The brain structure and genetic mechanisms underlying the nonlinear association between sleep duration, cognition and mental health

Abstract

Sleep duration, psychiatric disorders and dementias are closely interconnected in older adults. However, the underlying genetic mechanisms and brain structural changes are unknown. Using data from the UK Biobank for participants primarily of European ancestry aged 38–73 years, including 94% white people, we identified a nonlinear association between sleep, with approximately 7 h as the optimal sleep duration, and genetic and cognitive factors, brain structure, and mental health as key measures. The brain regions most significantly underlying this interconnection included the precentral cortex, the lateral orbitofrontal cortex and the hippocampus. Longitudinal analysis revealed that both insufficient and excessive sleep duration were significantly associated with a decline in cognition on follow up. Furthermore, mediation analysis and structural equation modeling identified a unified model incorporating polygenic risk score (PRS), sleep, brain structure, cognition and mental health. This indicates that possible genetic mechanisms and brain structural changes may underlie the nonlinear relationship between sleep duration and cognition and mental health.

Full text

Articles
https://doi.org/10.1038/s43587-022-00210-2
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China. 2Key Laboratory of Computational Neuroscience
and Brain-Inspired Intelligence, Fudan University, Ministry of Education, Shanghai, China. 3Department of Psychiatry, University of Cambridge, Cambridge,
UK. 4Behavioural and Clinical Neuroscience Institute, University of Cambridge, Cambridge, UK. 5MOE Frontiers Center for Brain Science, Fudan University,
Shanghai, China. 6Department of Neurology and Institute of Neurology, Huashan Hospital, Shanghai Medical College, Fudan University, Shanghai,
China. 7Fudan ISTBI—ZJNU Algorithm Centre for Brain-Inspired Intelligence, Zhejiang Normal University, Jinhua, China. 8Shanghai Medical College and
Zhongshan Hosptital Immunotherapy Technology Transfer Center, Shanghai, China. 9Department of Neurology, Huashan Hospital, Fudan University,
Shanghai, China. 10Zhangjiang Fudan International Innovation Center, Shanghai, China. 11Department of Computer Science, University of Warwick,
Coventry, UK. 12School of Data Science, Fudan University, Shanghai, China. 13These authors contributed equally: Yuzhu Li, Barbara J. Sahakian.
✉e-mail: wcheng@fudan.edu.cn; jianfeng64@gmail.com
Sleep serves critical functions in cognitive processing and
maintenance of psychological health, including consolidation
of memories1 and emotion processing2. Sleep also provides a
critical neuroprotective function through the clearance of waste
products3. Changes in sleep duration, a critical sleep characteris-
tic, have been linked to several diseases and psychiatric disorders,
including cardio-cerebral vascular disease and dementia4–6. Sleep
duration of less than 4–5 h per night is associated with increased
mortality7. Prolonged sleep duration has been recognized as a
potential marker of incident dementia8.
Alteration in sleep patterns, including difficulty falling asleep and
staying asleep, decreased quantity and quality of sleep and decreased
sleep efficiency are important characteristics of the aging process9–11.
Therefore, sleep disturbances are prevalent in the aging population
and may be accompanied by cognitive decline and poorer well-
being12,13. Relevant to this, a recent study showed an inverted U-shaped
association between sleep duration and global cognitive decline, with
sleep duration less than 4 h or more than 10 h being detrimental14.
In addition, a U-shaped association was observed between nocturnal
sleep duration and cerebrospinal fluid (CSF) biomarkers of amyloid
deposition in older adults, with optimal sleep duration around 6 h15.
However, the optimal level of sleep duration and its relationship
with genetics and brain mechanisms in addition to cognition and
mental health in a large cohort remains to be determined.
Abnormal sleep is associated with detrimental changes in brain
structures in older populations. Previous studies showed that each
hour of reduced sleep duration was associated with a 0.59% increase
in ventricular volume in participants aged over 55 years16. Shorter
total sleep duration in middle-aged and older adults was related to
impairment in white matter microstructure17. A longitudinal study
showed that age-related atrophy of the brain regions involved in
sleep regulation may contribute to the emergence of sleep dis orders
in the aging population18. Despite some previous discussion of
possible nonlinear relationships between sleep and behavioral
measures19–21, the previous reports considering sleep duration and
brain structure were focused on linear relationships.
To address whether brain and genetic mechanisms underlie
the nonlinear association between sleep duration, cognition and
mental health, this study focused on the sleep durations of mid-to-
late life adults using the large cohort of the UK Biobank. This large
cohort enables us to precisely determine the interaction between
age and sleep. The UK Biobank is a large-scale database contain-
ing cognitive assessments, mental health questionnaires (MHQs)
and brain imaging and in-depth genetic information from partici-
pants in the UK. The objectives of the current study were fourfold.
First, to investigate whether a nonlinear association exists between
sleep duration and various mental health conditions and cognitive
performance. Second, to investigate the relationship between sleep
duration and brain structure using neuroimaging data. Third, to
explore the relationship between sleep, PRS, brain structure, mental
health and cognitive functioning using structural equation model-
ing. Finally, to test the directional and direct association of sleep
The brain structure and genetic mechanisms
underlying the nonlinear association between
sleep duration, cognition and mental health
Yuzhu Li1,2,13, Barbara J. Sahakian1,3,4,13, Jujiao Kang1,2,5, Christelle Langley 1,3,4, Wei Zhang1,2,5,
Chao Xie1,2, Shitong Xiang1,2, Jintai Yu 1,6, Wei Cheng 1,2,7,8,9 ✉ and Jianfeng Feng 1,2,5,7,10,11,12 ✉
Sleep duration, psychiatric disorders and dementias are closely interconnected in older adults. However, the underlying genetic
mechanisms and brain structural changes are unknown. Using data from the UK Biobank for participants primarily of European
ancestry aged 38–73 years, including 94% white people, we identified a nonlinear association between sleep, with approxi-
mately 7 h as the optimal sleep duration, and genetic and cognitive factors, brain structure, and mental health as key measures.
The brain regions most significantly underlying this interconnection included the precentral cortex, the lateral orbitofrontal
cortex and the hippocampus. Longitudinal analysis revealed that both insufficient and excessive sleep duration were signifi-
cantly associated with a decline in cognition on follow up. Furthermore, mediation analysis and structural equation modeling
identified a unified model incorporating polygenic risk score (PRS), sleep, brain structure, cognition and mental health. This
indicates that possible genetic mechanisms and brain structural changes may underlie the nonlinear relationship between sleep
duration and cognition and mental health.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
duration, cognition and mental health via longitudinal analysis
and mediation analysis. We hypothesized that there is a nonlinear
association between sleep duration and mental health, cognition
and brain structure. Non-optimal sleep duration would be associ-
ated with subsequent inferior cognitive performance and mental
health symptoms. In addition, we hypothesized that this nonlinear
association between sleep duration and behavioral measures may be
supported by brain and genetic mechanisms.
Results
Population characteristics. Of the 502,536 participants in the UK
Biobank cohort, 498,277 participants aged 38–73 years (55% female)
completed touchscreen questions about sleep duration at baseline,
with mean sleep duration 7.15 ± 1.11 h (mean ± s.d.), of which
48,511 participants aged 44–82 years had data measured at the
follow-up neuroimaging visit (7.15 ± 1.05 h, mean ± s.d.; Extended
Data Fig. 1). The baseline data were assessed between 2006 and
2010, and neuroimaging data were collected from 2014, which
explains the difference in the age range. A total of 156,886 partici-
pants completed an online mental health questionnaire (MHQ) at
follow up. Brain imaging and genetic data of 39,692 participants
were used in the current study. Table 1 showed the demographic
information of the participants used in the study. Fig. 1 provides a
general schema of the current study.
Nonlinear association between sleep duration and key measures.
There were quadratic associations between cognitive function and
sleep duration (cognitive function ~sleep + sleep2, with covariates
adjusted; Extended Data Fig. 2), significant after Bonferroni cor-
rection (P < 0.001), encompassing fluid intelligence (F = 473.5,
P = 9.12 × 10−206), numeric memory (F = 111.5, P = 4.75 × 10−49),
pair matching (F = 79.3, P = 3.72 × 10−35), reaction time (F = 130.6,
P = 2.07 × 10−57) and trail making (F = 26.9, P = 2.14 × 10−12). Sleep
duration illustrated an inverted U shape with fluid intelligence and
numeric memory, and U-shaped associations were found for pair
matching, trail making, prospective memory and reaction time.
This demonstrated the positive association of both insufficient and
excessive sleep duration with inferior performance on cognitive
tasks (Fig. 2a and Table 2).
Sleep duration was significantly quadratically correlated with
multiple follow-up mental health measures (mental health ~sleep +
sleep2, with covariates adjusted) including anxiety symptoms
(F = 1,027.4, P < 1 × 10−300), depressive symptoms (F = 1,013.7,
P < 1 × 10−300), mania symptoms (F = 182.5, P = 6.71 × 10−80), mental
distress (F = 244.0, P = 1.52 × 10−106), psychotic experience (F = 328.3,
P = 4.94 × 10−143), self-harm (F = 444.3, P = 4.02 × 10−193), trauma
(F = 1,106.9, P < 1 × 10−300) and well-being (F = 923.2, P < 1 × 10−300)
after adjusting for covariates. Results were consistent after addi-
tionally adjusting interval years between sleep duration and online
follow-up mental health measurements. Specifically, sleep duration
showed a U-shaped association with anxiety symptoms, depres-
sive symptoms, mental distress, mania symptoms and self-harm
behaviors, whereas well-being showed an inverted U shape. This
indicated that both insufficient and excessive sleep duration were
positively correlated with mental health symptoms (Fig. 2b and
Table 2). Significant nonlinear associations were also revealed
between sleep duration and similar baseline mental health symp-
toms, including depressive symptoms (F = 4,283.0, P < 1 × 10−300),
mania symptoms (F = 668.9, P = 4.24 × 10−290) and well-being
(F = 1,850.4, P < 1 × 10−300). All the associations between sleep
duration and mental health scores were significant after Bonferroni
correction (P < 0.001).
Sleep duration was quadratically associated with brain struc-
tures. To determine how sleep duration and brain structure were
associated, measures of total area, mean thickness, total cortical gray
matter volume and total subcortical gray matter volume were used.
