Sleep duration, chronotype, health and lifestyle factors affect cognition: a UK Biobank cross-sectional study

Study design, data source and study population

The study data were derived from the UK Biobank, a population-based prospective study established by the UK Medical Research Council and Wellcome Trust. The study recruited approximately 501 718 men and women aged ≥40 years across the UK registered with the UK National Health Service. Details of the UK Biobank with Ethics Committee approval have been previously described.25 As this study involved secondary analysis of anonymised data, no additional ethical approval was needed. This report follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cross-sectional studies. Our analysis capitalised on the comprehensive data available from the UK Biobank, encompassing participants with complete information on key variables such as sex, birth year, body mass index (BMI), smoking status, frequency of alcohol intake and medical histories including diabetes, vascular/heart problems and cancer diagnosis. To optimise the analysis and enhance the representativeness of our findings, participants were divided into two cohorts based on the different combinations of cognitive assessments completed. Cohort 1, consisting of 10 067 participants (mean age 71 years, 56% female), included those who completed all four cognitive tests (Fluid Intelligence/reasoning, Pairs Matching, Reaction Time and Prospective Memory). In contrast, Cohort 2, comprising a larger group of 16 753 participants (mean age 72 years, 56% female), included individuals who completed only two of these cognitive assessments (Pairs Matching and Reaction Time). This division allowed us to incorporate a larger number of participants, thus making effective use of the extensive UK Biobank dataset. The two distinct cohorts were analysed separately. In selecting participants, we prioritised the inclusion of a comprehensive set of sleep parameters and relevant confounding variables. However, we did not adjust for educational attainment due to incomplete data in this area. Additionally, responses indicating unclear chronotypes such as “I don't know” were excluded from our analysis.

In the selection of our study participants from the UK Biobank, we prioritised the inclusion of a comprehensive set of sleep parameters and relevant confounding variables to rigorously assess their relationship with cognitive performance. As a result, the final sample size was determined based on the availability of complete data across these chosen variables. This approach, while resulting in a relatively smaller sample size compared with some previous UK Biobank sleep studies, ensured that our analysis was focused and specific to our research questions. It represents a trade-off between sample size and the depth and relevance of data for the specific variables under investigation. Furthermore, the cognitive assessments employed in the UK Biobank data are well validated and were administered through a novel brief computerised platform. The cognitive assessments—Fluid intelligence, Pairs Matching, Reaction Time and Prospective Memory—are used to evaluate different parameters of cognition including logic and reasoning, visual memory, processing speed and prospective memory, respectively.26 Therefore, given its large sample size and extensive cognitive assessments, the UK Biobank data has been used across numerous studies.26–28

Assessments were conducted between 2006 and 2010 across 22 centres in England, Scotland and Wales. Health information and sleep-related variables were obtained using self-report questionnaires, and cognitive assessments were conducted digitally.

Cognitive variables

Cognitive performance was assessed through four (Cohort 1) or two (Cohort 2) cognitive tests designed for the UK Biobank. The cognitive tests examined the performance of cognitive function and the stability of such ability over time has been well established previously.27

The assessment procedures of the four cognitive tests were as follows:

Fluid Intelligence/reasoning (Data Field ID: 20016): participants were given 13 verbal and numerical fluid intelligence questions designed by the UK Biobank. Each question was given a 2 min answering time to select five possible response answers. The dependent variable measured is the unweighted sum (0–13) of the number of correct answers from the 13 questions.

Pairs Matching (Data Field ID: 399): participants were presented with 12 cards consisting of six pairs of symbols. The cards then were turned face-down on the computer touchscreen and participants were tasked to identify and match as many pairs of cards as possible. The dependent variable was the number of errors made during the test.

Reaction Time (Data Field ID: 20023): Reaction time was assessed through the card game Snap. Participants were requested to press a snap button in response to the appearance of matched cards displayed on the computer touchscreen. The dependent variable was measured by the mean duration of time taken in milliseconds to react to the 12 matching trials. Values were rounded to the nearest whole number.

Prospective Memory (Data Field ID: 20018): participants were given the following instructions on the computer touchscreen: “At the end of the games, we will show you four coloured symbols and ask you to touch the blue square. However, to test your memory, we want you to actually touch the orange circle instead”. The participants were then given up to two attempts to recall the above instruction correctly after a filled interval. The dependent variable was measured by whether the participants had successfully recalled the instruction.

Sleep variables

The study focused on three sleep-related variables: sleep duration, sleep pattern (ie, chronotype) and sleep quality.

