Chronotype

Responses to two identical questions (“Are you naturally a night person or a morning person?”) were used to define the dichotomous morning person phenotype in the 23andMe cohort, with one question having a wider selection of neutral options. For the first instance, the possible answers were “Night owl”, “Early bird” and “Neither”, and for the second “Night person”, “Morning person”, “Neither”, “It depends” and “I’m not sure”. Individuals with discordant or neutral responses to both were excluded. For those with one neutral and one non-neutral response, their non-neutral response was used to define their phenotype. Morning people were coded as 1 (cases; N = 120,478) and evening people were coded as 0 (controls; N = 127,622).
The UK Biobank collected a single self-reported measure of Chronotype (“Morning/evening person (chronotype)”; data-field 1180). Participants were prompted to answer the question “Do you consider yourself to be?” with one of six possible answers: “Definitely a ‘morning’ person”, “More a ‘morning’ than ‘evening’ person”, “More an ‘evening’ than a ‘morning’ person”, “Definitely an ‘evening’ person”, “Do not know” or “Prefer not to answer”, which we coded as 2, 1, −1, −2, 0 and missing, respectively (distribution summarised in Table 1). Of the 451,454 white European participants with genetic data, 449,734 were included in the GWAS (had non-missing phenotype and covariates).
In order to provide interpretable ORs for our genome-wide significant variants, we also defined a binary phenotype using the same data-field as for Chronotype. Participants answering “Definitely an ‘evening’ person” and “More an ‘evening’ than a ‘morning’ person” were coded as 0 (controls) and those answering “Definitely a ‘morning’ person” and “More a ‘morning’ than ‘evening’ person” were coded as 1 (cases). Participants answering “Do not know” or “Prefer not to answer” were coded as missing. A total of 403,195 participants were included in the GWAS (252,287 cases and 150,908 controls).

We further examined the genes within genome-wide significant loci using gene-based pathway and tissue enrichment analyses^{45 (link),47 (link),69 (link)}. Gene-based analysis was performed using PASCAL, which estimated a combined association P value from the summary statistics of multiple SNPs in a gene^{45 (link)}. Pathway and ontology enrichment analyses were performed using FUMA^{69 (link)} and EnrichR^{47 (link)}. Tissue enrichment analysis was performed using MAGMA^{46 (link)} in FUMA, which controlled for gene size. Pathway and tissue enrichment analyses were also performed on genes within loci belonging to sleep propensity and sleep fragmentation clusters separately.
We constructed a weighted GRS comprising the 42 significant sleepiness loci and tested for associations with other self-reported sleep traits (sleep duration, long sleep duration, short sleep duration, insomnia, chronotype, and day naps), and 7-day accelerometry traits in the UK Biobank. Weighted GRS analyses were performed by summing the products or risk allele count multiplied by the effect estimate reported in the primary GWAS of self-reported daytime sleepiness using R package gds (https://cran.r-project.org/web/packages/gds/gds.pdf). We also tested the GRSs of reported loci for insomnia, sleep duration, short sleep, long sleep, day naps, chronotype, restless legs syndrome (RLS), narcolepsy, and coffee consumption associated with self-reported daytime sleepiness using the same approach. The SNPs selected for each trait include 57 genome-wide significant loci for frequent insomnia^{49 (link)}; 78, 27, and 8 loci for sleep duration, long sleep, and short sleep, respectively^{59 (link)}; 348 loci for chronotype^{67 (link)}; 125 loci for daytime napping; 20 genome-wide significant loci for RLS^{48 (link)}; 8 non-HLA suggestive significant loci (P < 10⁻⁴) in a narcolepsy case–control study of European Americans⁵¹ , and 8 loci for coffee consumption^{50 (link)}.

