Investigating the non‐linear association between sleep duration and type 2 diabetes: conventional and Mendelian randomization analyses from the UK Biobank

Hiroyuki Kuroda; Shiu Lun Au Yeung; Ryosuke Fujii; Masao Iwagami; Atsushi Goto

doi:10.1111/jdi.70039

. 2025 Apr 3;16(6):1126–1137. doi: 10.1111/jdi.70039

Investigating the non‐linear association between sleep duration and type 2 diabetes: conventional and Mendelian randomization analyses from the UK Biobank

Hiroyuki Kuroda ^1,², Shiu Lun Au Yeung ³, Ryosuke Fujii ^4,⁵, Masao Iwagami ^6,^7,⁸, Atsushi Goto ^1,^2,^✉

PMCID: PMC12131940 PMID: 40181521

ABSTRACT

Aims/Introduction

Previous observational studies have suggested an increased risk of type 2 diabetes associated with both short and long sleep duration. However, there remains uncertainty, particularly regarding the adverse effects of long sleep duration. We investigated the association between self‐reported questionnaire‐based and objectively measured accelerometer‐derived sleep duration and the risk of type 2 diabetes using data from the UK Biobank.

Materials and Methods

First, we performed conventional Cox regression analysis with restricted cubic splines to illustrate the potentially non‐linear association between sleep duration and the risk of type 2 diabetes. Second, we performed non‐linear Mendelian randomization (MR) analysis using the doubly‐ranked method with 85 and 20 genetic variants associated with questionnaire‐based and accelerometer‐based sleep duration, respectively. Third, we performed two‐sample MR analysis.

Results

The results of conventional analysis of accelerometer‐derived sleep duration did not suggest a strong association between longer sleep duration and type 2 diabetes risk (hazard ratio [HR] of ≥10 h compared with 7–8 h, 1.08; 95% confidence interval [CI], 0.92–1.27). The results of non‐linear MR showed no strong evidence for an increased risk of type 2 diabetes associated with questionnaire‐based longer sleep duration (HR of 9 h compared with 7 h, 0.77; 95% CI, 0.52–1.15). This finding was consistent with non‐linear MR of accelerometer‐derived sleep duration (HR of 9 h compared with 7 h, 0.78; 95% CI, 0.29–2.06).

Conclusions

Our findings suggest that longer sleep duration does not play a major role in the development of type 2 diabetes.

Keywords: Diabetes mellitus, type 2; Mendelian randomization; Sleep duration

Our non‐linear Mendelian randomization analysis using data from the UK Biobank showed no strong evidence of an increased risk of type 2 diabetes with longer sleep duration. Our findings suggest that longer sleep duration does not play a major role in the development of type 2 diabetes.

graphic file with name JDI-16-1126-g001.jpg

INTRODUCTION

The prevalence of type 2 diabetes is increasing globally; hence, the necessity for its prevention is being increasingly acknowledged. In 2021, it was estimated that 536.6 million individuals are affected by diabetes. Projections indicate that the number of individuals with diabetes will increase to 783.2 million by 2045 ¹ . Type 2 diabetes represents over 90% of diabetes cases with a major burden on both patients and society ² . Therefore, the imperative of preventing type 2 diabetes is being emphasized.

Sleep duration is a key modifiable lifestyle habit associated with various adverse outcomes ³ , ⁴ , ⁵ . Furthermore, recently, the potential impact of sleep duration on diabetes development has been increasingly examined. Since the 1990s, previous studies have indicated that both adults ⁶ and adolescents ⁷ are receiving 1 h less sleep. Consequently, there is growing concern that an increase in adverse health conditions, including type 2 diabetes, is linked to reduced sleep duration ⁸ .

Both short and long sleep duration have been previously suggested to be associated with an increased risk of type 2 diabetes ⁹ , ¹⁰ , ¹¹ . Biological investigations have indicated that there are multiple pathways through which short sleep duration may influence metabolic diseases ¹² , ¹³ , ¹⁴ . Conversely, there is limited evidence supporting the notion that longer sleep duration increases the risk of type 2 diabetes. It has been proposed that comorbid conditions and socioeconomic status could act as confounders or introduce reverse causation bias in these associations ¹⁵ . Consequently, conclusions regarding the non‐linear association between sleep duration and type 2 diabetes risk are still under discussion.

Although Mendelian randomization (MR) represents a valuable approach for addressing residual confounding and reverse causation, previous MR methods have limitations in examining non‐linear effects on exposures with coarse measurement units, including sleep duration. The direct stratification of exposure can induce collider bias, making this approach inappropriate for MR studies. To address this collider bias, a residual‐based non‐linear MR approach has been proposed. This method involves stratification by instrumental variable (IV)‐free exposure, which is calculated as the residual from regressing exposure on genetic variants. Following this, localized average causal effects (LACE) are estimated for each stratum ¹⁶ . An important assumption of this non‐linear MR approach is that the effect of genetic variants on exposure is constant across exposure strata. However, the categorized nature of coarse exposure, such as questionnaire‐based sleep duration, makes it challenging to calculate the precise IV‐free value for the exposure, potentially violating the previously mentioned assumption ¹⁷ . Conversely, a more recently proposed method, the doubly‐ranked method, can facilitate meaningful stratification even with coarse exposure ¹⁸ . In this method, individuals are first grouped into pre‐strata based on their genotype and then further categorized within each group according to their exposure values. Consequently, this approach addresses the issue of exposure coarseness and helps reduce potential bias.

In this study, we employed UK Biobank data and the non‐linear MR to investigate the association between sleep duration, as measured by both questionnaires and accelerometers, and type 2 diabetes risk. The use of objectively measured accelerometer‐derived sleep duration aimed to address the exposure stratification in non‐linear MR from another perspective. Although the doubly‐ranked method was used to address issues of collider bias and coarse exposure, the risk of bias cannot be eliminated ¹⁹ . Therefore, we further employed negative control analysis to address these issues ¹⁸ , ²⁰ .

