'sleep 427 types' associated with specific sleep quality profiles. We found that Younger adults are more likely 428 than older adults to display a pattern of sleep problems characterised by poor sleep quality and 429 longer sleep latency, whereas older adults are more likely to display inefficient sleeping, 430 characterised by long periods spent in bed whilst not asleep. Moreover, the probability of being a 431 'good' sleeper, unaffected by any adverse sleep symptoms, decreases considerably after age fifty. 432 Notably, closer investigation of the sleep classes reveals likely further complexities of age-433 related differences. The category 'poor sleepers', most prevalent in older adults, shows high 434 conditional likelihood of 'poor sleep' across all symptoms except 'daytime dysfunction'. One possible 435 explanation is that almost all individuals in this group are beyond retirement age. For this reason, 436 they likely have greater flexibility in tailoring their day to day activities to their energy levels (as 437 opposed to individuals working fulltime), and are therefore less likely to consider themselves 438 'disrupted' even in the presence of suboptimal sleep. Although more detailed, interview-based 439 investigations would be necessary to examine the precise nature of these findings, it stands to 440 reason that certain symptoms change not just in prevalence but also in meaning across the lifespan. 441  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  One key strength of our broad phenotypic assessment allows for direct comparison of the 442 different measures of sleep quality and four key health domains. We find strongest associations 443 between sleep quality and mental health, moderate relations between sleep quality and physical 444 health and cognitive health and sleep, virtually all such that poorer sleep is associated with poorer 445 health outcomes. We did not find evidence for associations between self-reported sleep and neural 446 health. Notably, the relationships we observe are mostly stable across the lifespan, affecting 447 younger and older individuals alike. A notable exception to these effects is the absence of any strong 448 relation (after controlling for age) between sleep quality and neural health as indexed by tract-based 449 average fractional anisotropy. Perhaps surprisingly, given we found strong relationships in the same 450 sample between sleep and other outcomes (e.g. mental health, Figure 10) we find that self-reported 451 sleep problems in a non-clinical sample are not associated with fractional anisotropy above and 452 beyond old age. This is despite the fact that previous work within the same cohort observed 453 moderate to strong associations between white matter and various cognitive outcomes (42,86,87). 454 However, although notable, our finding does not rule out that such associations do exist with other 455 white matter metrics, that they would be observed with objective measures of sleep such as 456 polysomnography, or that the co-occurrence of age-related declines in sleep quality and white 457 matter share an underlying causal association that cannot be teased apart in a cross-sectional 458 sample. 459 One strength of our study is the assessment of neuroimaging metrics, namely fractional 460 anisotropy, in a large, community-dwelling healthy population. Fractional anisotropy is often used in 461 studies of aging (e.g. Madden, is relatively reliable (88)) and is sensitive to clinical anomalies such as 462 white matter hyperintensities. However, the relationship between FA and white-matter health is 463 indirect (40,89) and drawbacks include its inability to distinguish crossing fibers (e.g. (40,89)  broadly accessible education, lower quality of healthcare, poorer nutrition and similar patterns). For 498 some of our measures, these are inherent limitations -truly longitudinal study of neural aging is 499 inherently impossible as scanner technology has not been around sufficiently long. This means our 500 findings likely reflect a combination of effects attributable to age-related changes as well as baseline 501 differences between subpopulations that may affect both mean differences as well as 502 developmental trajectories. 503 Second, our sample reflects an atypical population in the sense that they are willing and able 504 to visit the laboratory on multiple occasions for testing sessions. This subsample is likely a more 505 healthy subset of the full population, which will mean the range of (poor) sleep quality as well as 506 (poorer) health outcomes will likely be less extreme that in the full population. However, this 507 challenge is not specific to our sample. In fact, as the Cam-CAN cohort was developed using stratified 508 sampling based on primary healthcare providers, our sample is likely as population-representative as 509 is feasible for a cohort of this magnitude and phenotypic breadth (see (12) for further details). 510 Nonetheless, a healthier subsample may lead to restriction of range (96) 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59  60   F  o  r  p  e  e  r  r  e  v  i  e  w  o  n  l  y   24 Taken together, our study allows several conclusions. First, although we replicate the age-519 related deterioration in some aspects of sleep quality, other aspects remain stable or even improve. 520 Second, we show that the profile of sleep quality changes across the lifespan. This is important 521 methodologically, as it suggests that PSQI sum scores do not capture the full picture, especially in 522 age-heterogeneous samples. Moreover, it is important from a psychological standpoint: We show 523 that 'sleep quality' is a multidimensional construct and should be treated as such if we wish to 524 understand the complex effects and consequences of sleep quality across the lifespan. Third, 525 moderate to strong relations exist between sleep quality and cognitive, physical and mental health, 526 and these relations largely remain stable across the lifespan. In contrast, we show evidence that in 527 non-clinical populations, poorer self-reported sleep is not reliably associated with poorer neural 528 health. Finally, we find that for absolute sleep duration, we replicate previous findings that both 529 longer and shorter than average amounts of sleep are association with poorer self-reported general 530 health and higher levels of depression and anxiety. 531 Together with previous experimental and longitudinal evidence, our findings suggest that at 532 least some age-related decreases in health outcomes may be due to poorer sleep quality. We show 533 that self-reported sleep quality can be an important indicator of other aspects of healthy functioning 534 throughout the lifespan, especially for mental and general physical health. Our findings suggest 535 accurate understanding of sleep quality is essential in understanding and supporting healthy aging 536 across the lifespan.  1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59 Practise and Research .pdf. 1988. p. 193-213. 617 For peer review only -http://bmjopen.bmj.com/site/about/guidelines.xhtml 1  2  3  4  5  6  7  8  9  10  11  12  13  14  15  16  17  18  19  20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35  36  37  38  39  40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55  56  57  58  59

Study design 4
Present key elements of study design early in the paper 5 Setting 5 Describe the setting, locations, and relevant dates, including periods of recruitment, exposure, follow-up, and data collection 6 Participants 6 (a) Give the eligibility criteria, and the sources and methods of selection of participants. Describe methods of follow-up 6 (b) For matched studies, give matching criteria and number of exposed and unexposed NA  Discuss the generalisability (external validity) of the study results 25-26

Funding 22
Give the source of funding and the role of the funders for the present study and, if applicable, for the original study on which the present article is based 3 *Give information separately for cases and controls in case-control studies and, if applicable, for exposed and unexposed groups in cohort and cross-sectional studies.