Significant quadratic associations (brain structures ~ sleep + sleep2,
with covariates adjusted) were revealed between sleep duration and
total surface area (left hemisphere (lh), F = 28.88, P = 2.91 × 10−13;
right hemisphere (rh), F = 29.92, P = 1.04 × 10−13), global mean thick-
ness (lh, F = 23.88, P = 4.32 × 10−11; rh, F = 20.42, P = 1.37 × 10−9),
cortical gray matter volume (F = 41.13, P = 1.43 × 10−18) and sub-
cortical gray matter volume (F = 12.64, P = 3.24 × 10−6) (Fig. 3a and
Table 2). Inverted U-shaped associations were found between sleep
duration and the above brain structure measures. Restricted cubic
splines were also used to model the association between sleep dura-
tion and brain structures, which also demonstrated significant non-
linearity (Supplementary Table 5). Therefore, a nonlinear model
analysis was conducted to determine which brain regions had sig-
nificant nonlinear associations with sleep duration with intracranial
volume and other covariates adjusted. The most significant cortical
volumes nonlinearly associated with sleep duration included the
precentral cortex, the superior frontal gyrus (lh), the lateral orbito-
frontal cortex, the pars orbitalis, the frontal pole (lh) and the middle
temporal cortex (with all mentioned regions Bonferroni corrected
(P < 0.05)). Cortical areas of the isthmus cingulate gyrus and cor-
tical thicknesses of the superior frontal gyrus, the rostral middle
frontal gyrus, the superior temporal gyrus, the pars opercularis
and the triangularis and the frontal pole showed the most signifi-
cant nonlinear association with sleep duration (with all mentioned
regions Bonferroni corrected (P < 0.05); Fig. 3b). The correspond-
ing results for the cortical area and thickness are shown in Extended
Data Fig. 3. The subcortical volumes significantly quadratically
Table 1 | Demographic characteristics of participants
Variables Baseline
(2006–2010)
n=498,277
Imaging visit
(2014+)
n=48,511
Online follow
up (2016–2017)
n=156,884
Age (years;
mean ± s.d.) 56.5 ± 8.1 64.2 ± 7.7 63.9 ± 7.7
Sex (female,
percent) 270,814 (54.3%) 25,002 (51.5%) 88,774 (56.6%)
Townsend
deprivation
index
−1.3 ± 3.1 −1.9 ± 2.7 −1.7 ± 2.8
BMI 27.4 ± 4.8 26.6 ± 4.4 26.8 ± 4.6
Educational qualification (%)
Degree level 160,774 (32.3%) 23,493 (48.4%) 70,893 (45.2%)
Other 244,714 (49.1%) 21,662 (44.6%) 73,729 (47%)
Missing data 92,789 (18.6%) 3,356 (7.0%) 12,262 (7.8%)
Smoking status (%)
Never 271,798 (54.5%) 30,023 (61.9%) 90,135 (57.5%)
Previous 172,170 (34.6%) 16,639 (34.3%) 55,143 (35.1%)
Current 52,478 (10.5%) 1,693 (3.5%) 11,309 (7.2%)
Missing data 1,831 (0.4%) 156 (0.3%) 297 (0.2%)
Drinking status (%)
Never 22,018 (4.4%) 1,602 (3.3%) 4,493 (2.9%)
Previous 17,840 (3.6%) 1,640 (3.4%) 4,269 (2.7%)
Current 457,862 (91.9%) 45,251 (93.3%) 148,055 (94.4%)
Missing data 557 (0.1%) 18 (0.037%) 67 (0.043%)
Sleep duration
(h; mean ± s.d.) 7.15 ± 1.11 7.15 ± 1.05 NA
BMI, body mass index; NA, not available.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
correlated with sleep duration included the hippocampus (lh,
F = 13.7, P = 1.10 × 10−6; rh, F = 14.3, P = 6.0 × 10−7), the infe-
rior lateral ventricle (lh, F = 12.9, P = 2.38 × 10−6; rh, F = 11.7,
P = 8.15 × 10−6) and the corpus callosum anterior midbody (F = 11.3,
P = 1.22 × 10−5) (with all mentioned regions Bonferroni corrected
(P < 0.05); Fig. 3c).
Additionally, two-line tests were conducted to identify the break-
point of these quadratic regressions between sleep duration and
cognitive function, mental health symptoms and brain structure.
Consistent with the quadratic regression, the two-line test showed
that 7 h was the breakpoint with an opposite slope sign of regression
between sleep duration below and above 7 h and all these key mea-
sures (Supplementary Table 6). It is worth noting that only integer
values of sleep duration were available in the questionnaire.
The effect of age on the associations between sleep duration and
key measures. We further investigated whether there were differ-
ences in the association pattern between sleep duration, cognition,
mental health and brain structure between groups with different
ages. The participants were divided into three age groups, ensur-
ing approximately the same number of individuals in each sub-
group for baseline behavioral analysis and follow-up neuroimaging
analysis, respectively. The results revealed that, with increasing
age, there was a decrease in brain volume and a greater impairment
in cognitive function. However, the same pattern was not found for
mental health. The nonlinear relationship between sleep duration
and cortical volumes was most significant in the group aged 44–59
years, and the association curve gradually flattened with increas-
ing age (Fage 44–59 = 20.16, Fage 60–67 = 18.36, Fage 68–82 = 10.63), which was
also shown in subcortical volumes (Fage 44–59 = 10.15, Fage 60–67 = 6.79,
Fage 68–82 = 2.25). A similar trend was also observed in the asso-
ciation between sleep duration and several cognitive function and
mental health measures including reaction time (Fage 39–52 = 200.7,
Fage 53–61 = 194.4, Fage 62–70 = 48.8), depressive symptoms (Fage 39–52 = 420.3,
Fage 53–61 = 360.2, Fage 62–70 = 205.5) and mania (Fage 39–52 = 129.2,
Fage 53–61 = 94.0, Fage 62–70 = 47.9) (Fig . 4 and Supplementary Table 7).
These findings showed the frequently reported interaction between
sleep duration and age. Furthermore, ‘
sleep
×
age
’ and ‘
sleep2
×
age
’ terms were added in the nonlinear regression model to validate the
significance of the interaction terms. Details of results are provided
in Supplementary Table 8.
Noticeably, even for the older age groups with a relatively flat-
ter nonlinear curve, the nonlinear associations were still significant
between sleep duration and the above-mentioned cognitive func-
tion, mental health symptoms and cortical volumes compared to a
linear relationship (Supplementary Table 9).
We further explored the interaction between sleep duration and
sex; interaction terms ‘
sleep
×
sex
’ and ‘
sleep2
×
sex
’ were added to
the original nonlinear regression model (Extended Data Fig. 4 and
Supplementary Tables 10 and 11).
The longitudinal association between sleep duration and key
measures. To examine the association between sleep variability and
behavioral assessments, we calculated the difference in sleep dura-
tion between baseline and neuroimaging visits and associated this
difference with cognitive function and mental health. Similar non-
linear associations were found between sleep duration difference
and these assessments, with approximately 0 h being associated with
optimal cognitive performance (F-value range from 6.69 to 50.6, all
P values < 1.2 × 10−3) and mental health (F-value range from 38.9 to
158.9, all P values < 1.4 × 10−17). These findings highlight the close
association of stable sleep duration and health (Fig. 5). Detailed
F and P values are displayed in Supplementary Table 12.
Based on the quadratic associations between sleep duration and
mental health and cognitive function, participants with both lon-
gitudinal sleep and behavioral data were separated into two groups
with sleep duration ≤7 h and sleep duration >7 h. The longitudinal
analysis was conducted for depressive symptoms (Patient Health
Questionnaire (PHQ)-4, fluid intelligence and sleep duration at
baseline and at the neuroimaging visit. Histograms of the changes
in these variables over time are shown in Extended Data Fig. 5.
The results revealed that, for participants with sleep duration ≤7 h
Brain imaging
Cortical volume
Cortical area and thickness
Subcortical volume
distress, well-being
Mental health
Depression, mania,
anxiety, self-harm,
Genomics
PRS of sleep duration
Sleep duration
Baseline Follow up
At baseline (time 1) and
imaging visit (time 2)
Nonlinear association between sleep and other measures
Reaction time
Cognitive function
Fluid intelligence
Numeric memory
Pair matching
Longitudinal and mediation analysis
UK Biobank dataset: n = 498,277
Sleep (h)
Sleep Sleep
Mental
health
Mental
health
Baseline Follow up
Sleep
PRS
Brain
regions
Structural equation model
Cross-lagged panel model
Cognition
14710
1471014 710
0.8
0.55
0.40
0.25
0.6
0.4
Mental health Brain structure
2006–2010 2014+
Cognition Cognition
Cognition
Prospective memory
Quadratic regression model
Mental
health
Fig. 1 | Guideline of the study. Left, UK Biobank data used in the study including brain imaging, mental health, cognitive function, genomics and sleep
duration. Top right, nonlinear association between sleep duration and cognitive function, mental health and brain structure. Bottom right, longitudinal
analysis between sleep duration, cognition and mental health and structural equation model specifying the directional association between PRS, sleep,
mental health, cognition and brain structure.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
(n = 9,753) longer sleep duration was significantly associated with
lower PHQ-4 scores and higher fluid intelligence scores at follow up
(β = −0.038, P = 1.3 × 10−5 and β = 0.021, P = 0.01, Fig. 6a). Baseline
PHQ-4 scores were significantly associated with follow-up fluid
intelligence scores (β = −0.018, P = 0.029). All mentioned longitu-
dinal associations were significant after false discovery rate (FDR)
Cognitive function
Mental health
Fluid intelligence
Anxiety symptoms
Trail making
Numeric memory
1.0
0.5
0.2
0.1
0
1.0
b
0.5
0
Depressive symptoms
1.2
0.6
0
Mania symptoms
1.2
0.6
0
Self-harm
1.2
0.6
0
Well-being
1.2
0.6
0
Mental distress
1.4
0.8
0.2
Prospective memory
1.0
0.5
0
Reaction time
0.4
0.2
0
0
1.0
0.5
0
Pair matching
0.2
0.1
0
1 7 13
Sleep duration (h)
1 7 13
Sleep duration (h)
1 7 13
Sleep duration (h)
1 7 13
Sleep duration (h)
1713
Sleep duration (h)
1713
Sleep duration (h)
1 7 13
Sleep duration (h)
1 7 13
Sleep duration (h)
1713
Sleep duration (h)
1713
Sleep duration (h)
1 7 13
Sleep duration (h)
1 7 13
Sleep duration (h)
a
Fig. 2 | Nonlinear association of sleep duration with mental health and with cognitive function. a, Significant nonlinear association between sleep
duration and cognitive function including fluid intelligence, numeric memory, pair matching, trail making, prospective memory and reaction time
(Bonferroni corrected, P < 0.01). b, Significant nonlinear association between sleep duration and mental health including anxiety symptoms, depressive
symptoms, mental distress, mania symptoms, self-harm and well-being (Bonferroni corrected, P < 0.01). The variables shown in the figure were adjusted
for covariates comprising age, sex, body mass index, Townsend deprivation index, educational qualification, smoking status and drinking status. F-tests
were used to assess statistical significance and derive F statistics and corresponding one-sided P values adjusted for multiple comparisons. Lines are fitted
nonlinear models indicating fitted mean values, and shaded areas are 95% confidence intervals (CIs); gray points are individual data points.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
correction (FDR corrected P < 0.05; Supplementary Table. 13). For
participants with sleep duration of more than 7 h (n = 5,247), longer
sleep duration at baseline correlated with lower fluid intelligence
scores at follow up (β = 0.025, P = 0.027, Extended Data Fig. 6a).