MATERIALS AND METHODS

Study population

Data from the UK Biobank were analyzed. The details of the UK Biobank have been published elsewhere ²¹ , ²² , ²³ . Our study included participants who self‐reported as White British and had similar ancestral backgrounds from genetic principal components (PCs). The detailed cohort selection flow is provided in Figure S1.

Definition of exposure and outcome

The primary exposure was questionnaire‐based sleep duration, recorded as integer values in 1‐h increments. Participants who reported extremely short (≤3 h) or long (≥13 h) sleep duration, as well as those who did not respond, were excluded. We also analyzed accelerometer‐derived sleep duration. The details of a sub‐study of accelerometers have been published elsewhere ²⁴ . Approximately 100,000 participants wore accelerometers for 7 days. Accelerometer‐derived sleep duration was estimated using a previously developed and validated machine learning‐based method ²⁵ , ²⁶ . In this study, accelerometer‐derived sleep duration was defined as the mean daily sleep duration recorded within the 4 to 12‐h range.

Outcomes were extracted from hospital inpatient data. The primary outcome was the incidence of type 2 diabetes, defined in accordance with previous studies as the recording of the International Classification of Diseases 10th edition (ICD‐10) code E11 ²⁷ , ²⁸ . Participants with prevalent diabetes (as defined by previously reported algorithms ²⁹ and ICD‐9 or 10 codes as detailed in Table S1) were excluded. Participants who were lost to follow‐up or died were considered censored. Based on the hospital inpatient data available at the time of this study's analysis, the analysis period was defined as the period from each participant's study participation date till October 31, 2022.

Selection of genetic variants

The full data release contains the cohort of successfully genotyped samples (n = 488,377). 49,979 individuals were genotyped using the UK BiLEVE array and 438,398 using the UK Biobank axiom array. Pre‐imputation quality control, phasing, and imputation are described elsewhere ³⁰ .

Based on previous genome‐wide association studies (GWASs) investigating questionnaire‐based or accelerometer‐derived sleep duration in European individuals and the subsequent selection criteria, we used 85 single nucleotide polymorphisms (SNPs) associated with questionnaire‐based sleep duration ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ and 20 SNPs associated with accelerometer‐derived sleep duration ³⁴ , ³⁵ , ³⁷ , ³⁸ . Eligible SNPs were selected with genome‐wide significant P‐values <5.0 × 10⁻⁸ for these GWASs. The following SNPs were excluded: SNPs not included in the 1000 Genomes reference panel, SNPs located on sex chromosomes, indel polymorphisms, multi‐allelic SNPs, SNPs with a minor allele frequency <0.01, and an INFO score <0.7 for imputed SNPs. The candidate SNPs underwent clumping using the “clump_data” function from the TwoSampleMR package with a linkage disequilibrium threshold r ² < 0.001 within a distance of 10,000 kb ³⁹ . To reduce the winner's curse bias, we elected to retain SNPs, including those with P‐values from our data exceeding 5.0 × 10⁻⁸ ⁴⁰ . No proxy SNPs were used in the two‐sample MR.

Statistical analysis

First, we performed conventional Cox regression analysis. In accordance with previous reviews, we adjusted the following factors: age, sex, body mass index (BMI), parental history of diabetes, Townsend deprivation index, household income, educational attainment, employment status, physical activity, tobacco use, alcohol consumption, living with a partner or children, and a history of hypertension ¹⁰ , ¹¹ . The definition of educational attainment is provided in Table S4, ⁴¹ . A history of hypertension was defined as having at least one of the following conditions: a self‐reported diagnosis of hypertension, self‐reported use of antihypertensive medication, or a systolic blood pressure (BP) ≥140 mmHg or a diastolic BP ≥90 mmHg at baseline assessment ⁴² . Any missing values of covariates were imputed using chained random forests ⁴³ . A proportional hazard assumption was evaluated by the scaled Schoenfeld residuals plots. To illustrate the potentially non‐linear association, we further fitted restricted cubic splines with knots at the 5th, 35th, 65th, and 95th percentiles of exposure ⁴⁴ . Sensitivity analyses were conducted using the complete case data.

Second, we performed non‐linear MR analysis. The categorized nature of coarse exposure makes it challenging to precisely estimate IV‐free values. Consequently, the assumption of residual‐based methods that the effect of genetic variants is constant across exposure strata is violated ¹⁷ . Accordingly, we employed the doubly‐ranked method, which allows for effective stratification even with coarse exposure because of the relaxation of the assumptions ¹⁸ . The weighted genetic risk score (GRS) for sleep duration was calculated. Samples were initially stratified into 10 pre‐strata based on their GRS rankings. Within each pre‐stratum, the samples were further re‐stratified into 10 strata based on their exposure ranks. Subsequently, LACE was estimated within each stratum, adjusting for age, sex, assessment center, genotyping batch, and the top 20 genetic PC ⁴⁵ , ⁴⁶ . Fractional polynomial models were fitted to the meta‐analysis of LACE estimates among each stratum. Although the doubly‐ranked method has certain advantages, it may also be susceptible to selection bias ¹⁹ . To ascertain serious violations of the exposure stratification assumptions, a negative control analysis was conducted using sex and hair color as negative control outcomes ²⁰ , ⁴⁷ . Furthermore, excessive strata could lead to some strata having identical exposure values, thus violating the relevance assumption. The number of strata was therefore assessed using the Gelman‐Rubin (GR) uniformity statistic with the threshold of 1.02 ¹⁸ . An analysis of the questionnaire‐based sleep duration, stratified by sex, age, and BMI, was also conducted. To examine the reverse causation bias, a sensitivity analysis was conducted, whereby participants who developed type 2 diabetes within 2 years of study participation were excluded. In the MR analysis, none of the participants had missing covariate values after cohort selection.