Sleep duration (Data Field ID: 1160) was collected from the touchscreen question “How many hours of sleep do you get in every 24 hours (please include naps)?”. The inputted responses were then systematically filtered by the following criteria: if the answer was <1 hour or >23 hours it was rejected; if the answer was <3 hours or >12 hours then the participants were asked to confirm. For this study, sleep duration was categorised into short (<7 hours), normal (7–9 hours) and long (>9 hours) in accordance with the guidelines of the American Academy of Sleep Medicine and Sleep Research Society.29

Sleep pattern (Data Field ID: 1180) was determined by an individual’s chronotype (ie, a morningness person is active and alert predominantly in the morning while dormant at night while an eveningness person is active and alert predominantly at night while dormant in the morning). This was assessed through the touchscreen question: “Do you consider yourself to be?” Participants were then given six answer options to select: ‘definitely a morning person’, ‘more a morning than an evening person’, ‘more an evening than a morning person’, ‘definitely an evening person’, ‘do not know’ and ‘prefer not to answer’. For this study the data were re-categorised into three groups: ‘Morningness’, which consisted of participants who answered ‘definitely a morning person’; ‘Intermediate’, which consisted of participants who answered ‘more a morning than evening person’ and ‘more an evening than morning person’; and ‘Eveningness’, which consisted of participants who replied ‘definitely an evening person’.

Sleep quality (Data Field ID: 1200) was assessed through the degree of sleeplessness/insomnia a participant experienced. Participants were asked the question “Do you have trouble falling asleep at night or do you wake up in the middle of the night?” and given the four answer choices of ‘never/rarely’, ‘sometimes’, ‘usually’ and ‘prefer not to answer’. For this study, the data were re-categorised into two categories of sleeplessness/insomnia: ‘never/rarely’ or ‘sometimes/usually’.

Covariates

The study included covariates to account for potential factors that might confound the association between the three sleep parameters and cognition in the analyses. These covariates encompassed sex (male/female) (Data Field ID: 31), year of birth (Data Field ID: 34), BMI (Data Field ID: 21001), smoking status (Data Field ID: 20116), alcohol intake frequency (Data Field ID: 1558), diabetes diagnosis (Data Field ID: 2443), vascular/heart diagnosis (heart attack, angina, stroke and high BP) (Data Field ID:6150) and cancer diagnosis (Data Field ID: 2453).

Statistical analysis

Descriptive statistics were used for data characterisation. To normalise the cognitive test scores and ensure comparability, we transformed the raw scores into z-scores using the formula: z-score = (raw score – mean baseline)/baseline standard deviation (SD). This transformation standardised scores from different tests into a unified scale.

Our primary analytical approach was the ordinary least squares (OLS) regression. This model was chosen for its ability to handle multiple predictors and to assess their independent effect on cognitive scores. We included sleep parameters and a range of other covariates such as demographic, health and lifestyle factors to control for potential confounding effects.

To address heteroskedasticity, which can lead to biased standard error (SE) estimates and affect the reliability of test statistics, robust SE were integrated into the regression model. This approach strengthens the validity of our inferences under potential heteroskedastic conditions.

Multicollinearity among independent variables was rigorously evaluated using the Variance Inflation Factor (VIF), with all values confirming below the threshold of 10, indicating that our model did not suffer from multicollinearity issues.

Residual diagnostics were thoroughly conducted. We performed skewness and kurtosis tests to assess the normality of the residuals. Additionally, we visually inspected residual plots against fitted values to confirm the assumptions of linear regression were met. Cook’s distance was used to identify and evaluate the impact of potentially influential observations on the regression model, thereby ensuring its robustness and reliability.

All statistical analyses were performed using Stata/BE 18 software.

Data availability

Full study data cannot be openly shared under the material transfer agreement with UK Biobank. Prospective researchers can apply for access to the UK Biobank data from www.ukbiobank.ac.uk.

Demographics and sample sizes of the UK Biobank data

The present study included 26 820 participants who completed all baseline surveys, including all demographic, lifestyle and comorbidity covariates. Of the 26 820 participants, 10 067 who completed all four cognitive assessments were grouped as Cohort 1 (mean (SD) age 71 (8) years; 56% female (5650/10 067)) while 16 753 who completed only two cognitive assessments were grouped as Cohort 2 (mean (SD) age 72 (8) years; 56% female (9440/16 753)). The baseline demographic background covariates and sleep-related variables of the two independent cohorts from the UK Biobank data are shown in table 1.

Regression diagnostics

Our regression analysis for both cohorts involved several diagnostic checks to confirm the appropriateness of the OLS regression model. For Cohort 1, the Breusch–Pagan test suggested the presence of heteroskedasticity (χ²(1) = 242.37, p<0.0001), prompting the use of robust SE. Cohort 2 showed similar signs of heteroskedasticity with a χ² value of 709.60 (p<0.0001) from the Breusch–Pagan test, indicating the need for robust SE in this group as well.

The VIF was computed for each predictor in both cohorts, with all VIF values remaining below the threshold of 10, suggesting that multicollinearity was not a concern for our regression models. Skewness and kurtosis tests for normality indicated a departure from normality in the residuals for both cohorts (p<0.0001); however, the central limit theorem was anticipated to mitigate the impact of this non-normality due to the large sample sizes.