Serial mediation between PRSs of sleep and behavioral measures.
Based on the longitudinal association of baseline sleep duration with
subsequent depressive symptoms and the significant association
between sleep duration and brain structure, we further examined
whether sleep duration and brain structure contributed to the asso-
ciation between PRSs for sleep and behavioral measures. Therefore,
we first conducted three mediation pathway analyses for depres-
sive symptoms, namely (1) PRS → sleep duration → brain struc-
ture → depressive symptoms, (2) PRS → sleep duration → depressive
symptoms and (3) PRS → brain structure → depressive symptoms.
For participants with sleep duration ≤7 h, the results of first
model demonstrated that PRS for sleep had a significant nega-
tive effect on depressive symptoms (β = −0.033, P = 2.3 × 10−4),
PRS was associated with sleep duration (β = 0.058, P = 1.4 × 10−10)
and sleep duration was associated with brain structure (β = 0.046,
P = 3.7 × 10−7), and, in addition, brain structure was associated with
depressive score (β = −0.027, P = 2.2 × 10−3). The indirect pathway
of the effect of PRS on depressive symptoms via sleep duration
and brain structure was significant (Fig. 6b, path β1 = −7.3 × 10−5,
P = 0.025). The results of the second model showed that sleep
duration significantly mediated the association between PRS and
depressive symptoms (Fig. 6b, path β2 = −0.0071, P = 9.3 × 10−9).
The third model revealed that brain structure was also a significant
mediator for the association between PRS and depressive symptoms
(Fig. 6b, path β3 = −7.7 × 10−4, P = 0.035). A mediation analysis of
PRS → depressive symptoms → brain structures → sleep was also
conducted (Extended Data Fig. 7a), with depressive symptoms and
brain structures serially mediating the association between PRS and
sleep duration (β = 4.11 × 10−5, P = 0.049).
Three mediation pathway analyses were also conducted for the
cognitive function of fluid intelligence for participants with sleep
duration ≤7 h: (1) PRS → sleep duration → brain structure → fluid
intelligence, (2) PRS → sleep duration → fluid intelligence and (3)
PRS → brain structure → fluid intelligence. PRSs of sleep showed a
significant positive association with fluid intelligence in the model
(β = 0.044, P = 2.5 × 10−8). The serial mediation pathway via sleep
and brain structure was significant (β1 = 3.8 × 10−4, P = 1.3 × 10−5).
Specifically, sleep duration was significantly associated with PRS
(β = 0.061, P = 1.2 × 10−14), and brain volume was positively associ-
ated with sleep duration (β = 0.045, P = 1.3 × 10−8) and, in addition,
brain volume was significantly associated with fluid intelligence
(β = 0.14, P = 0). The result of model 2 and model 3 showed that
sleep duration and brain structure were also separately significant
mediators for this association (Fig. 6b; β2 = 0.004, P = 2.8 × 10−8;
β3 = 0.004, P = 7.9 × 10−5). Additionally, mediation analyses exami-
ning how depressive symptoms mediated the association between
sleep duration and fluid intelligence were also conducted and are
presented in Extended Data Fig. 7b.
For participants with sleep duration >7 h, mediation analysis
was also conducted (sleep duration → brain structure → fluid intelli-
gence). The results revealed that the association between sleep
duration and fluid intelligence was also significantly mediated by
brain structure (Extended Data Fig. 6b; β = −0.004, P = 1.4 × 10−5).
The mediation analysis between sleep duration and depressive
symptoms did not yield significant results for participants with
sleep duration >7 h.
Five other cognitive functions were also used to conduct media-
tion analysis (Extended Data Fig. 8). For participants with sleep
duration ≤7 h, brain structure related to sleep significantly medi-
ated the association between sleep duration and numeric memory
(path β = 0.006, P = 1.4 × 10−11), trail making (path β = −0.003,
P = 7.8 × 10−7) and prospective memory (path β = −8.8 × 10−4,
P = 0.021). The association between these cognitive functions and
sleep duration was also significantly mediated by brain structures
related to sleep for participants with sleep duration >7 h, includ-
ing symbol–digit substitution (path β = −0.002, P = 0.019), numeric
memory (path β = −0.003, P = 0.0039), etc.
Structural equation model. Confirmatory factor analysis was used
to examine the latent variables in the structural equation model
including brain structure, mental health and cognitive function. For
participants with sleep duration ≤7 h, the results demonstrated that
depressive symptoms and anxiety symptoms were the main compo-
nents of the mental health latent variable (β = 0.84 and 0.71, respec-
tively, P < 0.001). The volume of the cortex was the most significant
predictor of brain volume (β = 0.98, P < 0.001). The latent variable
cognitive function was represented by fluid intelligence, prospective
memory, the reaction time test and the pair-matching test (β = 1,
0.15, 0.10 and 0.12, respectively; P < 0.001).
Structural equation modeling was used to specify the directional
association between PRS and sleep duration, brain structure, mental
Table 2 | The nonlinear correlation between sleep duration and
key measures
Cognitive function (baseline) F value r value P value
Fluid intelligence 473.5 0.0057 9.12 × 10−206
Matrix pattern completion 30.2 0.0022 8.21 × 10−14
Numeric memory 111.5 0.0043 4.75 × 10−49
Pair matching 79.3 0.0007 3.72 × 10−35
Prospective memory 223.1 0.0026 1.76 × 10−97
Reaction time 130.6 0.0012 2.07 × 10−57
Symbol–digit substitution 28.3 0.0017 5.10 × 10−13
Tower rearranging 17.5 0.0011 2.58 × 10−8
Trail making 26.9 0.0016 2.14 × 10−12
Mental health
Anxiety symptoms 1, 027.4 0.0131 <1 × 10−300
Depressive symptoms 1,013.7 0.0128 <1 × 10−300
Mania symptoms 280.9 0.0037 1.70 × 10−122
Mental distress 244.0 0.0031 1.52 × 10−106
Psychotic experience 328.3 0.0042 4.94 × 10−143
Self-harm 444.3 0.0056 4.02 × 10−193
Trauma 1,106.9 0.0139 <1 × 10−300
Well-being 923.2 0.0116 <1 × 10−300
Brain structure
Area of total surface (lh, mm2) 28.88 0.0015 2.91 × 10−13
Area of total surface (rh, mm2)29.92 0.0015 1.04 × 10−13
Mean thickness (lh, mm) 23.88 0.0012 4.32 × 10−11
Mean thickness (rh, mm) 20.42 0.0010 1.37 × 10−9
Total gray volume (mm3) 39.81 0.0020 5.36 × 10−18
Subcortical gray volume (mm3) 12.64 0.0006 3.24 × 10−6
Cortex volume (lh, mm3) 42.68 0.0021 3.04 × 10−19
Cortex volume (rh, mm3) 35.68 0.0018 3.29 × 10−16
Cortical gray volume (mm3) 41.13 0.0021 1.43 × 10−18
F-tests were used to assess statistical significance and derive F statistics and corresponding
one-sided P values adjusted for multiple comparisons (Bonferroni correction). F statistics were
converted to effect-size r values with the equation
r
=
√F
×
df1
(
F
×
df1
+
df2
)−1, where
df1 is the numerator degrees of freedom and df2 is the denominator degrees of freedom. lh, left
hemisphere; rh, right hemisphere.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
0.75
a
b
c
0.50
Cortical area (lh)
0.25
1.2
0.8
Cortical thickness (lh)
0.4
1.2
0.8
Cortical thickness (rh)
0.4
1.0
0.7
Subcortical gray volume
0.4
1.0
0.7
Cortical gray volume
0.4
20
15
10
0.75
0.50
Cortical area (rh)
0.25
1713 1713 1 7 13
1 7
Sleep duration (h)
13 17
Sleep duration (h)
13 17
Sleep duration (h)
F value
14
12
10
8
F value
13
Fig. 3 | Nonlinear association between sleep duration and brain structure. a, Significant nonlinear association of sleep duration with area of total surface,
global mean thickness and cortical and subcortical gray matter volumes (Bonferroni corrected, P < 0.005). Lines are fitted nonlinear models indicating
fitted mean values, and shaded areas are 95% CIs; gray points are individual data points. b, Cortical regions with their volume significantly and nonlinearly
associated with sleep duration adjusted for intracranial volume, age, sex, body mass index, Townsend deprivation index, educational qualification, smoking
status, drinking status and imaging scanning sites (Bonferroni corrected, P < 0.05). c, Subcortical regions with their volumes significantly nonlinearly
associated with sleep duration adjusted for intracranial volume, age, sex, body mass index, Townsend deprivation index, educational qualification,
smoking status, drinking status and imaging scanning sites (Bonferroni corrected, P < 0.05). F-tests were used to assess statistical significance and
derive F statistics and corresponding one-sided P values adjusted for multiple comparisons.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
health and cognitive function. All associations in the path model
(Fig. 6c) were in the expected direction. All paths were significant
(with P < 0.01). PRS was significantly associated with sleep duration
(β = 0.059, P = 9.2 × 10−10), mental health (β = −0.033, P = 9.0 × 10−8)
and brain gray matter volume (β = 0.034, P < 1.0 × 10−20). Brain
volume was a better predictor of cognitive function (β = −0.14,
P < 1.0 × 10−20) than mental health (β = 0.041, P = 5.0 × 10−11)
or sleep duration (β = −0.049, P = 5.3 × 10−7). Sleep duration
was the most significant predictor of mental health (β = −0.15,
P < 1.0 × 10−20) compared to PRS and brain volume (β = −0.016,
P = 5.8 × 10−11). Sleep duration was a significant predictor of brain
volume (β = 0.048, P < 1.0 × 10−20). The results of the analysis of data
from sleep duration >7 h are provided in Extended Data Fig. 6c.