Third, we performed a two‐sample MR analysis. SNP‐exposure data were obtained from previously published GWAS on short or long sleep duration ³⁵ . From candidate SNPs with genome‐wide significance, 25 that were associated with short sleep duration and seven that were associated with long sleep duration were selected. SNP‐outcome data were derived from two summary statistics, and their respective MR estimates were synthesized using a meta‐analysis of the random‐effects model. The first SNP‐outcome data were obtained from the DIAGRAM consortium ⁴⁸ . Summary statistics, which excluded the UK Biobank participants, were employed to reduce sample overlap bias ⁴⁰ . Another set of SNP‐outcome data was obtained from the FinnGen study ⁴⁹ .

All analyses were conducted using R software version 4.2.3 (R Foundation for Statistical Computing, Vienna, Austria). The SUMnlmr package was employed to perform non‐linear MR, and the TwoSampleMR package was utilized for SNP selection and two‐sample MR ³⁹ , ⁵⁰ . This study was conducted in accordance with STROBE ⁵¹ and STROBE‐MR ⁵² statements. The corresponding checklists are presented in Tables S11 and S12, respectively.

RESULTS

Cohort selection and baseline characteristics

Table 1 summarizes the baseline characteristics of the overall participants and those with accelerometer‐derived data. As illustrated in the cohort selection flow diagram (Figure S1), 117,231 participants were excluded from the initial 502,366 participants. Consequently, subsequent analyses were conducted on 385,135 participants, of whom 75,094 had accelerometer‐derived data. The details are provided in Tables S2 and S3. During the median follow‐up period of 13.6 years, 16,699 participants (4.3%) were newly diagnosed with type 2 diabetes. The incidence of type 2 diabetes among the participants with accelerometer‐derived data was 3.0% (2,258/75,094), which was lower than that in the overall cohort. Furthermore, this subgroup demonstrated a lower prevalence of hypertension, lower BMI, and lower smoking rates than the overall cohort.

Table 1.

Summary of baseline characteristics of the overall participants and those with accelerometer‐derived data

Variables	Overall	With accelerometer‐derived data
Participants	385,135	75,094
With complete covariates	213,789	46,935
Median (IQR) age at baseline (years)	58 (50, 63)	57 (50, 62)
Female (%)	211,277 (55)	42,485 (57)
Median (IQR) questionnaire‐based sleep duration (h)	7 (7, 8)	7 (7, 8)
Median (IQR) accelerometer‐derived sleep duration (h)	NA	8.40 (7.75, 9.09)
History of hypertension (%)	200,883 (52)	36,106 (48)
Missing	10	0
Median (IQR) BMI	26.6 (24.1, 29.6)	25.9 (23.5, 28.8)
Missing	1114	125
Annual household income before tax (£) (%)
Less than 18,000	71,402 (21)	9,680 (14)
18,000 to 30,999	85,459 (26)	16,614 (24)
31,000 to 51,999	88,912 (27)	19,652 (29)
52,000 to 100,000	68,996 (21)	17,098 (25)
Greater than 100,000	17,747 (5.3)	4,778 (7.0)
Missing	52,619	7,272
Educational attainment as ISCED category (%)
Category 1 (lower)	65,273 (17)	6,322 (8.5)
Category 2	67,014 (18)	11,440 (15)
Category 3	20,596 (5.4)	4,557 (6.1)
Category 4	47,205 (12)	9,780 (13)
Category 5 (higher)	181,628 (48)	42,582 (57)
Missing	3,419	413
Current employment status (%)
In paid employment or self‐employed	221,560 (58)	45,950 (62)
Full‐ or part‐time student	748 (0.2)	170 (0.2)
Retired	133,159 (35)	24,308 (33)
Unemployed	26,914 (7.0)	4,246 (5.7)
Missing	2,754	420
Townsend deprivation index (%)
Q1 (least deprived)	77,106 (20)	16,690 (22)
Q2	76,805 (20)	15,971 (21)
Q3	76,901 (20)	15,236 (20)
Q4	76,945 (20)	14,765 (20)
Q5 (most deprived)	76,928 (20)	12,350 (16)
Missing	450	82
Median (IQR) MET minutes per week for MVPA	960 (240, 2,160)	960 (280, 2,160)
Missing	83,980	12,508
Smoking status (%)
Current	38,505 (10)	4,870 (6.5)
Previous	133,395 (35)	26,593 (35)
Never	211,999 (55)	43,472 (58)
Missing	1,236	159
Alcohol drinker status (%)
Current	361,099 (94)	71,406 (95)
Previous	12,204 (3.2)	1,884 (2.5)
Never	11,559 (3.0)	1,774 (2.4)
Missing	273	30
Father's history of diabetes (%)	30,650 (8.2)	6,255 (8.5)
Missing	9162	1519
Mother's history of diabetes (%)	30,891 (8.1)	5,784 (7.8)
Missing	5,601	946
Living with partner (%)	286,423 (91)	57,978 (92)
Missing	69,158	12,068
Living with child (%)	130,553 (41)	26,754 (42)
Missing	69,158	12,068

Open in a new tab

IQR, interquartile range; ISCED, International Standard Classification for Education; MET, metabolic equivalent task; MVPA, moderate‐to‐vigorous physical activity; NA, not applicable.

Conventional analysis

Tables S5 and S6 provide the results of the Cox regression analyses treating questionnaire‐based and accelerometer‐derived sleep durations as categorical variables, respectively. Relative to 7 h questionnaire‐based sleep duration, both shorter (≤5 h) (hazard ratio [HR], 1.30; 95% confidence interval [CI], 1.23–1.39) and longer (≥10 h) (HR, 1.46; 95% CI, 1.33–1.59) sleep durations were associated with type 2 diabetes risk. By contrast, relative to 7–8 h accelerometer‐derived sleep duration, shorter (<6 h) sleep duration corresponded to an increased type 2 diabetes risk (HR, 1.52; 95% CI, 1.09–2.13), whereas longer (≥10 h) sleep duration did not exhibit a strong association with type 2 diabetes risk (HR, 1.08; 95% CI, 0.92–1.27). The spline curve in Figure 1 shows a U‐shaped association between questionnaire‐based sleep duration and type 2 diabetes risk. By contrast, a J‐shaped association was observed between accelerometer‐derived sleep duration and type 2 diabetes risk, indicating that there is no strong evidence for the association between longer sleep duration and type 2 diabetes risk. P‐values for non‐linearity were less than 0.001 and 0.024 for the questionnaire‐based and accelerometer‐derived sleep duration analyses, respectively. Figures S3 and S4 present sensitivity analyses using complete case data and 8 h as the reference sleep duration, respectively. There were no major changes in the shape of the association compared with the main analysis. The scaled Schoenfeld residuals plots (Figure S2) did not indicate a serious violation of the proportionality assumption.