We performed all association test using BOLT-LMM^{71 (link)} v2.3, which applies a linear mixed model (LMM) to adjust for the effects of population structure and individual relatedness, and allowed us to include all related individuals in our white European subset, boosting our power to detect associations. This meant a sample size of up to 449,734 individuals, as opposed to the set of 379,768 unrelated individuals. BOLT-LMM approximates relatedness within a cohort by using LD blocks and avoids the requirement of building a genetic-relationship matrix, with which calculations are intractable in cohorts of this size. From the ~805,000 directly genotyped (non-imputed) variants available, we identified 524,307 high-quality variants (bi-allelic SNPs; MAF ≥ 1%; HWE P > 1 × 10⁻⁶; non-missing in all genotype batches, total missingness < 1.5% and not in a region of long-range LD^{72 (link)}) that BOLT-LMM used to build its relatedness model. For LD structure information, we used the default 1000 Genomes LD-Score table provided with the software. We forced BOLT-LMM to apply a non-infinitesimal model, which provides better effect size estimates for variants with moderate to large effect sizes, in exchange for increased computing time. At runtime, the chronotype and morning person phenotypes were adjusted for age (field 21003), sex (field 31), study centre (field 54; categorical) and a derived variable representing genotyping release (categorical; UKBiLEVE array, UKB Axiom array interim release and UKB Axiom array full release). Accelerometer-based phenotypes were adjusted at runtime for age activity monitor worn (derived from month and year of birth and date activity monitor worn), sex, season activity monitor worn (categorical; winter, spring, summer or autumn; derived from date activity monitor worn) and number of valid measurements (SPT-windows for sleep phenotypes, number of valid days for diurnal inactivity or number of L5 or M10 detections for L5 or M10 timing). The GWA analysis for the number of sleep episodes phenotype was also adjusted for the mean length of SPT-window (across all included SPT-windows) to account for the fact that individuals have a greater number of sleep episodes the longer they spend in bed.
In the 23andMe morning person GWAS, the summary statistics were generated through logistic regression (using an additive model) of the phenotype against the genotype, adjusting for age, sex, the first four principal components and a categorical variable representing genotyping platform. Genotyping batches in which particular variants failed to meet minimum quality control were not included in association testing for those variants, resulting in a range of sample sizes over the whole set of results. A λ_GC of 1.325 was reported for this GWAS. Lead variants for the 23andMe only morning person GWAS are provided in Supplementary Data 15.

Chronotype was measured using the Korean version [45 (link)] of the Morningness-Eveningness Questionnaire (MEQ-K) [46 ]. The MEQ-K comprises 19 items, with a total scoring range of 16–86 points (11 items [scoring range 1–4 points]; 2 items [scored 0, 2, 4, and 6]; 1 item [scored 0, 2, 3, and 5]; 5 items [scoring range 1–5 points); a higher score indicates a more extreme morning type [46 ]. Based on the criteria of total scores, chronotype was categorized into three groups: morning chronotype (59–86), neither chronotype (42–58), and evening chronotype (16–41) [47 (link)]. The reliability (Cronbach’s α) of the MEQ was 0.82 at the time of development [46 ], 0.77 for MEQ-K [45 (link)], and 0.64 in this study.
The SJL was measured using the Munich Chronotype Questionnaire (MCTQ) [48 (link)]. The MCTQ comprises 14 items about bedtime, sleep onset, sleep latency, time of awakening, time to get up, use of an alarm, and outdoor activity time, to assess weekday and weekend sleep-wake cycles. Cheng and Hang [49 (link)] established the reliability of the MCTQ by confirming that the sleep-wake patterns measured using the scale are closely linked to actigraphy results. SJL was calculated based on the sleep-corrected SJL formula by Jankowski [50 (link)].
Daily light exposure was calculated based on the response on the MCTQ about the time spent outdoors without a roof during the daytime on weekdays and holidays. Sleep duration was calculated as the time from sleep onset to time of awakening based on the response on the MCTQ.