Discussion
The present study revealed consistent nonlinear associations
between sleep duration and cognitive function, mental health
and brain structure in middle-aged to older adults, with approxi-
mately 7 h as the optimal sleep duration. Our study demonstrated a
coherent mediated pathway involving genetics, sleep duration,
brain structure, cognition and mental health using structural
equation modeling. The brain areas that were significantly quadra-
tically associated with sleep duration included the precentral
cortex, the lateral orbitofrontal cortex, the superior frontal cortex
and the hippocampus. Initially, we took genetics and brain structure
into consideration to explore the nonlinear association between
sleep duration, mental health and cognition. The result of the
longitudinal analysis supported our hypothesis that both insuffi-
cient and excessive sleep duration are associated with follow-up
impaired cognitive performance of a middle-aged to older adult
population. The mediation analyses further suggested that the
genetic constructs of sleep may contribute to behavioral measures,
such as depressive symptoms and cognition, through the mediation
of brain structure.
Nonlinear associations between sleep duration and behavioral
measures. We found a beneficial association with cognitive func-
tion and mental health with a sleep duration of approximately
7 h in a middle-aged to older adult population. A previous study
reported a nonlinear association between sleep duration and
cognitive decline in memory and executive functions14. The current
findings were consistent with this study and extend the findings to
a broader range of cognitive functions including processing speed,
visual attention, memory and problem-solving ability as well as a
much larger cohort of middle-aged to older adults22. This associa-
tion between sleep duration and cognitive performance potentially
suggests that insufficient or excessive sleep duration may be a risk
factor for cognitive decline in aging. This is supported by previous
reports of a nonlinear relationship between nocturnal sleep dura-
tion and the risk of developing Alzheimer’s disease and dementia,
in which cognitive decline is a hallmark symptom23,24. A possible
reason for the association between insufficient sleep duration and
cognitive decline may be due to the disruption of slow-wave sleep,
which has been identified as having a close association with mem-
ory consolidation25 as well as amyloid deposition26–28. A reduction
in sleep time may have detrimental consequences to the clearance
of toxins29. It is possible that prolonged sleep duration results from
poor-quality and fragmented sleep30.
Our results also indicate the close association between optimal
sleep duration and mental health including symptoms of anxiety,
symptoms of depression, mania and well-being. Inadequate or
excessive sleep duration are both considered criteria for severity
of depressive symptoms in the PHQ-9 (ref. 31). Insomnia and
depression have been shown to share overlapping genetic and envi-
ronmental causal influences in twin studies32,33. Our longitudinal
association between baseline sleep duration and follow-up depres-
sive symptoms supported the finding that non-optimal sleep dura-
tion may contribute to psychiatric disorders in middle-aged to older
adults. Our findings further support the idea that interventions that
4710
47
Sleep duration (h)
10 47
Sleep duration (h)
10 47
Sleep duration (h)
10 47
Sleep duration (h)
10
4710 4710 4710
0.26
0.28
0.30
0.18
0.24
0.30
0.1
0.2
0.3
62–70 years old
53–61 years old
39–52 years old
68–82 years ol
d
60–67 years ol
d
44–59 years ol
d
0.20
0.35
0.50
Reaction time
0.32
0.36
0.40
Cortical volume
0.36
0.39
0.42
Subcortical volume
0.45
0.48
0.51
Lateral OFC
0.35
0.38
0.41
Hippocampus
Prospective memory
Mania
Depression
Fig. 4 | The interaction between age and sleep duration. Participants were divided into three age groups: 39–52, 53–61 and 62–70 years of age for
behavioral measures; and 44–59, 60–67 and 68–82 years of age for imaging data collected at follow up. Top, each age group showed nonlinear
associations between sleep duration and cognitive function and mental health. With increasing age, the nonlinear curve gradually tended to flatten, and
the F value decreased simultaneously, especially for reaction time, depressive symptoms and mania symptoms. Bottom, each age group showed nonlinear
associations between sleep duration and brain structures. The nonlinear association between sleep duration and brain structures gradually flattened with
aging. Lines are fitted nonlinear models. OFC, orbitofrontal cortex.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
are able to optimize a 7-h sleep duration on a regular basis may be
beneficial for mental health and reducing psychiatric symptoms34.
Nonlinear associations between sleep duration and brain struc-
ture. Nonlinear associations between sleep duration and brain
structure were found in the current study and suggested that insuffi-
cient or excessive sleep duration was associated with smaller brain
volume, area and thickness. The brain regions involved were the
precentral gyrus, the lateral orbitofrontal cortex, the left insula and
the hippocampus. Previous studies have shown that reduced gray
matter in the lateral orbitofrontal cortex and hippocampal damage
were associated with disrupted sleep patterns in older adults35,36.
Furthermore, poorer sleep quality and efficiency have been associ-
ated with a greater rate of decline in hippocampal volume37. Our
results are consistent with these findings. Moreover, the involvement
of the lateral orbitofrontal cortex and the hippocampus corresponds
1.0
a
b
Cognitive function
Mental health
Fluid intelligence
Trail making
Anxiety symptoms
Mania symptoms
Self-harm Prospective memory
Reaction time
0.5
0.4
1.0
0.5
1.2
0.6
0
1.2
0.6
0
Well-being
1.2
0.6
0
0
Depressive symptoms
Mental distress
1.2 1.4
0.8
0.2
0.6
0
0.5 0.4
0.2
0
–0.5
0
0.2
0
0
1.0
Numeric memory
0.5
0
0.6
Pair matching
0.3
0
04–4
04–4
04–4
04–4
04–4
0
Sleep variability (h)
4–4
04–4
0
Sleep variability (h)
4
–4
04–4
04–4
04–4
04
–4
Fig. 5 | Nonlinear association between sleep variability, mental health, cognitive function and brain structures. a, Sleep variability between baseline
and imaging follow up showed nonlinear significant associations with cognitive function including fluid intelligence, numeric memory, pair matching,
trail making, prospective memory and reaction time (Bonferroni corrected, P < 0.05), with almost 0 h as the inflection point. b, Sleep variability between
baseline and imaging follow up showed nonlinear significant associations with mental health including anxiety symptoms, depressive symptoms, mental
distress, mania symptoms, self-harm and well-being (Bonferroni corrected, P < 0.05), with almost 0 h as the inflection point. F-tests were used to assess
statistical significance and derive F statistics and corresponding one-sided P values adjusted for multiple comparisons. Lines are fitted nonlinear models
indicating fitted mean values, and shaded areas are 95% CIs; gray points are individual data points.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
to our finding of a nonlinear association between sleep duration and
cognitive function and mental health. A previous study also showed
that the lateral orbitofrontal cortex, the insula and the hippocampus
mediate the relationship between sleep quality and depressive
symptoms38, further emphasizing the importance of these brain
regions. Our study also demonstrated that brain structure may play
Cognitive function
–0.049***
–0.14***
0.041***
–0.024**
0.048***
–0.15***
0.059***
–0.016***
–0.033***
0.034***
Sleep
≤7 h
Polygenetic
risk score
Mental health Brain regions
Well-being
Anxiety symptoms
Depressive symptoms
Mania symptoms
Mental distress
Psychotic experience
Self-harm
Trauma
0.23
0.54
0.53
0.86
0.47
0.56
0.50
0.58
Mental health
ab
C
Pair matching 0.12
Cognitive function
Prospective memory
Fluid intelligence
Reaction time 0.10
1.00
0.15
Brain structure
Hippocampus
Hippocampus
Volume of cortex
Volume of subcortex
Frontal polelh
rh
lh
0.67
Lateral OFC rh
Superiorfrontal lh
Pars orbitalis lh
0.61
0.66
0.53
0.56
0.37
0.56
0.74
0.53
Precentral lh
Precentral rh
0.74
0.70
0.77
Middle temporal lh
Lateral OFC lh
rh
Middle temporal rh
Sleep duration
β = 0.45***
β
= 0.012
β
= –0.04***
β = 0.50***
–0.13***
Depressive
symptoms
Depressive
symptoms
Sleep duration
≤7 h
Baseline Follow up
Fluid
intelligence
Fluid
intelligence
–0.03**
β
= –0.004
β
= 0.019
β = –0.018*
β
= 0.021*
–0.08***
Brain
regions
Depressive
symptoms
β = 0.046***
PRS of sleep
Sleep
≤7 h
β = 0.058***
β = –0.033***
β
= 0.059***
β = –0.12***
β = 0.028**
β
= –0.027**
β
= –0.027**
β
1
= –7.3 × 10
–5
*
Fluid
intelligence
Brain
regions
β = 0.044***
β = 0.045***
β = 0.063***
β
= 0.031***
β = 0.061***
β = 0.061***
β
= 0.14***
β
= 0.14***
Sleep
≤7 h
PRS of sleep
Pars orbitalis
0.68
0.98
β
2
= –
0
.
00
7**
*
β
3
= –0.001*
β
1
= 3.8 × 10
–4
*** β
2
= 0.004
***
β
3
= 0.004***
–0.02
Fig. 6 | Structural equation model, longitudinal analysis and mediation analysis. a, The longitudinal association between sleep duration, depressive
symptoms and fluid intelligence revealed by a cross-lagged panel model. The baseline sleep duration was significantly associated with severe depressive
symptoms and fluid intelligence at follow up (β = −0.038, P = 1.3 × 10−5 and β = 0.021, P = 0.01, respectively); baseline depressive symptoms were significantly
associated with follow-up fluid intelligence (β = −0.018, P = 0.029). The reverse (dashed line) was not significant. The associations between baseline
and follow-up sleep duration (β = 0.45, P < 1.0 × 10−20), depressive symptoms (β = 0.50, P < 1.0 × 10−20) and fluid intelligence (β = 0.59, P < 1.0 × 10−20)
were significant. b, Mediation analysis. Three mediation models were conducted to analyze the direct relationship between PRSs of sleep and depressive
symptoms simultaneously, with sleep duration, brain structure and both of them as mediators, respectively. The indirect pathway of the effect of PRS on
depressive symptoms via sleep duration and brain structure was significant (path β1 = −7.3 × 10−5, P = 0.025). Meanwhile, sleep duration and brain structure
significantly mediated the association between PRS and depressive symptoms, respectively (path β2 = −0.0071, P = 9.3 × 10−9; path β3 = −0.001, P = 0.035).