Non‐linear MR analysis

Figure 2 provides the non‐linear MR analysis of questionnaire‐based sleep duration. The effect of sleep duration on type 2 diabetes risk was linear (P‐value for non‐linearity = 0.94). Furthermore, no strong association between longer sleep duration and type 2 diabetes risk was observed (e.g., HR of 9 h compared with 7 h, 0.77; 95% CI, 0.52–1.15). The GRS, calculated from 85 selected SNPs, explained 0.67% of the variance in exposure, with an F‐statistic of 2589. The GR statistic was 1.001 (<1.02), indicating that the number of strata (10 strata) was not excessive. Figure 3 presents the non‐linear MR analysis of questionnaire‐based sleep duration stratified by sex, age, and BMI. Analysis stratified by any of the variables revealed no major changes. Figure S5 presents sensitivity analyses using 8 h as the reference sleep duration. There was no major change in the shape of the association. Figure S9 shows the non‐linear MR analysis of questionnaire‐based sleep duration, excluding participants who developed the disease within 2 years of study participation. There were no major changes compared with the main analysis. The LACE estimates for Figures 2 and 3 are shown in Figures S6 and S7, respectively. The details of the SNPs associated with the questionnaire‐based sleep duration are provided in Table S7.

Non‐linear MR analysis for questionnaire‐based sleep duration and type 2 diabetes risk stratified by (a) sex, (b) age, and (c) BMI. The figure shows the HRs with a reference sleep duration of 7 h. The doubly‐ranked method was used to stratify exposure. Estimates were adjusted for sex, age, assessment center, genotyping batch, and the top 20 genetic principal components; however, sex or age was excluded from the adjustment when stratified by sex or age, respectively. The y‐axis is on a log scale. BMI, body mass index; HR, hazard ratio; MR, Mendelian randomization.

Figure S8 shows the non‐linear MR analysis of accelerometer‐derived sleep duration. Although the estimates were highly uncertain, the effect of sleep duration on type 2 diabetes risk was linear (P‐value for non‐linearity = 0.72), and no strong effect of longer sleep duration on type 2 diabetes risk was suggested (e.g., HR of 9 h compared with 7 h, 0.78; 95% CI, 0.29–2.06). The GRS calculated from the 20 selected SNPs explained 0.56% of the variance in exposure with an F‐statistic of 427. The GR statistic was 1.000 (<1.02), indicating that the number of strata (10 strata) was not excessive. Figure S10 shows the same analysis using the residual‐based method. There was no major change compared with the rank‐based method. The details of the SNPs associated with accelerometer‐derived sleep duration are provided in Table S8.

Figure S11 illustrates the negative control analysis with sex and hair color as negative control outcomes. The CIs for the odds ratios (ORs) in most strata contained 1.0, indicating a low risk of serious collider bias.

Two‐sample MR analysis

Two palindromic SNPs (rs2186122 and rs9367621) associated with short sleep duration were excluded during the data harmonization phase, and the remaining 23 SNPs were included in the analysis. Seven SNPs associated with long sleep duration were all included in the analysis. The details of the SNPs are provided in Tables S9 and S10. Figure 4 shows forest plots of the two‐sample MR analyses for short or long sleep duration and type 2 diabetes risk. Short sleep duration was associated with type 2 diabetes risk for most methods (OR of inverse variance weighting [IVW], 1.12; 95% CI, 1.02–1.23). Conversely, no strong effect of long sleep duration on type 2 diabetes risk were suggested (OR of IVW, 0.80; 95% CI, 0.60–1.07), although the MR Egger estimate was highly uncertain. P‐values for the MR Egger intercept test were 0.29 and 0.14 for short and long sleep duration, respectively, suggesting no significant average pleiotropic effect.

DISCUSSION

In this study, we investigated the potentially non‐linear association between sleep duration and type 2 diabetes risk with conventional and MR analyses. In contrast to the previous conventional analyses of questionnaire‐based sleep duration, neither the conventional analysis of accelerometer‐derived sleep duration nor the MR analysis provided strong evidence for the association between longer sleep duration and type 2 diabetes risk.

The differences between our findings, which were derived from non‐linear MR analyses and conventional analysis of accelerometers, and those of previous observational studies of questionnaires may be explained by the following reasons. First, the over‐reporting of questionnaire‐based sleep duration may have biased estimates of conventional analysis of questionnaires ⁵³ . Individuals with sedentary behavior or poor sleep quality may report the time spent in bed as sleep time, even when they are not actually asleep. Previous studies have indicated an association between sedentary behavior and self‐reported long sleep duration ⁵⁴ . Therefore, the previously observed association between (over‐reported) questionnaire‐based sleep duration and type 2 diabetes risk may be attributed to these unhealthy habits. We initially conducted a conventional analysis of accelerometer‐derived sleep duration. This analysis indicated that longer sleep duration did not strongly affect type 2 diabetes risk. These results align with a previous study ⁵⁵ , suggesting the potential for bias in the use of self‐reported questionnaire‐based sleep duration in conventional analyses. Second, residual confounding may also explain the increased risk of type 2 diabetes associated with questionnaire‐based longer sleep duration. Our non‐linear MR analysis yielded no strong evidence for an association between longer sleep duration and type 2 diabetes risk. This lack of association was consistently observed in both questionnaire‐based and accelerometer‐derived sleep durations. One advantage of the MR analysis is its robustness against residual confounding and measurement errors ⁵⁶ . Therefore, the MR analysis helped reduce residual confounding related to the aforementioned over‐reporting of questionnaire‐based sleep duration.