Cook’s distance was used to detect influential cases, identifying 5.93% of the data points in Cohort 1 and 4.83% in Cohort 2 as potential outliers. These points were, however, retained for analysis as they did not present with data entry errors or anomalies that would warrant exclusion. Visual inspection of residual plots against fitted values did not reveal any discernible patterns, suggesting that the assumptions of linearity were met for the regression models applied to both cohorts.

The study used a multivariable linear regression model with robust SE to evaluate predictors and their relationship with the global cognitive z-score, ensuring a robust inference in the presence of heteroskedasticity as indicated by our diagnostics checks for both cohorts. A comprehensive summary of the regression model results for Cohorts 1 and 2 is presented in table 2.

Association between sleep factors and cognitive performance

In Cohort 1, normal sleep duration was associated with a slightly higher global cognitive score (β coefficient of 0.0567, p<0.001) than short sleepers. In contrast, long sleepers had significantly lower scores (β coefficient of −0.1880, p<0.001). The pattern was similar for long sleepers in Cohort 2 (β coefficient of −0.2619, p<0.001), although normal sleep duration had no significant effect.

Chronotype had a considerable impact on cognitive performance in both cohorts. Intermediate sleepers had higher z-scores (β coefficients of 0.1061 and 0.0632 for Cohorts 1 and 2, p<0.001) compared with morningness, as did the eveningness types (β coefficients of 0.1351 and 0.0750 for Cohort 1, p<0.001 and Cohort 2, p=0.002, respectively). Sleeplessness/insomnia, however, showed no significant association in either cohort.

Association between health and lifestyle factors and cognitive performance

Health and lifestyle factors showed variable correlations with cognitive performance. Gender, age, diabetes and alcohol consumption influenced cognitive performance in both cohorts. Women had lower scores than men (β coefficients of −0.0649 and −0.0761 for Cohorts 1 and 2, p<0.001). Age was inversely correlated with cognitive performance (β coefficients of −0.0177 and −0.0264 for Cohorts 1 and 2, p<0.001). Participants with diabetes had lower cognitive scores in both cohorts (β coefficients of −0.1204, p<0.001 and −0.0847, p=0.011 for Cohorts 1 and 2, respectively). Never or occasional alcohol consumers had significantly lower cognitive scores (β coefficients of −0.2971 and −0.1644 for Cohorts 1 and 2, p<0.001), compared with daily or almost daily consumers. For those who consumed alcohol up to four times a week (weekly), β coefficients of −0.0607 (p<0.001) and −0.0396 (p=0.006) for Cohorts 1 and 2 or up to three times a month (monthly), β coefficients of −0.0802 (p<0.001) and −0.0433 (p=0.034) for Cohorts 1 and 2, respectively, also scored lower in the cognitive tests.

Participants with angina and high BP did significantly worse in the cognitive test only in Cohort 1 (β coefficients −0.1513 and −0.0551, respectively, p<0.001). Current smokers in Cohort 1 had lower scores (β coefficient of −0.1182, p<0.001), while former smokers in Cohort 2 had higher scores (β coefficients of 0.0297, p=0.019) than never smoking. A diagnosis of cancer, heart attack, stroke and BMI did not significantly correlate with the cognitive scores.

Our analysis using predictive margins delineates the relationship between sleep duration and cognitive function as measured by the global z-score (figure 1). For Cohort 1, the data show an inverted U-shaped curve, signifying optimal cognitive function among individuals reporting 6–9 hours of sleep, with performance diminishing among those with shorter or longer sleep durations. In contrast, Cohort 2 shows a relatively stable cognitive function across individuals sleeping 5–9 hours. Notably, there is a pronounced decline in function for those reporting sleep durations beyond 9 hours, with a marked decrease at 3 hours followed by a modest recovery at the 2-hour mark.

Request permission

The predictive margins analysis for chronotype revealed a consistent pattern in cognitive performance in relation to circadian preferences across both cohorts (figure 2). Morning types consistently showed the lowest cognitive scores in both cohorts, with scores improving for intermediate types and reaching higher levels for evening types. The pattern indicates that later chronotypes correlate with better cognitive performance, with this effect being more pronounced in Cohort 1 compared with Cohort 2, where the increase in scores from intermediate to evening types is less substantial.

Request permission

The predictive margins analysis for sleeplessness/insomnia shows that in Cohort 1 there is a minimal decrease in cognitive performance when moving from ‘never/rarely’ experiencing sleeplessness/insomnia to ‘sometimes/usually’ experiencing it (figure 3). For Cohort 2, a similar trend is observed, with a modest decline in cognitive performance for those who ‘sometimes/usually’ experience sleeplessness/insomnia compared with those who ‘never/rarely’ do. However, in both cohorts the change in cognitive performance is minimal and the confidence intervals are broad, suggesting a lack of significant impact of sleeplessness/insomnia on the global cognitive z-score.

Request permission

Transparency

The lead author (RW) affirms that the manuscript is an honest, accurate and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned have been explained.