Similarly, three mediation models were conducted to analyze the direct relationship between sleep and fluid intelligence simultaneously, with brain
structure and depressive symptoms as mediators, respectively. The serial mediation pathway via sleep and brain structure was significant (β1 = 3.8 × 10−4,
P = 1.3 × 10−5). Meanwhile, sleep duration and brain structure were also separately significant mediators for this association (β2 = 0.004, P = 2.8 × 10−8;
β3 = 0.004, P = 7.9 × 10−5). These three models are presented using orange, green and blue lines. Data from participants with sleep duration less than 8 h
were used. c, Full frame model. Standardized coefficients are shown. PRS was significantly associated with sleep duration (β = 0.059, P = 9.2 × 10−10), mental
health (β = −0.033, P = 9.0 × 10−8) and brain regions (β = 0.034, P < 1.0 × 10−20). Brain volumes were a better predictor of cognitive function (β = −0.14,
P < 1.0 × 10−20) than mental health (β = 0.041, P = 5.0 × 10−11) or sleep duration (β = −0.049, P = 7.2 × 10−7). Sleep duration was the most significant predictor
of mental health (β = −0.15, P < 1.0 × 10−20) compared to PRS and brain volume (β = −0.016, P = 5.8 × 10−11). Sleep duration was a significant predictor of
brain volume (β = 0.048, P < 1.0 × 10−20). All paths represent significant associations except for the one between PRS and cognitive function. Latent variables
including brain structure, mental health and cognitive function were estimated in the model, which are shown with orange, green and blue boxes, respectively.
*P < 0.05, **P < 0.01 and ***P < 0.001. Wald tests were used to derive the two-sided P values adjusted for multiple comparisons (FDR correction).
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
a critical role in the mediation of the association between genetics,
cognitive function and mental health.
Additionally, the above-mentioned nonlinear association was
also demonstrated between sleep variability and cognition and
mental health. A stable sleep pattern, indicated by a 0-h differ-
ence between baseline and neuroimaging visit sleep durations, was
closely associated with cognition and mental health. Consistent
with our findings, a previous study showed that increased intra-
individual variability in sleep duration is related to psychosocial and
physiological stress39. Moreover, high variability in sleep behaviors
has been associated with increased inflammation, indicating sus-
ceptibility to age-related diseases in older people40. Our findings
further indicate that a consistent sleep duration of approximately
7 h should be maintained long term. For those who have occupa-
tions that require shift work or traveling, this recommendation may
be particularly important to preserve mental health and well-being
as well as cognition41.
Age-varying nonlinear association between sleep duration and
other measures. Our study identifies the distinct nonlinear associa-
tion between sleep duration and cognition and mental health across
different age groups. The gradually flattening nonlinear curve with
increasing age illustrates that the association between sleep dura-
tion and cognitive function and mental health gradually diminishes
in a population aged above 65 years compared with a middle-aged
population of 40 years in age. This distinction might be explained
by the gradual increase in sleep disturbances with age and the fre-
quent occurrence of fragmented sleep42, which may contribute to
the attenuated nonlinear association between optimal sleep dura-
tion and mental and cognitive health in a population aged over 66
years. Furthermore, age-related atrophy of brain regions involved
in the regulation of sleep and wakefulness may contribute to circa-
dian dysfunction and decreased production and secretion of mela-
tonin in older adults43,44. Our results demonstrate that optimal sleep
duration may be more beneficial to the middle-aged population,
which is likely related to their engagement in occupational activities
and skills.
Strengths and limitations. One of the strengths of the current
study is the large sample size from the UK Biobank. In addition, we
comprehensively describe the nonlinear association between sleep
duration and brain structure, mental health and cognitive function.
Based on the large cohort from the UK Biobank, we demonstrate
that these nonlinear associations vary from middle-aged to older
adults. Our study also suggested a longitudinal association between
baseline sleep duration and depressive symptoms and cognition
8 years later. Finally, the possible mechanistic paths underlying
this process ranging from genetics, brain structure and eventual
behavior were specified in current study.
Our study also has some limitations. One limitation of the
current study is that we only used total sleep duration and did not
have access to other measures of sleep hygiene. Future investiga-
tions could focus on enriching sleep measures. In addition, sleep
duration was assessed via a self-reported questionnaire, which may
introduce some bias. Nevertheless, given the large sample size of
the UK Biobank, the measures of sleep duration used in the current
study should be robust. Supporting the robustness of this measure, a
previous study found that self-reported sleep duration showed con-
sistent direction with accelerometer-based sleep duration in asso-
ciation with genetic variants in the UK Biobank sample45. Second,
compared with the general UK population, the UK Biobank has
a ‘healthy volunteer’ selection bias46. Sleep durations used in our
study were mainly reported by healthier people, and future inves-
tigations could further focus on sleep pattern in patients with brain
disorders. Third, online MHQs were used in our study and these
provided quantitative measures of mental health symptoms but not a
Diagnostic and Statistical Manual (DSM)-5 diagnosis. Additionally,
the MHQ scores were obtained several years after the baseline
assessment; nonetheless, the nonlinear associations between sleep
duration and mental health were consistent after adjusting for the
interval years between these two assessment time points. MHQ
participants were better educated, of higher socioeconomic status
and healthier than the entire UK Biobank cohort47. To minimize
bias from this demographic difference, we have adjusted for related
confounders such as educational qualifications, Townsend depriva-
tion index, etc. in all statistical analyses. Finally, the results of our
study reflect the demographic makeup of the UK Biobank and may
not fully extrapolate to other populations.
Future studies could investigate the potential different mecha-
nisms for the association between excessive or insufficient sleep
with mental health and cognitive function. Previous investigations
mainly focused on sleep deprivation; therefore, the mechanism for
the association between excessive sleep and well-being should be
further investigated. In addition, in view of brain atrophy as a critical
characteristic of the aging process, longitudinal neuroimaging
data and sleep measures are needed for further investigation of the
possible contribution of non-optimal sleep duration to brain atrophy
in older adults. As mentioned above, more detailed sleep hygiene
measures, including sleep timing, sleep efficiency and circadian
rhythm, as well as objective sleep measures could be combined in
future studies to provide more detailed sleep recommendations for
the general population. Interestingly, in contrast to the non-mono-
tonic relationship identified in adults, previous studies have reported
a monotonic relationship between sleep duration and behavioral
and neuroimaging measures in adolescents47. Therefore, we would
dedicate future studies to exploring the lifespan associations of
sleep duration with physical and mental health, particularly inves-
tigating the critical transition period when the relationship shifts
from monotonic to non-monotonic.
Conclusion. In conclusion, nonlinear associations between sleep
duration and mental health, cognitive function and brain structure
were found in a large cohort of middle-aged to older participants
from the UK Biobank. The most significant brain structures were
found to include the precentral cortex, the lateral orbitofrontal cor-
tex and the hippocampus. Given the role of the hippocampus in
memory processes and in Alzheimer’s disease, the nonlinear asso-
ciation between sleep duration and this brain region is of particu-
lar importance. Furthermore, baseline non-optimal sleep duration
was significantly associated with decreased cognitive function and
increased psychiatric symptoms on follow up. Our findings have
emphasized the importance of sleep regulation for cognition, men-
tal health and well-being of adults. In addition, we identified a pos-
sible unified pathway that includes genetics and brain mechanisms.
Methods
Participants. We used data from the UK Biobank with application ID 19542,
which included 498,277 participants primarily of European ancestry aged between
38 and 73 years. e participants include 94.3% white people, 0.6% mixed people,
2.0% Asian people, 1.6% Black people, 0.3% Chinese people, 0.9% other ethnic
groups and 0.3% with missing data. e UK Biobank has research tissue bank
approval from the North West Multi-centre Research Ethics Committee
(https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/about-us/ethics)
and provided oversight for this study. Written informed consent was obtained from
all participants. Participation is voluntary, and participants are free to withdraw at
any time without giving any reason. e data consisted of detailed demographic,
health, behavioral and cognitive assessments at baseline and ongoing longitudinal
follow up. Neuroimaging data were collected from 48,511 participants, and 156,884
participants completed online follow-up MHQs 6–8 years aer the baseline
assessment. Neuroimaging data of 39,692 participants were available and used in
the current analyses under the application number 19542. All participants provided
written informed consent.
Sleep measures. Sleep duration was recorded through touchscreen questionnaires
including questions such as ‘About how many hours sleep do you get in every
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
24 hours? (please include naps)’. Answers <1 h or >23 h were rejected, and
answers <3 h or >12 h required confirmation by the participants. If the participant
activated the help button, they were shown the message: ‘If the time you spend
sleeping varies a lot, give the average time for a 24 hour day in the last 4 weeks’.
The sleep-duration data from the baseline assessment (2006–2010, n = 498,277)
and neuroimaging visit (2014+, n = 48,511) were used in the analyses. Sleep
duration assessed at baseline was used to determine the association between
cognitive function and online follow-up mental health assessments. Sleep duration
assessed at the neuroimaging visit was used to determine the association with brain
structure. Histograms of sleep duration are shown in Supplementary Fig. 1; only
integer values of sleep duration were available in the questionnaire.
Mental health. Measurement of depressive symptoms via the four-item PHQ-4 was
first assessed in the UK Biobank Assessment Centre (2006–2010, n = 499,585) and
then repeated at the neuroimaging visit (2014–2017, n = 48,571). Furthermore, a
detailed and comprehensive mental health questionnaire (MHQ) was administered
online (2016–2017, n = 157,366), which assessed self-reported symptoms of mental
disorders and major environmental exposures for mental disorders including
mental distress, depressive symptoms, mania symptoms, anxiety symptoms, alcohol
use, cannabis use, psychotic experiences, traumatic events, self-harm, addiction
and well-being (https://biobank.ndph.ox.ac.uk/showcase/label.cgi?id=136). Based
on these questionnaires, we obtained quantitative measures of various mental
health symptoms. The items used to assess each mental health symptom are
provided in Supplementary Table 2. Briefly, the scores of items in one subcategory
of the MHQ were adjusted to same direction such that higher values indicated
more symptoms of mental disorder (except for well-being, with a higher value
indicating better well-being). Next, each item was normalized by mapping the
minimum and maximum to (0, 1) using the MATLAB function ‘mapminmax’ to
maintain the same scale between items. Finally, all items in one subcategory were
averaged to measure overall symptoms of mental health. Summary information for
each mental health symptom is provided in Supplementary Table 3.
Cognitive testing. Cognitive tests were first administered via a touchscreen
interface in the UK Biobank Assessment Centre at the baseline visit and repeated
at the neuroimaging visit. Six cognitive tests including reaction time, numeric
memory, fluid intelligence, trail-making, prospective memory and pair-matching
tests were used in the current study. Supplementary Table 1 illustrates the sample
sizes for each cognitive task used in the current study. The scores were normalized
by mapping the minimum and maximum to (0, 1) using the MATLAB function
‘mapminmax’ before analysis.