Previous studies examined the potentially non‐linear association between sleep duration and type 2 diabetes risk using MR analysis. Liang et al. ⁵⁷ investigated the association between questionnaire‐based sleep duration and hyperglycemia utilizing UK Biobank data and the residual‐based non‐linear MR. They suggested a J‐shaped association in which short sleep duration was associated with hyperglycemia, whereas long sleep duration was not. However, for methodological reasons, their non‐linear MR was performed using only three strata. A small number of strata may lead to limited precision. Moreover, the application of a residual‐based method to questionnaire‐based sleep duration, which is a coarse exposure, is prone to collider bias. Jia et al. ⁵⁸ investigated the association between short or long sleep duration and type 2 diabetes in a two‐sample MR. Their findings indicated that shorter sleep duration has an adverse effect, whereas longer sleep duration had a beneficial effect on type 2 diabetes. However, these estimates are uncertain. We used data from both the DIAGRAM consortium and the FinnGen study as SNP‐outcome data to improve the precision of the estimates.

The principal strength of our study is the triangulation of evidence using a combination of sleep duration data and analytical methods. Given the aforementioned possibility of over‐reporting of self‐reported sleep duration, we employed objectively measured accelerometer‐derived sleep duration. Moreover, we compared the results of the conventional analysis and MR analysis, which are based on different assumptions. Conventional analyses require a no‐residual‐confounding assumption, while MR analysis is robust against residual confounding but requires different assumptions, such as no pleiotropic effects. It has been reported that non‐linear MR analysis should be interpreted using a triangulation framework of evidence in combination with analyses based on different perspectives ⁵⁹ . We believe that the comparable results regarding longer sleep duration obtained from the conventional analysis of accelerometers and non‐linear MR analysis, which were based on different assumptions, support the validity of our findings. Second, the reliability of the non‐linear MR analysis was rigorously evaluated. Non‐linear MR is susceptible to collider bias, and some studies have identified cases in which biased results have been estimated due to inappropriate exposure stratification ¹⁷ . A series of sensitivity analyses, including negative control analyses, were conducted to ascertain the potential for serious selection bias. Moreover, non‐linear MR analysis offers the benefit of enabling stratified analysis. We conducted a stratified analysis of key variables and observed no notable alterations in the association. Third, we leveraged accessible large‐scale data, UK Biobank and FinnGen study, to enhance the precision of the estimates.

Our study has some limitations. First, the incidence of type 2 diabetes may have been underestimated compared with the actual situation. The outcomes of our study were derived from hospital inpatient data. Given that type 2 diabetes is commonly diagnosed in an outpatient setting, it is not feasible to ascertain the incidence of outcomes in non‐hospitalized participants. Nevertheless, this definition has been employed in several studies of type 2 diabetes utilizing UK Biobank data ²⁷ , ²⁸ and is considered reasonable in terms of comparability. Second, the precision of estimates was insufficient, particularly in the non‐linear MR analysis of accelerometer‐derived sleep duration. Although no serious violation of the relevance assumption was identified, the proportion explaining exposure variance was relatively low. Furthermore, the limited sample size of the accelerometer data was considered a factor contributing to the reduced precision. However, the proportion explaining exposure variance was consistent with previous studies ⁶⁰ , and the criteria for SNP selection appeared appropriate. Third, although sleep duration can vary, the time‐varying effects of exposure could not be examined because of exposure information at a single time point. Fourth, our study was conducted with Caucasian participants, which potentially limits the generalizability of our findings if there is an effect modification by ethnicity. Fifth, despite careful consideration, residual bias due to violations of assumptions, particularly in the non‐linear MR analyses, may still exist.

Our findings may enable healthcare providers to offer patients clearer lifestyle advice. Decreases in sleep duration have been reported in several countries ⁶ , ⁷ , underscoring the importance of lifestyle advice to increase sleep in clinical practice. However, advice on avoiding both long and short sleep duration lacks clarity. Based on our findings, healthcare providers can offer clearer advice to sleep‐deprived individuals by simply instructing them to increase sleep duration. This clarity may encourage lifestyle improvements, and our study is considered to have clinical importance in this regard.

In conclusion, our study provides evidence that a longer sleep duration does not play a major role in the development of type 2 diabetes. Previously proposed U‐shaped associations based on conventional analyses of questionnaire‐based sleep duration may be influenced by biases such as systematic over‐reporting. Our findings can provide clearer advice for patients on improving sleep habits to reduce the risk of type 2 diabetes.

DISCLOSURE

The authors declare no conflict of interest.

Approval of the research protocol: The UK Biobank protocol has approval from the North West Multi‐centre Research Ethics Committee (MREC) as a Research Tissue Bank (RTB) approval. This approval means that researchers do not require separate ethical clearance and can operate under the RTB approval (https://www.ukbiobank.ac.uk/learn‐more‐about‐uk‐biobank/about‐us/ethics).

Informed consent: Written informed consent was obtained from all study participants by the UK Biobank.

Registry and the registration no. of the study/trial: The latest ethics renewal of the UK Biobank protocol was approved on June 29, 2021 (REC reference, 21/NW/0157).

Animal studies: N/A.

Supporting information

Table S1 | Definitions of prevalent diabetes and incident type 2 diabetes.

Table S2 | Detailed baseline characteristics of the overall participants (N = 385,135).

Table S3 | Detailed baseline characteristics of participants with accelerometer‐derived data (N = 75,094).

Table S4 | International Standard for Classification of Education codes mapped to UK Biobank self‐report highest qualification.