The surveys collected the following meal-timing information during the week and on weekends: timing of breakfast, lunch and dinner (as drop down menu with 1-h intervals, e.g., from 12:00 till 13:00), snacks between meals (“yes”/”no”), snack between breakfast and lunch (“yes”/”no”), snack between lunch and dinner (“yes”/”no”), snack between dinner and breakfast (“yes”/”no”) and timing of last snack of the day (hours, minutes). We used the midpoint of the intervals of the hourly bins (e.g., “12:00–13:00″ was substituted by 12:30) to create pseudocontinuous variables for the time of breakfast, lunch and dinner. The continuous variable nighttime fasting was defined as the time elapsed between the last and the first meal of the day. We created two additional continuous variables: last meal to bed time (hours) and eating midpoint (hours), defined as the midpoint between the first and the last meal of the day. We also generated the discrete variable number of eating occasions (ranging from 0 to 6, the maximum number of eating occasions that could be reported in the surveys), which included main meals and snacks, and a dichotomous variable, breakfast skipping (“skipping” vs. “eating”). Each variable was calculated independently for weekdays and weekends.
The survey collected detailed information on sleep duration, sleep timing and sleep quality. We used the 3rd edition of the International Classification of Sleep Disorders (ICSD-III) [27 ] to define chronic insomnia (“yes”/”no”), as explained in Weitzer et al. [28 (link)]. The surveys also collected information on self-rated health status (“In your opinion: How is your health status in general?”, one answer possible: “very good”/“good”/“moderate”/“bad”/“very bad”) and diagnosed medical conditions (“During the past 12 months, did you have any of the following diseases or conditions?”; multiple answers possible). With this information, we defined the following dichotomous (“yes”/“no”) outcome variables: depression, diabetes, hypertension and bad or very vad self-rated health status. Participants also reported their height and weight, which were used to calculate BMI and define obesity [“yes” (BMI ≥ 30 kg/m²)/ “no” (BMI < 30 kg/m²)].
The survey also collected information on sex and age (“How old are you?”, with respondents asked to fill in their age in the 2017 survey, and to choose one of the following categories in the 2020 survey: “ < 20” / “20–24” / “25–29” / “30–34” / “35–39” / “40–44” / “45–49” / “50–54” / “55–59” / “60–64” / “65–69” / “ ≥ 70”) and other confounders and effect modifiers of interest, i.e. self-rated chronotype (“One hears about “morning” and “evening” types of people. Which ONE of these types do you consider yourself to be?”, one answer possible: “definitely a morning type”/ “rather more a morning than an evening type”/ “rather more an evening than a morning type”/ “definitely an evening type”), marital status (“What is your current marital status?”, one answer possible: “single”/ “married or in a partnership”/ “divorced”/ “widowed”), work status [“What is your current work status?”, multiple answers possible: “(self-) employed full-time” / “(self-) employed part-time” / “retired” / “unemployed” / “student, further training, unpaid work experience” / “disabled” / “in compulsory military or community service” / “household”], alcohol consumption [“How much alcohol do you drink per week? (Please give approximate/average amounts)”, with respondents asked to fill in the number of glasses of beer and wine and shots of liquor/whiskey/gin etc. consumed per week], smoking status (“Do you currently smoke?”, one answer possible: “No, never”/“No, not anymore”/“Yes, I currently smoke”) and history of nightshifts (“Have you ever worked night shifts (schedule including ≥ 3 h of work between 12 pm and 6 am and at least 3 nights/month)?”, one answer possible: “No” / “Yes, in the past” / “Yes, currently”).

Summary statistics [medians and interquartile ranges (IQRs), and frequency (N and %)] were used to describe baseline characteristics and meal-timing patterns, independently for each survey. The correlation between numerical meal-timing variables during the week and during the weekend and between different numerical variables in each survey was analysed graphically (with matrix scatter plots) and the significance of correlation was tested using the Pearson correlation coefficient. The association of these variables with breakfast skipping was assessed using point biserial correlation. The association between breakfast skipping during the week and the weekend was analysed using the χ² test.
Within surveys, cluster analysis was performed to group individuals with similar meal-timing behaviours. Nighttime fasting, last meal to bed time and eating midpoint during the week were standardized and included as indicators. To establish the cluster groups, a combination of hierarchical and non-hierarchical clustering methods was applied. Firstly, we used Ward’s method (hierarchical method) removing univariate outliers (values > 3 SD above or below the mean) and generated the resulting dendrogram, in order to select the optimum number of clusters (two for each survey). Using the initial cluster centres obtained by hierarchical clustering and including also outliers of the variables, an iterative non-hierarchical K-means clustering procedure was applied. The Cohen’s ҡ coefficient for the solutions obtained by hierarchical methods and by non-hierarchical methods (final cluster solution) was 0.96 for the 2017 survey, indicating almost perfect agreement, and 0.77 for the 2020 survey (substantial agreement).
To describe characteristics or predictors of the different cluster groups, as well as participants’ sociodemographic and lifestyle characteristics, summary statistics [medians and interquartile ranges (IQRs), and frequency (N and %)] were used. Differences on the indicators between cluster groups were analysed using the Wilcoxon rank-sum test. Within surveys and using the largest cluster as reference category, unconditional logistic regression analysis was performed to study the association of meal-timing behaviours and chronic insomnia, depression, obesity, diabetes, hypertension and self-rated health status. Logistic regression models were used and odds ratios (OR) with 95% confidence intervals were calculated. Age and sex-adjusted ORs (AORs) and multivariable-adjusted ORs (MV-ORs) are presented. In addition to age and sex and based on a directed acyclic graph, we considered the following potential confounders for the multivariable adjusted model: self-rated chronotype, marital status, work status, alcohol consumption, smoking status and history of nightshifts.
Risk estimates were compared across strata of sex and chronotype profiles (early/late). In sensitivity analyses, only participants without report of heavy alcohol drinking (drinking ≤ 12 standard glasses of alcohol a week) or those who had no history of nightshift were included.
All statistical analyses were performed using STATA 16.