Structural magnetic resonance imaging data. Quality-controlled T1-weighted
neuroimaging data, processed with FreeSurfer, were used in current study. Details
of the imaging protocol can be found in an open-source document (https://
biobank.ndph.ox.ac.uk/showcase/showcase/docs/brain_mri.pdf). Neuroimaging
data were collected with a standard Siemens Skyra 3T scanner with a 32-channel
head coil. T1 images were processed with FreeSurfer; surface templates were
used to extract imaging-derived phenotypes referred to as atlas regions’ surface
area, volume and mean cortical thickness48. Subcortical regions were extracted
via FreeSurfer’s aseg tool49. FreeSurfer aparc (ID 192) and aseg (ID 190) atlases
corresponding to 68 cortical regions and 40 subcortical regions were used in
this study. The intracranial volume (field ID 26521) generated by aseg was used
as a covariate in the neuroimaging analyses. The Qoala-T approach was used to
check FreeSurfer outputs, supplemented by manual checking of outputs close to
the threshold. Any FreeSurfer outputs that failed to pass quality control were not
included in the FreeSurfer imaging-derived phenotypes.
Polygenic risk score for sleep duration. Genotype data were available for all
500,000 participants in the UK Biobank cohort. Detailed genotyping and quality-
control procedures for the UK Biobank are available in a previous publication50.
We excluded single-nucleotide polymorphisms (SNPs) with call rates <95%, minor
allele frequency <0.1% or deviation from the Hardy–Weinberg equilibrium with
P < 1 × 10−10 and selected individuals that were estimated to have recent British
ancestry and have no more than ten putative third-degree relatives in the kinship
table, consistent with the previous study51. After the quality-control procedures, we
obtained a total of 616,339 SNPs and 337,199 participants. To avoided the issue of
circular analysis, participants with neuroimaging data (n = 39,692) were used for
subsequent estimation of PRS and therefore were removed from the genome-wide
association study (GWAS) sample.
We performed genome-wide association analysis, adjusting for age, sex and
the top 20 ancestry principal components, using PLINK 1.90 (ref. 52) to assess the
association between genotype and sleep duration. To determine the nonlinear
effects of sleep duration, we performed genome-wide association analyses for sleep
duration in 114,419 individuals who sleep >7 h and 193,056 individuals who sleep
≤7 h separately. To avoid data-overlap bias, individuals for whom brain MRI was
collected were removed from the above genome-wide association analysis.
LD-score regression (GenomicSEM version 0.0.3 in R) was used to assess
the SNP-based heritability of sleep duration ≤7 h and >7 h, respectively, and its
genetic correlation with a previous published GWAS of sleep duration (http://www.
t2diabetesgenes.org/data/). HapMap3 SNPs were used as the reference SNP list.
European ancestral background LD scores from the 1000 Genomes Project were
used as the reference panel. The heritability for sleep duration ≤7 h and >7 h was
0.0636 (s.e.m. = 0.0051) and 0.0194 (s.e.m. = 0.0068), respectively. The heritability
for sleep duration >7 h was lower than the suggested threshold of z (h2z = 2.83),
which implied potential bias of genetic correlation. The genetic correlation
between sleep duration ≤7 h and the previous GWAS of sleep duration was 0.687
(s.e.m. = 0.074, P = 1.67 × 10−20). A positive genetic correlation between sleep
duration >7 h and previous GWAS results was also found (0.3382; s.e.m. = 0.1416,
P = 0.0168).
PRSs for sleep duration in individuals with brain MRI measures were
calculated using PRSice software (http://www.prsice.info). P-value-informed
clumping with a cutoff of r2 = 0.1 in a 250-kb window was used in the analysis.
PRSs were calculated using the mean of P values at the threshold ranging from
0.005 to 0.5 with 0.005 as the step size.
Statistical analysis. Nonlinear association analysis. A nonlinear regression
model (y = bx2 + ax + c) was used to investigate the association of sleep duration
(x) with the measures of interest (y), including mental health variables (online
follow up), cognitive tests scores and brain morphometric measures. e
following variables were used as covariates of no interest in the model: age, sex,
body mass index, the scanning site of imaging, Townsend deprivation index
measuring socioeconomic status, educational qualications, smoking status
and drinking status (Supplementary Fig. 2). Furthermore, we adjusted for
intracranial volumes, derived using the FreeSurfer aseg tool in the regression
model examining sleep and brain structures. An F statistic was obtained for each
quadratic model to reect the association of sleep duration and the measures
of interest. F statistics were transferred to eect-size r values using the equation
r
=
√F
×
df1
(
F
×
df1
+
df2
)−
1
, where df1 is the numerator degrees of freedom
and df2 is the denominator degrees of freedom53. Bonferroni corrections were
conducted for multiple comparisons. Restricted cubic splines (package rms 6.2-0
in R) with three knots at the tenth, 50th and 90th percentiles were also used to
model the association between sleep duration and brain structures and validate
the nonlinearity.
Two-line tests were conducted to estimate an interrupted regression and to
identify the breakpoint between lines with opposite sign of slope54. The breakpoint
was set to maximize the power of detecting nonlinear relationships (two-lines test
version 0.52 implemented in R).
Longitudinal analysis. The longitudinal association of sleep duration with
depressive scores (PHQ-4) and with cognitive function was explored using a classic
two-wave cross-lagged panel model (implemented with the lavaan 0.8 package
in R). The analysis was conducted separately for participants with sleep duration
≤7 h and >7 h at baseline assessment. PHQ-4 scores and fluid intelligence scores
at the baseline assessment and at the neuroimaging visit assessment were used.
Covariates including age, sex, body mass index, Townsend deprivation index,
educational qualification, smoking status and drinking status were regressed out
before the analysis. Model parameters were estimated by maximum likelihood
estimation. Standardized regression coefficients and their standard errors were
reported throughout.
Structural equation model. A structural equation model was estimated separately
for participants with sleep duration ≤7 h and >7 h (implemented in R (lavaan
0.8)). Three latent variables were estimated in the model using confirmatory
factor analysis. A latent variable representing cognitive function was estimated
via reaction time, fluid intelligence, prospective memory and pair-matching
performance, which were all significantly quadratically associated with sleep
duration. The latent variable of mental health was also measured in the model
using anxiety symptom, cannabis, depressive symptom, mania symptom, mental
distress, psychotic experience, self-harm, trauma and well-being scores in the
MHQ. Finally, the latent variable for brain structures was derived from the first ten
cortical and five subcortical brain gray matter volumes significantly correlated with
sleep duration, adjusted for intracranial volume and the other specified covariates.
These three latent variables were investigated to determine the directional
dependencies with PRS and sleep duration via path modeling.
Mediation analysis. Three mediation models were used in the current study. First,
the serial mediation model was used to investigate whether the association of PRS
with depressive symptoms was mediated by brain structures and sleep duration,
adjusting for age, sex, educational qualification, body mass index, scanning sites,
PRS components, smoking status and drinking status. The mean values of brain
gray matter volumes significantly associated with sleep duration that survived
Bonferroni correction (P < 0.05 adjusted for intracranial volume) were used in
the model. PRS was calculated separately based on sleep duration >7 h or ≤7 h.
Depressive symptoms used in the model were measured (category ID 138) in an
online follow-up questionnaire (Supplementary Table 2). Two other mediation
analyses were conducted, specifically to determine whether the association
between PRS and depressive symptoms could be mediated only by brain regions
or sleep duration; the same covariates as in the first model were used in the second
model and the third model excluded scanning sites.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Similarly, three mediation models were also conducted to investigate the
association between PRS for sleep and fluid intelligence. The first model was
used to investigate whether the association between PRS and fluid intelligence
could be serially mediated by sleep and brain structures. The second and third
models were used to investigate whether the association between PRS and fluid
intelligence could be mediated by sleep or brain structures, respectively. The
mediation model used the same covariates as mentioned above. Analyses were
conducted separately for sleep duration less than 8 h and greater than 7 h. Finally,
serial mediation analysis of the path sleep duration → brain structure → depressive
symptoms → fluid intelligence was also conducted. Additionally, five other
cognitive functions were also used to conduct mediation analysis via the path
sleep duration → brain structure → cognitive function including symbol–digit
substitution, numeric memory, trail making, prospective memory and reaction
time. Total, direct and indirect associations were estimated by the 10,000-iteration
nonparametric bootstrap approach. Analysis was performed in R with lavaan 0.8.
Interaction between age and sleep duration. To explore whether sleep duration
was associated with mental health, cognitive function and brain structure across
various ages, participants were first divided into three age groups, ensuring similar
numbers of participants in each group; these were 39–52 years old, 53–61 years
old and 62–70 years old for behavioral data. Binomial fitting was conducted
for each age group to observe the interaction between age and sleep duration.
Neuroimaging data were collected at a follow-up visit around 4 years later, and
participants were also divided into three age groups to determine the association
between sleep duration and brain structures; these were 44–59, 60–67 and
68–82 years old.
Furthermore, to test the significance of the interaction between sleep duration
and age, the linear interaction term ‘
sleep
×
age
’ and the nonlinear interaction
term ‘
sleep2
×
age
’ were added to the original nonlinear regression model55
(implemented with the AER 1.2-9 package in R). t-tests were conducted to identify
the significance of the coefficient of each interaction term. Additionally, F-tests
were conducted to test the joint hypotheses that both coefficients of interaction
terms were zero. A similar method was also used to explore the interaction
between sleep duration and sex, with ‘
sleep
×
sex
’ and ‘
sleep2
×
sex
’ added to the
nonlinear regression model.
Reporting Summary. Further information on research design is available in the
Nature Research Reporting Summary linked to this article.
Data availability
This project corresponds to UK Biobank application ID 19542. Neuroimaging,
genotype and behavioral data from the UK Biobank dataset are available at
https://biobank.ndph.ox.ac.uk/ by application. The variables used here are detailed
in Supplementary Table 1. The previously published GWAS of sleep duration
was downloaded from http://www.t2diabetesgenes.org/data/. European ancestral
background LD scores from the 1000 Genomes Project were downloaded from
https://alkesgroup.broadinstitute.org/LDSCORE/.
Code availability
MATLAB 2018b was used to perform nonlinear association analysis. FreeSurfer
version 6.0 was used to process imaging data. PLINK 1.90 and PRSice (http://www.
prsice.info) were used to perform genome-wide association analysis and calculate
the PRS, respectively. lavaan 0.8 in R version 3.6.0 was used to perform longitudinal
and mediation analyses and make the structural equation model. AER 1.2-9 in
R version 3.6.0 was used to perform the interaction test; rms 6.2-0 was used to
conduct restricted cubic spine analysis; GenomicSEM version 0.0.3 was used to
calculate heritability and genetic correlation; two-lines test version 0.52 was used
to identify the breakpoints of the nonlinear model. Scripts used to perform the
analyses are available at https://github.com/yuzhulineu/UKB_sleep.
Received: 6 July 2021; Accepted: 17 March 2022;
Published: xx xx xxxx
References
1. Stickgold, R. Sleep-dependent memory consolidation. Nature 437, 1272–1278
(2005).
2. Walker, M. P. & van der Helm, E. Overnight therapy? e role of sleep in
emotional brain processing. Psychol. Bull. 135, 731–748 (2009).
3. Xie, L. L. et al. Sleep drives metabolite clearance from the adult brain. Science
342, 373–377 (2013).
4. He, Q., Zhang, P., Li, G. X., Dai, H. X. & Shi, J. P. e association between
insomnia symptoms and risk of cardio-cerebral vascular events: a meta-
analysis of prospective cohort studies. Eur. J. Prev. Cardiol. 24, 1071–1082
(2017).
5. Spira, A. P., Chen-Edinboro, L. P., Wu, M. N. & Yae, K. Impact of sleep
on the risk of cognitive decline and dementia. Curr. Opin. Psychiatry 27,
478–483 (2014).
6. Sabia, S. et al. Association of sleep duration in middle and old age with
incidence of dementia. Nat. Commun. 12, 2289 (2021).
7. Ikehara, S. et al. Association of sleep duration with mortality from
cardiovascular disease and other causes for Japanese men and women: the
JACC study. Sleep 32, 295–301 (2009).
8. Westwood, A. J. et al. Prolonged sleep duration as a marker of early
neurodegeneration predicting incident dementia. Neurology 88, 1172–1179
(2017).
9. Crowley, K. Sleep and sleep disorders in older adults. Neuropsychol. Rev. 21,
41–53 (2011).
10. Ohayon, M. M., Carskadon, M. A., Guilleminault, C. & Vitiello, M. V.
Meta-analysis of quantitative sleep parameters from childhood to old age in
healthy individuals: developing normative sleep values across the human
lifespan. Sleep 27, 1255–1273 (2004).
11. Leng, Y. et al. Self-reported sleep patterns in a British population cohort.
Sleep Med. 15, 295–302 (2014).
12. Gulia, K. K. & Kumar, V. M. Sleep disorders in the elderly: a growing
challenge. Psychogeriatrics 18, 155–165 (2018).
13. Irwin, M. R. & Vitiello, M. V. Implications of sleep disturbance and
inammation for Alzheimer’s disease dementia. Lancet Neurol. 18, 296–306
(2019).
14. Ma, Y. J. et al. Association between sleep duration and cognitive decline.
JAMA Netw. Open 3, e2013573 (2020).
15. Xu, W. et al. Sleep characteristics and cerebrospinal uid biomarkers of
Alzheimer’s disease pathology in cognitively intact older adults: the CABLE
study. Alzheimers Dement. 16, 1146–1152 (2020).
16. Lo, J. C., Loh, K. K., Zheng, H., Sim, S. K. Y. & Chee, M. W. L. Sleep duration
and age-related changes in brain structure and cognitive performance. Sleep
37, 1171–1178 (2014).
17. Kocevska, D. et al. e prospective association of objectively measured sleep
and cerebral white matter microstructure in middle-aged and older persons.
Sleep 42, zsz140 (2019).
18. Wu, Y. H. & Swaab, D. F. Disturbance and strategies for reactivation of the
circadian rhythm system in aging and Alzheimer’s disease. Sleep Med. 8,
623–636 (2007).
19. Grumbach, P. et al. Sleep duration is associated with white matter
microstructure and cognitive performance in healthy adults. Hum. Brain
Mapp. 41, 4397–4405 (2020).
20. Shi, H. Y. et al. Sleep duration and snoring at midlife in relation to healthy
aging in women 70 years of age or older. Nat. Sci. Sleep 13, 411–422 (2021).
21. El-Sheikh, M., Philbrook, L. E., Kelly, R. J., Hinnant, J. B. & Buckhalt, J. A.
What does a good night’s sleep mean? Nonlinear relations between sleep and
children’s cognitive functioning and mental health. Sleep 42, zsz078 (2019).
22. Kendall, K. M. et al. Cognitive performance among carriers of pathogenic
copy number variants: analysis of 152,000 UK Biobank subjects. Biol.
Psychiatry 82, 103–110 (2017).
23. Lutsey, P. L. et al. Sleep characteristics and risk of dementia and Alzheimer’s
disease: the Atherosclerosis Risk in Communities Study. Alzheimers Dement.
14, 157–166 (2018).
24. Liang, Y., Qu, L. B. & Liu, H. Non-linear associations between sleep duration
and the risks of mild cognitive impairment/dementia and cognitive decline: a
dose–response meta-analysis of observational studies. Aging Clin. Exp. Res.
31, 309–320 (2019).
25. Helfrich, R. F., Mander, B. A., Jagust, W. J., Knight, R. T. & Walker, M. P. Old
brains come uncoupled in sleep: slow wave–spindle synchrony, brain atrophy,
and forgetting. Neuron 97, 221–230 (2018).
26. Klinzing, J. G., Niethard, N. & Born, J. Mechanisms of systems memory
consolidation during sleep. Nat. Neurosci. 22, 1598–1610 (2019).
27. Olsson, M., Arlig, J., Hedner, J., Blennow, K. & Zetterberg, H. Sleep
deprivation and cerebrospinal uid biomarkers for Alzheimer’s disease. Sleep
41, zsy025 (2018).
28. Winer, J. R. et al. Sleep disturbance forecasts β-amyloid accumulation across
subsequent years. Curr. Biol. 30, 4291–4298 (2020).
29. Holth, J. K. et al. e sleep–wake cycle regulates brain interstitial uid tau in
mice and CSF tau in humans. Science 363, 880–884 (2019).
30. Grandner, M. A. & Kripke, D. F. Self-reported sleep complaints with long
and short sleep: a nationally representative sample. Psychosom. Med. 66,
239–241 (2004).
31. Kroenke, K., Spitzer, R. L. & Williams, J. B. e PHQ-9: validity of a brief
depression severity measure. J. Gen. Intern. Med. 16, 606–613 (2001).
32. Gehrman, P. R. et al. Heritability of insomnia symptoms in youth and their
relationship to depression and anxiety. Sleep 34, 1641–1646 (2011).
33. Gregory, A. M. et al. A longitudinal twin and sibling study of associations
between insomnia and depression symptoms in young adults. Sleep 39,
1985–1992 (2016).
34. Freeman, D., Sheaves, B., Waite, F., Harvey, A. G. & Harrison, P. J. Sleep
disturbance and psychiatric disorders. Lancet Psychiatry 7, 628–637 (2020).
35. Lim, A. S. P. et al. Regional neocortical gray matter structure and sleep
fragmentation in older adults. Sleep 39, 227–235 (2016).
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg
36. Spano, G. et al. Sleeping with hippocampal damage. Curr. Biol. 30,
523–529 (2020).
37. Fjell, A. M. et al. Self-reported sleep relates to hippocampal atrophy across
the adult lifespan: results from the Lifebrain consortium. Sleep 43, zsz280
(2020).
38. Cheng, W., Rolls, E. T., Ruan, H. T. & Feng, J. F. Functional connectivities in
the brain that mediate the association between depressive problems and sleep
quality. JAMA Psychiatry 75, 1052–1061 (2018).
39. Mezick, E. J. et al. Intra-individual variability in sleep duration and fragmen-
tation: associations with stress. Psychoneuroendocrinology 34, 1346–1354
(2009).
40. Okun, M. L. et al. Sleep variability, health-related practices, and inammatory
markers in a community dwelling sample of older adults. Psychosom. Med.
73, 142–150 (2011).
41. Vetter, C., Fischer, D., Matera, J. L. & Roenneberg, T. Aligning work and
circadian time in shi workers improves sleep and reduces circadian
disruption. Curr. Biol. 25, 907–911 (2015).
42. Mander, B. A., Winer, J. R. & Walker, M. P. Sleep and human aging. Neuron
94, 19–36 (2017).
43. Musiek, E. S., Xiong, D. D. & Holtzman, D. M. Sleep, circadian rhythms,
and the pathogenesis of Alzheimer disease. Exp. Mol. Med. 47, e148 (2015).
44. Skene, D. J. & Swaab, D. F. Melatonin rhythmicity: eect of age and
Alzheimer’s disease. Exp. Gerontol. 38, 199–206 (2003).
45. Jones, S. E. et al. Genetic studies of accelerometer-based sleep measures yield
new insights into human sleep behaviour. Nat. Commun. 10, 1585 (2019).
46. Fry, A. et al. Comparison of sociodemographic and health-related
characteristics of UK Biobank participants with those of the general
population. Am. J. Epidemiol. 186, 1026–1034 (2017).
47. Davis, K. A. S. et al. Mental health in UK Biobank—development,
implementation and results from an online questionnaire completed by
157 366 participants: a reanalysis. BJPsych Open 6, e18 (2020).
48. Desikan, R. S. et al. An automated labeling system for subdividing the
human cerebral cortex on MRI scans into gyral based regions of interest.
Neuroimage 31, 968–980 (2006).
49. Fischl, B. et al. Whole brain segmentation: automated labeling of
neuroanatomical structures in the human brain. Neuron 33, 341–355 (2002).
50. Bycro, C. et al. e UK Biobank resource with deep phenotyping and
genomic data. Nature 562, 203–209 (2018).
51. Kang, J. et al. Increased brain volume from higher cereal and lower coee
intake: shared genetic determinants and impacts on cognition and
metabolism. Cereb. Cortex https://doi.org/10.1093/cercor/bhac005 (2022).
52. Purcell, S. et al. PLINK: a tool set for whole-genome association and
population-based linkage analyses. Am. J. Hum. Genet. 81, 559–575 (2007).
53. Cooper, H. M., Hedges, L. V. & Valentine, J. C. e Handbook of Research
Synthesis and Meta-Analysis 3rd edn (Russell Sage Foundation, 2019).
54. Simonsohn, U. Two lines: a valid alternative to the invalid testing of
U-shaped relationships with quadratic regressions. Adv. Methods Pract.
Psychol. Sci. 1, 538–555 (2018).
55. Kleiber, C. & Zeileis, A. Applied Econometrics with R (Springer, 2008).
Acknowledgements
This study used the UK Biobank Resource under application number 19542. We thank
all participants and researchers from the UK Biobank. J.F. was supported by the
National Key R&D Program of China (nos. 2018YFC1312900 and 2019YFA0709502),
the Shanghai Municipal Science and Technology Major Project (no. 2018SHZDZX01),
the ZJ Lab, Shanghai Center for Brain Science and Brain-Inspired Technology and
the 111 Project (no. B18015). W.C. was supported by grants from the National Natural
Sciences Foundation of China (no. 82071997) and the Shanghai Rising-Star Program
(no. 21QA1408700).
Author contributions
J.F. and W.C. proposed the study. Y.L., J.K. and W.Z. analyzed data. S.X. preprocessed data.
W.C., J.F. and B.J.S. contributed to interpretation of results. Y.L. drafted the manuscript.
B.J.S., C.L., J.Y. and W.C. edited the manuscript. Y.L., C.X. and W.C. contributed to
visualization. All authors considered how to analyze data and approved the manuscript.
Competing interests
The authors declare no competing interests.
Additional information
Extended data is available for this paper at https://doi.org/10.1038/s43587-022-00210-2.
Supplementary information The online version contains supplementary material
available at https://doi.org/10.1038/s43587-022-00210-2.
Correspondence and requests for materials should be addressed to Wei Cheng
or Jianfeng Feng.
Peer review information Nature Aging thanks Naiara Demnitz, Cathryn Lewis and the
other, anonymous, reviewer(s) for their contribution to the peer review of this work.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional affiliations.
© The Author(s), under exclusive licence to Springer Nature America, Inc. 2022,
corrected publication 2022
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Articles Nature agiNg
Extended Data Fig. 1 | Histograms of sleep duration in baseline and imaging assessment. The sleep duration data from the baseline assessment
(2006–2010, n = 498,277) and neuroimaging visit (2014+, n = 48,511) were used in the analyses. Sleep duration assessed at baseline was utilized to
determine the association between cognitive function and online follow-up mental health assessments. Sleep duration assessed at the neuroimaging visit
was used to determine the association with brain structure.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg Articles
Nature agiNg
Extended Data Fig. 2 | Covariates utilized in the statistical analyses. Age, sex, body mass index, Townsend deprivation index, educational qualification,
smoking status and drinking status were adjusted in all analyses. In addition, for analysis involving neuroimaging data and polygenetic risk score,
intracranial volumes, neuroimaging scanning sites and PRS components were further added as covariates respectively.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Articles Nature agiNg
Extended Data Fig. 3 | Nonlinear association between sleep duration and cortical area and thickness. Cortical regions with their a) area and b) thickness
significantly and nonlinearly associated with sleep duration adjusted for sleep duration with intracranial volume, age, sex, sex, body mass index, Townsend
deprivation index, educational qualification, smoking status and drinking status, imaging scanning sites (Bonferroni corrected, p < 0.05). F-tests were
utilized to access statistical significance and derive F-statistics and corresponding one-sided p values adjusted for multiple comparisons.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg Articles
Nature agiNg
Extended Data Fig. 4 | Sex difference of the association between sleep duration and mental health, cognitive function and brain structure. The nonlinear
association between sleep duration and anxiety symptom was more significant in female participants (F female = 622.6, n = 533, 2878, 15240, 36239,
26712, 4712 and 1000 participants respectively; F male = 417.2, n = 347, 2126, 12587, 29710, 18677, 3229 and 665 participants respectively), whereas
mania symptoms showed more significant association with sleep duration for male participants (F female = 140.3, n = 550, 2928, 15466, 36672, 26988,
4774 and 1025 participants respectively; F male = 145.0 respectively, n = 354, 2148, 12673, 29893, 18780, 3252 and 670 participants respectively).Fluid
intelligence were found to have a greater nonlinear association with sleep duration in females compared with males (F female = 272.7, n = 940, 3981, 16606,
32724, 25625, 5051 and 1502 participants respectively; F male = 205.4, n = 673, 3144, 14934, 29192, 20019, 4018 and 1223 participants respectively) while
pair matching were more associated with sleep duration in males (F female = 85.8, n = 3087, 11892, 48704, 98567, 79070, 15934 and 4922 participants
respectively; F male = 104.1, n = 2367, 9356, 44236, 88501, 61182, 12315 and 3980 participants respectively). For brain structure, female participants
demonstrated a more significant association between sleep duration and cortical volumes (rh, F female = 29.1, n = 192, 991, 4221, 8375, 5746, 1158 and
249 participants respectively; F male = 14.7, n = 118, 631, 3445, 7523, 5592, 1231 and 238 participants respectively) while cortical thickness was more
significantly associated with sleep duration for males (F female = 2.89, n = 192, 991, 4221, 8375, 5746, 1158 and 249 participants respectively; F male = 20.0,
n = 118, 631, 3445, 7523, 5592, 1231 and 238 participants respectively). Lines are fitted nonlinear model indicating fitted mean value and error bar is
standard error of the mean.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Articles Nature agiNg
Extended Data Fig. 5 | Histograms of the change of variables over time in the longitudinal analysis. Baseline sleep duration is 0.031 hours longer than the
follow-up sleep duration (std = 0.94). At baseline, participants were more depressed compared with the measurement at follow-up (difference = 0.0069,
std = 0.11). Fluid intelligence of participants at baseline was also higher than at follow-up (difference = 0.043, std = 1.73).
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg Articles
Nature agiNg
Extended Data Fig. 6 | Structural equation model, longitudinal analysis and mediation analysis for participants with more than 7hours sleep. a. The
longitudinal association between the sleep duration, depression and fluid intelligence revealed by cross-lagged panel model. The baseline sleep duration
(β= 0.025, p = 1.3 × 10−5) and depressive symptom (β= −0.023, p = 1.3 × 10−5) was significantly associated with fluid intelligence in the follow-up.
b. Mediation analysis. The mediation models were conducted to analyze the direct relationship between sleep duration and fluid intelligence, with sleep
duration, brain structure and both of them as mediator respectively. Brain regions significantly mediated the association between sleep duration and fluid
intelligence (β= −0.0046, p = 1.4 × 10−5). These figures utilized participants with sleep duration more than 7 hours. c. Full frame model. Standardized
coefficients were showed in the figure. PRS was significantly associated with mental health (β = −0.034, p = 4.7 × 10−5). Brain volumes were a better
predictor of cognitive function (β= −0.198, p < 1.0 × 10−20) compared to mental health (β = 0.048, p = 3.5 × 10−6). Sleep duration was the most significant
predictor of mental health (β = 0.167, p < 1.0 × 10−20) and brain regions (β= −0.044, p < 1.0 × 10−20). Latent variable including brain structure, mental health
and cognitive function were estimated in the model which showed in the figure with orange, green and blue box respectively. Wald tests were utilized
to derive the two-sided p value adjusted for multiple comparisons (FDR correction). * represented p < 0.05, ** represented p < 0.01 and *** represented
p < 0.001.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Articles Nature agiNg
Extended Data Fig. 7 | Mediation analysis. a. Three mediation analysis were conducted between PRS and sleep duration, 1) PRS →depressive
symptoms→brain structure →sleep, 2) PRS →depressive symptoms→sleep, 3) PRS→brain structure →sleep. Depressive symptoms and brain structure
serially mediated the association between PRS and sleep duration (β=4.14 × 10−5, p = 0.044). Specifically, with depression significantly associated with
PRS (β = −0.033, p = 1.6 × 10−4) and brain volumes positively associated with depression (β = −0.028, p = 1.4 × 10−3), and in addition, brain volumes
significantly associated with sleep duration (β = 0.044, p = 1.4 × 10−6). Meanwhile, depressive symptoms and brain structure also separately significantly
mediated the association between PRS and sleep duration (β2 = 0.004, p = 3.3 × 10−4, β3 = 0.001, p = 0.01). b. Three mediation pathway analyses were
conducted for the cognitive function of fluid intelligence for participants with less than 8 hours sleep duration, 1) sleep duration→brain structure→
depression→ fluid intelligence, 2) sleep duration→ brain structure → fluid intelligence, 3) sleep duration→ depression→ fluid intelligence. Sleep duration
showed a significant positive association with fluid intelligence in the model (β = 0.062, p = 2 × 10−15). The serial mediation pathway via brain structure
and depression was not significant (β1 = 3 × 10−5, p = 0.06), but brain structure and depression were separately significant mediators for this association.
Brain structure accounted for the association between sleep duration and fluid intelligence (β2 = 0.009, p = 2 × 10−7; β3 = 0.004, p = 5.6 × 10−5). Wald tests
were utilized to derive the two-sided p value adjusted for multiple comparisons (FDR correction). * represented p < 0.05, ** represented p < 0.01 and
*** represented p < 0.001.
NATURE AGING | www.nature.com/nataging

Articles
Nature agiNg Articles
Nature agiNg
Extended Data Fig. 8 | See next page for caption.
NATURE AGING | www.nature.com/nataging

Articles Nature agiNg
Articles Nature agiNg
Extended Data Fig. 8 | Mediation analysis between sleep duration and cognitive functions. For participants with sleep duration ≤ 7 hours, brain structure
related to sleep significantly mediated the association between sleep duration and numeric memory (path β= 0.006, p = 1.4 × 10−11), trail making (path
β= −0.003, p = 7.8 × 10−7), prospective memory (path β= −8.8 × 10−4, p = 0.02) and tower rearranging (path β= 0.004, p = 9.5 × 10−9). Meanwhile, sleep
duration and brain regions related to sleep significantly mediated the association between PRS of sleep and symbol digit substitution (path β = 1.5 × 10−4,
p = 0.001). Specifically, with sleep duration significantly associated with PRS (β = 0.058, p = 4.5 × 10−10) and brain volumes positively associated with
sleep duration (β = 0.053, p = 1.1 × 10−8), and in addition, brain volumes significantly associated with symbol digit substitution (β = 0.049, p = 1.1 × 10−7).
Sleep duration (β = 0.003, p = 6.3 × 10−5) and brain volumes (β = 0.002, p = 4.2 × 10−3) also separately mediated the association between PRS and symbol
digit substitution. The association between these cognitive functions and sleep duration were also significantly mediated by brain structure related to
sleep for participants with sleep duration > 7 hours, including symbol digit substitution (path β= −0.002, p = 0.019), numeric memory (path β= −0.003,
p = 0.004) and trail making (path β= 0.002, p = 0.017). Reaction time and sleep duration were also mediated by brain structure for participants with
sleep duration > 7 hours (path β= 0.001, p = 0.031). Wald tests were utilized to derive the two-sided p value adjusted for multiple comparisons
(FDR correction). * represented p < 0.05, ** represented p < 0.01 and *** represented p < 0.001.
NATURE AGING | www.nature.com/nataging