Table S5 | Cox regression analysis for the association between questionnaire‐based sleep duration as a categorical variable and the risk of type 2 diabetes.

Table S6 | Cox regression analysis for the association between accelerometer‐derived sleep duration as a categorical variable and the risk of type 2 diabetes.

Table S7 | Associations between questionnaire‐based sleep duration and each SNP that had genome‐wide significance in previous GWASs.

Table S8 | Associations between accelerometer‐derived sleep duration and each SNP that had genome‐wide significance in previous GWASs.

Table S9 | Associations between short (<7 h) relative to normal (7–8 h) sleep duration and each SNP that had genome‐wide significance in previous GWAS.

Table S10 | Associations between long (≥9 h) relative to normal (7–8 h) sleep duration and each SNP that had genome‐wide significance in previous GWAS.

Table S11 | STROBE checklist of items that should be included in reports of cohort studies.

Table S12 | STROBE‐MR checklist of recommended items to address in reports of Mendelian randomization studies.

Figure S1 | Flow diagram of the UK Biobank cohort selection.

Figure S2 | Scaled Schoenfeld residuals plots of the sleep duration.

Figure S3 | Shape of the association between sleep duration and the risk of type 2 diabetes based on the conventional analysis using the complete case data.

Figure S4 | Shape of the association between sleep duration and the risk of type 2 diabetes based on the conventional analysis with reference sleep duration of 8 h.

Figure S5 | Non‐linear MR analysis for questionnaire‐based sleep duration and type 2 diabetes risk with reference sleep duration of 8 h.

Figure S6 | Local average causal effect estimates of non‐linear MR analysis for questionnaire‐based sleep duration and the risk of type 2 diabetes.

Figure S7 | Local average causal effect estimates of non‐linear MR analysis for questionnaire‐based sleep duration and the risk of type 2 diabetes stratified by sex, age, and BMI.

Figure S8 | Non‐linear MR analysis for accelerometer‐derived sleep duration and type 2 diabetes risk.

Figure S9 | Non‐linear MR analysis for questionnaire‐based sleep duration and the risk of type 2 diabetes, excluding participants who developed the disease within 2 years of study participation.

Figure S10 | Non‐linear MR analysis for accelerometer‐derived sleep duration and the risk of type 2 diabetes, employing the residual method.

Figure S11 | Negative control analysis for non‐linear MR analysis.

JDI-16-1126-s001.pdf^{(2.2MB, pdf)}

ACKNOWLEDGMENTS

This work was supported by Japan Society for the Promotion of Science (JSPS) Grants‐in‐Aid for Scientific Research (KAKENHI) Grant Number 24K23752 and 24K02708, and the National Cancer Center Research and Development Fund. We would like to thank Editage (www.editage.jp) for English language editing.

REFERENCES

1. Sun H, Saeedi P, Karuranga S, et al. IDF diabetes atlas: global, regional and country‐level diabetes prevalence estimates for 2021 and projections for 2045. Diabetes Res Clin Pract 2022; 183: 109119. [DOI] [PMC free article] [PubMed] [Google Scholar]
2. Chatterjee S, Khunti K, Davies MJ. Type 2 diabetes. Lancet 2017; 389: 2239–2251. [DOI] [PubMed] [Google Scholar]
3. Yin J, Jin X, Shan Z, et al. Relationship of sleep duration with all‐cause mortality and cardiovascular events: a systematic review and dose‐response meta‐analysis of prospective cohort studies. J Am Heart Assoc 2017; 6: 6. [DOI] [PMC free article] [PubMed] [Google Scholar]
4. Cappuccio FP, Cooper D, D'Elia L, et al. Sleep duration predicts cardiovascular outcomes: a systematic review and meta‐analysis of prospective studies. Eur Heart J 2011; 32: 1484–1492. [DOI] [PubMed] [Google Scholar]
5. Cappuccio FP, D'Elia L, Strazzullo P, et al. Sleep duration and all‐cause mortality: a systematic review and meta‐analysis of prospective studies. Sleep 2010; 33: 585–592. [DOI] [PMC free article] [PubMed] [Google Scholar]
6. Bin YS, Marshall NS, Glozier N. Secular trends in adult sleep duration: a systematic review. Sleep Med Rev 2012; 16: 223–230. [DOI] [PubMed] [Google Scholar]
7. Matricciani L, Olds T, Petkov J. In search of lost sleep: secular trends in the sleep time of school‐aged children and adolescents. Sleep Med Rev 2012; 16: 203–211. [DOI] [PubMed] [Google Scholar]
8. St‐Onge MP, Grandner MA, Brown D, et al. Sleep duration and quality: impact on lifestyle behaviors and cardiometabolic health: a scientific statement from the American Heart Association. Circulation 2016; 134: e367–e386. [DOI] [PMC free article] [PubMed] [Google Scholar]
9. Lu H, Yang Q, Tian F, et al. A meta‐analysis of a cohort study on the association between sleep duration and type 2 diabetes mellitus. J Diabetes Res 2021; 2021: 8861038. [DOI] [PMC free article] [PubMed] [Google Scholar]
10. Shan Z, Ma H, Xie M, et al. Sleep duration and risk of type 2 diabetes: a meta‐analysis of prospective studies. Diabetes Care 2015; 38: 529–537. [DOI] [PubMed] [Google Scholar]
11. Cappuccio FP, D'Elia L, Strazzullo P, et al. Quantity and quality of sleep and incidence of type 2 diabetes: a systematic review and meta‐analysis. Diabetes Care 2010; 33: 414–420. [DOI] [PMC free article] [PubMed] [Google Scholar]
12. Spiegel K, Tasali E, Penev P, et al. Brief communication: sleep curtailment in healthy young men is associated with decreased leptin levels, elevated ghrelin levels, and increased hunger and appetite. Ann Intern Med 2004; 141: 846–850. [DOI] [PubMed] [Google Scholar]
13. Irwin MR, Opp MR. Sleep health: reciprocal regulation of sleep and innate immunity. Neuropsychopharmacology 2017; 42: 129–155. [DOI] [PMC free article] [PubMed] [Google Scholar]
14. Tobaldini E, Costantino G, Solbiati M, et al. Sleep, sleep deprivation, autonomic nervous system and cardiovascular diseases. Neurosci Biobehav Rev 2017; 74: 321–329. [DOI] [PubMed] [Google Scholar]
15. Stranges S, Dorn JM, Shipley MJ, et al. Correlates of short and long sleep duration: a cross‐cultural comparison between the United Kingdom and the United States: the Whitehall II study and the Western New York health study. Am J Epidemiol 2008; 168: 1353–1364. [DOI] [PMC free article] [PubMed] [Google Scholar]
16. Burgess S, Davies NM, Thompson SG, et al. Instrumental variable analysis with a nonlinear exposure‐outcome relationship. Epidemiology 2014; 25: 877–885. [DOI] [PMC free article] [PubMed] [Google Scholar]
17. Davey Smith G. Non‐linear Mendelian randomization publications on vitamin D report spurious findings and require major correction. Eur Heart J 2024; 45: 2677–2678. [DOI] [PubMed] [Google Scholar]
18. Tian H, Mason AM, Liu C, et al. Relaxing parametric assumptions for non‐linear Mendelian randomization using a doubly‐ranked stratification method. PLoS Genet 2023; 19: e1010823. [DOI] [PMC free article] [PubMed] [Google Scholar]
19. Burgess S, Sun YQ, Zhou A, et al. Body mass index and all‐cause mortality in HUNT and UK biobank studies: revised non‐linear Mendelian randomisation analyses. BMJ Open 2024; 14: e081399. [DOI] [PMC free article] [PubMed] [Google Scholar]
20. Hamilton FW, Hughes DA, Spiller W, et al. Non‐linear Mendelian randomization: detection of biases using negative controls with a focus on BMI, vitamin D and LDL cholesterol. Eur J Epidemiol 2024; 39: 451–465. [DOI] [PMC free article] [PubMed] [Google Scholar]
21. Sudlow C, Gallacher J, Allen N, et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med 2015; 12: e1001779. [DOI] [PMC free article] [PubMed] [Google Scholar]
22. Fry A, Littlejohns TJ, Sudlow C, et al. Comparison of sociodemographic and health‐related characteristics of UK biobank participants with those of the general population. Am J Epidemiol 2017; 186: 1026–1034. [DOI] [PMC free article] [PubMed] [Google Scholar]
23. Collins R. What makes UK biobank special? Lancet 2012; 379: 1173–1174. [DOI] [PubMed] [Google Scholar]
24. Doherty A, Jackson D, Hammerla N, et al. Large scale population assessment of physical activity using wrist worn accelerometers: the UK biobank study. PLoS One 2017; 12: e0169649. [DOI] [PMC free article] [PubMed] [Google Scholar]
25. Willetts M, Hollowell S, Aslett L, et al. Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK biobank participants. Sci Rep 2018; 8: 7961. [DOI] [PMC free article] [PubMed] [Google Scholar]
26. Walmsley R, Chan S, Smith‐Byrne K, et al. Reallocation of time between device‐measured movement behaviours and risk of incident cardiovascular disease. Br J Sports Med 2021; 56: 1008–1017. [DOI] [PMC free article] [PubMed] [Google Scholar]
27. Ma H, Li X, Zhou T, et al. Glucosamine use, inflammation, and genetic susceptibility, and incidence of type 2 diabetes: a prospective study in UK biobank. Diabetes Care 2020; 43: 719–725. [DOI] [PMC free article] [PubMed] [Google Scholar]
28. Wang M, Zhou T, Li X, et al. Baseline vitamin D status, sleep patterns, and the risk of incident type 2 diabetes in data from the UK biobank study. Diabetes Care 2020; 43: 2776–2784. [DOI] [PMC free article] [PubMed] [Google Scholar]
29. Eastwood SV, Mathur R, Atkinson M, et al. Algorithms for the capture and adjudication of prevalent and incident diabetes in UK biobank. PLoS One 2016; 11: e0162388. [DOI] [PMC free article] [PubMed] [Google Scholar]
30. Bycroft C, Freeman C, Petkova D, et al. The UK biobank resource with deep phenotyping and genomic data. Nature 2018; 562: 203–209. [DOI] [PMC free article] [PubMed] [Google Scholar]
31. Gottlieb DJ, Hek K, Chen TH, et al. Novel loci associated with usual sleep duration: the CHARGE consortium genome‐wide association study. Mol Psychiatry 2015; 20: 1232–1239. [DOI] [PMC free article] [PubMed] [Google Scholar]
32. Jones SE, Tyrrell J, Wood AR, et al. Genome‐wide association analyses in 128,266 individuals identifies new Morningness and sleep duration loci. PLoS Genet 2016; 12: e1006125. [DOI] [PMC free article] [PubMed] [Google Scholar]
33. Hammerschlag AR, Stringer S, de Leeuw CA, et al. Genome‐wide association analysis of insomnia complaints identifies risk genes and genetic overlap with psychiatric and metabolic traits. Nat Genet 2017; 49: 1584–1592. [DOI] [PMC free article] [PubMed] [Google Scholar]
34. Doherty A, Smith‐Byrne K, Ferreira T, et al. GWAS identifies 14 loci for device‐measured physical activity and sleep duration. Nat Commun 2018; 9: 5257. [DOI] [PMC free article] [PubMed] [Google Scholar]
35. Dashti HS, Jones SE, Wood AR, et al. Genome‐wide association study identifies genetic loci for self‐reported habitual sleep duration supported by accelerometer‐derived estimates. Nat Commun 2019; 10: 1100. [DOI] [PMC free article] [PubMed] [Google Scholar]
36. Jansen PR, Watanabe K, Stringer S, et al. Genome‐wide analysis of insomnia in 1,331,010 individuals identifies new risk loci and functional pathways. Nat Genet 2019; 51: 394–403. [DOI] [PubMed] [Google Scholar]
37. Jones SE, van Hees VT, Mazzotti DR, et al. Genetic studies of accelerometer‐based sleep measures yield new insights into human sleep behaviour. Nat Commun 2019; 10: 1585. [DOI] [PMC free article] [PubMed] [Google Scholar]
38. Li X, Zhao H. Automated feature extraction from population wearable device data identified novel loci associated with sleep and circadian rhythms. PLoS Genet 2020; 16: e1009089. [DOI] [PMC free article] [PubMed] [Google Scholar]
39. Hemani G, Zheng J, Elsworth B, et al. The MR‐base platform supports systematic causal inference across the human phenome. elife 2018; 7: 7. [DOI] [PMC free article] [PubMed] [Google Scholar]
40. Burgess S, Davies NM, Thompson SG. Bias due to participant overlap in two‐sample Mendelian randomization. Genet Epidemiol 2016; 40: 597–608. [DOI] [PMC free article] [PubMed] [Google Scholar]
41. Carter AR, Gill D, Davies NM, et al. Understanding the consequences of education inequality on cardiovascular disease: Mendelian randomisation study. BMJ 2019; 365: l1855. [DOI] [PMC free article] [PubMed] [Google Scholar]
42. Tapela N, Collister J, Clifton L, et al. Prevalence and determinants of hypertension control among almost 100 000 treated adults in the UK. Open Heart 2021; 8: e001461. [DOI] [PMC free article] [PubMed] [Google Scholar]
43. Stekhoven DJ, Buhlmann P. MissForest – non‐parametric missing value imputation for mixed‐type data. Bioinformatics 2012; 28: 112–118. [DOI] [PubMed] [Google Scholar]
44. Harrell FE. In: Harrell FE (ed). Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. Cham: Springer International Publishing, 2015; 13–44. [Google Scholar]
45. Haworth S, Mitchell R, Corbin L, et al. Apparent latent structure within the UK biobank sample has implications for epidemiological analysis. Nat Commun 2019; 10: 333. [DOI] [PMC free article] [PubMed] [Google Scholar]
46. Burgess S, Davey Smith G, Davies NM, et al. Guidelines for performing Mendelian randomization investigations: update for summer 2023. Wellcome Open Res 2019; 4: 186. [DOI] [PMC free article] [PubMed] [Google Scholar]
47. Sanderson E, Richardson TG, Hemani G, et al. The use of negative control outcomes in Mendelian randomization to detect potential population stratification. Int J Epidemiol 2021; 50: 1350–1361. [DOI] [PMC free article] [PubMed] [Google Scholar]
48. Mahajan A, Taliun D, Thurner M, et al. Fine‐mapping type 2 diabetes loci to single‐variant resolution using high‐density imputation and islet‐specific epigenome maps. Nat Genet 2018; 50: 1505–1513. [DOI] [PMC free article] [PubMed] [Google Scholar]
49. Kurki MI, Karjalainen J, Palta P, et al. FinnGen provides genetic insights from a well‐phenotyped isolated population. Nature 2023; 613: 508–518. [DOI] [PMC free article] [PubMed] [Google Scholar]
50. Hemani G, Tilling K, Davey Smith G. Orienting the causal relationship between imprecisely measured traits using GWAS summary data. PLoS Genet 2017; 13: e1007081. [DOI] [PMC free article] [PubMed] [Google Scholar]
51. von Elm E, Altman DG, Egger M, et al. The strengthening the reporting of observational studies in epidemiology (STROBE) statement: guidelines for reporting observational studies. Lancet 2007; 370: 1453–1457. [DOI] [PubMed] [Google Scholar]
52. Skrivankova VW, Richmond RC, Woolf BAR, et al. Strengthening the reporting of observational studies in epidemiology using Mendelian randomization: the STROBE‐MR statement. JAMA 2021; 326: 1614–1621. [DOI] [PubMed] [Google Scholar]
53. Lauderdale DS, Knutson KL, Yan LL, et al. Self‐reported and measured sleep duration: how similar are they? Epidemiology 2008; 19: 838–845. [DOI] [PMC free article] [PubMed] [Google Scholar]
54. Lakerveld J, Mackenbach JD, Horvath E, et al. The relation between sleep duration and sedentary behaviours in European adults. Obes Rev 2016; 17(Suppl 1): 62–67. [DOI] [PubMed] [Google Scholar]
55. Jin X, Chen Y, Feng H, et al. Association of accelerometer‐measured sleep duration and different intensities of physical activity with incident type 2 diabetes in a population‐based cohort study. J Sport Health Sci 2024; 13: 222‐232. [DOI] [PMC free article] [PubMed] [Google Scholar]
56. Smith GD, Ebrahim S. ‘Mendelian randomization’: can genetic epidemiology contribute to understanding environmental determinants of disease? Int J Epidemiol 2003; 32: 1–22. [DOI] [PubMed] [Google Scholar]
57. Liang YY, Chen J, Peng M, et al. Association between sleep duration and metabolic syndrome: linear and nonlinear Mendelian randomization analyses. J Transl Med 2023; 21: 90. [DOI] [PMC free article] [PubMed] [Google Scholar]
58. Jia Y, Guo D, Sun L, et al. Self‐reported daytime napping, daytime sleepiness, and other sleep phenotypes in the development of cardiometabolic diseases: a Mendelian randomization study. Eur J Prev Cardiol 2022; 29: 1982–1991. [DOI] [PubMed] [Google Scholar]
59. Burgess S. Towards more reliable non‐linear Mendelian randomization investigations. Eur J Epidemiol 2024; 39: 447–449. [DOI] [PMC free article] [PubMed] [Google Scholar]
60. Henry A, Katsoulis M, Masi S, et al. The relationship between sleep duration, cognition and dementia: a Mendelian randomization study. Int J Epidemiol 2019; 48: 849–860. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials