Sleep Duration and Health Outcomes by Sex in the KoGES Ansan Cohort: A Cross-Sectional Interaction Analysis
Article information
Abstract
Objective
To test whether sex modifies the association between habitual sleep duration and health outcomes in middle-aged Korean adults.
Methods
We analyzed cross-sectional data from 3,709 participants in the Ansan arm of the Korean Genome and Epidemiology Study Phase 2 (2003–2004). Sleep duration was calculated as a weighted 7-day average and categorized as short (<7 hr), reference (7 to <9 hr), or long (≥9 hr). Primary outcomes were body mass index, fasting plasma glucose, a fatigue-related symptom score, and forced expiratory volume in 1 second/forced vital capacity (FEV1/FVC) ratio. The primary analysis tested the sleep category×sex interaction using 2-df likelihood-ratio tests adjusted for age, income, marital status, occupation, smoking, drinking, and exercise. Secondary outcomes were controlled using the Benjamini–Hochberg false discovery rate.
Results
In the primary fully adjusted model, sleep duration×sex interactions were not significant for body mass index (p=0.930), fasting glucose (p=0.388), fatigue-related symptom score (p=0.104), or FEV1/FVC ratio (p=0.855). A complementary short-versus-reference 1-df interaction contrast was nominally significant for fatigue-related symptom score (p=0.047). In an explicitly post hoc age-stratified analysis, the fatigue-related symptom score interaction was concentrated among participants aged <50 years (p=0.001): short sleep was associated with higher adjusted fatigue-related symptom scores in women (+4.75 points; 95% confidence interval [CI], +2.33 to +7.16) but not in men (-0.50; 95% CI, -2.52 to +1.52).
Conclusion
This analysis found no evidence of broad sex modification for cardiometabolic or pulmonary outcomes. The younger-adult fatigue-related symptom score signal should be considered hypothesis-generating pendingreplication with validated fatigue measures and longitudinal data. The fatigue-related symptom score used in this study was not a validated fatigue instrument, which limits the interpretability of the fatigue findings.
INTRODUCTION
Habitual sleep duration is a modifiable behavior with established links to morbidity and mortality [1,2]. The National Sleep Foundation recommends 7–9 hours for healthy adults [3]. Three prospective meta-analyses anchor the effect-direction and magnitude of this recommendation. Cappuccio and colleagues [1] found that short sleepers had 48% higher odds of incident coronary heart disease and 15% higher odds of stroke, with symmetric elevations for long sleepers. The all-cause mortality meta-analysis from the same group reported pooled relative risks of 1.12 for short and 1.30 for long sleep [2]. Yin and colleagues [4], in a dose-response synthesis, showed a J-shaped curve nadir near 7 hours for both all-cause mortality and cardiovascular events. These findings are robust across European and North American cohorts but rest heavily on sex-aggregated estimates or sex-stratified estimates that were never formally compared.
The following mechanistic discussion is exploratory and hypothesis-generating; none of the pathways described below were directly tested in the present data. A sex-medicine reading of the same evidence suggests several distinct biological pathways by which a sleep×sex interaction could arise, each most plausible at a particular life-stage. First, gonadal hormones influence slowwave sleep and rapid-eye-movement sleep architecture, with estradiol and progesterone fluctuations across the menstrual cycle and the menopausal transition producing measurable sleep-stage differences between women and men of the same age [5,6]. Second, women show a stronger cortisol response to psychosocial stress across the reproductive years, and the hypothalamic–pituitary–adrenal (HPA) axis is both an output of and an input into sleep regulation [7,8]. Third, autonomic tone during sleep differs between sexes: women exhibit higher vagal dominance in nonrapid eye movement (REM) sleep, which modulates sympathetic-parasympathetic balance in ways that could plausibly moderate the cardiometabolic cost of short sleep. Fourth, experimental sleep-restriction studies show sex differences in the inflammatory response—particularly in interleukin-6 and C-reactive protein (CRP)—that may compound over habitual short sleep and reach detectable thresholds only in younger adults whose baseline inflammatory tone is lower. These four mechanisms map onto the life-stage of peak activity (reproductive-age effects are most pronounced in women under 50; autonomic and inflammatory effects persist but attenuate) and therefore predict that a sex×sleep interaction, if present, should be concentrated in neurobehavioral outcomes and in younger adults.
In Korea, sleep epidemiology has matured rapidly over the past two decades. Community surveys report high prevalence of short sleep and insomnia [9,10], and population data demonstrate bidirectional links between sleep duration and quality of life [11]. The Ansan arm of the Korean Genome and Epidemiology Study (KoGES) [12] has been analyzed extensively for sleep-related questions: prior analyses characterized Ansan sleep duration and daytime sleepiness and their cardiovascular correlates in the early 2000s; Kim and colleagues [12] documented the cohort profile and follow-up design; and Lee and colleagues [13] reported a stress–insomnia–quality-of-life mediation pathway. None of these analyses has formally pre-specified a sleep×sex interaction test across cardiometabolic, pulmonary, and neurobehavioral domains simultaneously. That gap matters both epidemiologically—sex-modification claims in the Korean sleep literature have rested on sex-stratified descriptive statistics rather than on interaction inference—and methodologically, because formal interaction testing is the only statistically defensible way to compare between-stratum slopes [14,15].
We therefore conducted an a priori interaction analysis in Ansan KoGES Phase 2 to test whether the association between three-category sleep duration and each of four a-priori health outcomes—body mass index (BMI), fasting glucose, a fatigue-related symptom score, and forced expiratory volume in 1 second/forced vital capacity (FEV1/FVC) ratio—differs between men and women. We hypothesized that 1) cardiometabolic and pulmonary sleep–health associations would not be materially modified by sex in the Korean middle-aged population given the modest magnitude of sex differences reported for these phenotypes in prior Korean work, and 2) any sex-specific signal, if present, would concentrate in the fatigue domain and in younger adults. We report the primary interaction tests, sex-stratified effect estimates, and seven sensitivity analyses. Reporting follows STrengthening the Reporting of OBservational studies in Epidemiology [16] and Sex and Gender Equity in Research [17] guidelines.
METHODS
Study design and participants
KoGES is a nationwide prospective community-based cohort of adults aged 40–69 years residing in the Ansan (urban) and Ansung (rural) areas, enrolled in 2001–2002 with biennial follow-up [12]. The two sites differ in occupational structure, physicalactivity patterns, and examination logistics, including spirometry equipment models and interview scheduling. The present analysis used Phase 2 data (2003–2004), the earliest wave with simultaneous availability of sleep-duration items, anthropometrics, spirometry, fasting biochemistry, and psychosocial questionnaires.
We restricted the primary analysis to the Ansan site for three reasons. First, the Ansan and Ansung sites used different spirometry devices, and pooling FEV1/FVC data across devices would introduce measurement heterogeneity into one of the four primary outcomes. Second, interview protocols and scheduling differed between sites, which may affect self-reported sleep-duration estimates. Third, restricting to Ansan maintains continuity with the existing Ansan-focused KoGES literature on sleep and quality of life [13]. We recognize that this decision reduces sample size—particularly in the long-sleep stratum—and limits generalizability to rural populations. To address this, a pooled Ansan+Ansung analysis with site as a covariate is pre-registered as a sensitivity component of the planned longitudinal follow-up.
Of 7,586 Phase 2 participants, we applied pre-specified exclusions in sequence (Figure 1): 1) non-Ansan residence (n=3,855 excluded); 2) missing sex (n=0); 3) age outside 30–80 years (n= 0); 4) missing self-reported sleep duration (n=2); and 5) implausible sleep duration <3 or >14 hours/day (n=20). The final analytic sample comprised 3,709 participants (1,959 men, 1,750 women).
Sleep-duration assessment
Habitual sleep duration was self-reported in the Phase 2 questionnaire with separate items for weekdays and weekends. Participants responded in whole- or half-hour increments. A weighted 7-day average was computed as (5×weekday+2×weekend)/7. Following the three anchor meta-analyses [1,4,18], the continuous variable was categorized a priori as short (<7 hr), reference (7 to <9 hr), and long (≥9 hr); the exact boundary at 7 hours places a 6.99 hours respondent in the short category. The reference interval was literature-anchored to avoid post hoc category optimization.
Outcomes
Primary outcomes
The primary outcomes were assessed as follows: 1) BMI (kg/m2) from directly measured height and weight. 2) Fasting plasma glucose (FPG, mg/dL) after an overnight fast. 3) A fatigue-related symptom score derived from the KoGES Phase 2 “tiredness/ fatigue” symptom block, in which participants rated multiple tiredness- and energy-related items on a Likert intensity scale; the items were summed without reverse-coding to yield a composite where higher values indicate greater subjective fatigue. In this sample the observed sum-score ranged 30–120 (integer), with mean 69.3 and standard deviation (SD) 18.0 (n=3,705). 4) FEV1/FVC ratio (%) by standardized spirometry performed prospectively per the ATS/ERS 2005 standardization [19].
Rationale for the four-outcome primary tier
The selection spans three organ-system axes implicated in the sleep-duration literature: BMI and fasting glucose represent the cardiometabolic axis most directly tied to the Cappuccio et al. [1], Yin et al. [4], and Shan et al. [20], meta-analyses; FEV1/FVC represents the pulmonary axis, underrepresented in the sex-difference literature but measured to ATS/ERS specifications in KoGES; and the fatigue-related symptom score represents the neurobehavioral axis, in which sex differences are most consistently reported in prior work [21,22]. The outcome list was fixed before any interaction test was run.
Limitation of the fatigue-related symptom score
The sum-score is not a validated fatigue instrument such as the Fatigue Severity Scale or the Epworth Sleepiness Scale; KoGES Phase 2 does not include either. We report the sum-score because it is the best available Phase 2 proxy for subjective fatigue and is directly comparable across participants, but we explicitly caveat this limitation in the Discussion. Cronbach’s α at the item level cannot be computed from the derived analytic file used here because only the sum-score was retained in the preprocessing pipeline; item-level reliability reporting is committed to as part of the planned longitudinal follow-up.
Secondary outcomes
The secondary outcomes were assessed as follows: 1) Anthropometry and blood pressure (waist circumference, systolic blood pressure [SBP)], diastolic blood pressure [DBP]); 2) lipid/metabolic panel (hemoglobin A1c [HbA1c], total cholesterol, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, CRP); 3) pulmonary function (FEV1, FVC, FEV1 %predicted); 4) self-rated health and quality of life (World Health Organization Quality of Life-Brief version [WHOQOL-BREF] total mean score [23] using the Korean validation [24] and three individual items); plus four binary clinical endpoints (measured hypertension [HTN] [SBP ≥140 or DBP ≥90], fasting diabetes [FPG ≥126 mg/dL], obesity [BMI ≥25 kg/m2], airflow obstruction [FEV1/FVC <0.70]).
Covariates
Covariates were pre-specified and grouped into two adjustment sets. Model 1 (M1; minimally adjusted) included age (continuous, years). Model 2 (M2; primary, fully adjusted) additionally included household income (four ordered bands), marital status (married vs. other), occupational group in five categories (1: managerial/professional; 2: clerical; 3: service/sales; 4: skilled/manual; 5: homemaker/unemployed/other), smoking status (never smoked/former/occasional/current), drinking status (never/past/current), and current regular exercise (yes/no). We used drinking status rather than drinking quantity (g/day) because the derived analytic file retained status only; a drinking-quantity sensitivity analysis is committed to in the planned longitudinal follow-up after re-extraction from raw source. BMI was intentionally excluded from M2 for all non-BMI outcomes because it is on the causal path from sleep duration to cardiometabolic outcomes (directed acyclic graph); its inclusion in a marginal model would bias the total-effect estimate away from the causal quantity of interest [25]. BMI adjustment was separately examined as sensitivity analysis S4.
Statistical analysis
Descriptive analyses
Baseline characteristics were tabulated by sex as mean±SD (continuous) or n (%) (categorical). Sex differences were tested by Welch’s t-test and Pearson’s χ2 test, respectively.
Primary interaction test
For each primary outcome under each adjustment set, we tested the sleep category×sex interaction using a likelihood-ratio (LR) χ2 test with 2 degrees of freedom. Two nested generalized linear models were compared: a reduced model containing sleep category, sex, and covariates, and a full model that added the sleep × sex interaction terms. An identity link was used for continuous outcomes and a logit link for binary outcomes. The LR test was chosen because it is invariant to the choice of reference category, unlike coefficient-level Wald tests [14]. All models used completecase observations; analytic N was 3,652 for BMI, 3,425 for fasting glucose, 3,700 for the fatigue-related symptom score, and 3,542 for FEV1/FVC. Full results for all 24 outcomes are provided in Table 1 (primary) and Supplementary Table 1 (secondary).
Complementary 1-df short-vs.-reference contrast.
Because the long-sleep stratum was sparse (n=53), the 2-df omnibus test distributes power across both the short and long contrasts. We therefore additionally tested a 1-df short-vs.-reference sex-interaction contrast by restricting the sample to short and reference sleepers only. This analysis was considered supportive and is interpreted cautiously.
Minimum detectable effect
Using the observed stratum sizes and outcome standard deviations, we computed the smallest sex-difference in each sleep contrast detectable at α=0.05 and 80% power (Supplementary Table 2). These values provide a concrete floor for interpreting null results.
Sex-stratified effect estimates
For interpretation, M1 and M2 were also fitted separately in men and women, with β (continuous) or odds ratios (binary) and 95% confidence intervals (CIs) for short-vs.-reference and long-vs.-reference contrasts. Estimates are displayed as paired forest plots (Figure 2).
Sex-stratified forest plot of primary outcomes. β coefficients (95% CI) for short (<7 hr) and long (≥9 hr) sleep vs. reference (7 to <9 hr) are shown separately for men (blue) and women (red) under the fully adjusted model (M2). Sample sizes were 2,393, 1,263, and 53 for the short-sleep, reference-sleep, and long-sleep groups, respectively. Among men, the corresponding sample sizes were 1,176, 749, and 34; among women, they were 1,217, 514, and 19. BMI, body mass index; FEV1, forced expiratory volume in 1 second; FVC, forced vital capacity; p_int = 2-df LR interaction p-value; CI, confidence interval; LR, likelihood-ratio.
Multiple-testing control
The four primary outcomes were tested without correction as a pre-specified primary tier. For the 20 secondary outcomes, the Benjamini–Hochberg false discovery rate (FDR) procedure [26] was applied within M1 and M2; secondary findings were reported as statistically noteworthy at q<0.10. The post hoc age-stratified analysis (sensitivity S7) involves 8 tests (2 strata×4 outcomes); a Bonferroni-adjusted threshold of α/8=0.00625 is applied and discussed explicitly.
Sensitivity analyses
Six pre-specified sensitivity analyses tested robustness: (S1) weekday-only sleep duration; (S2) widened plausible range (2–16 h); (S3) exclusion of prevalent hypertension or fasting diabetes; (S4) additional BMI adjustment (except when BMI was the outcome); (S5) mean/mode imputation of M2 covariates; (S6) finer 5-category sleep scheme (<6, 6 to <7, 7 to <8 [reference], 8 to <9, ≥9 hr). One exploratory age-stratified analysis (sensitivity S7, <50 vs. ≥50 years) was added post hoc after the 1-df fatigue-related symptom score contrast emerged as nominally significant; its post hoc status is disclosed throughout.
Ethics
The KoGES was approved by the Institutional Review Board of the Korea Centers for Disease Control and Prevention; all participants provided written informed consent at baseline and each follow-up. The present secondary analysis of de-identified KoGES data was reviewed and approved by the Institutional Review Board of Korea University Anam Hospital (IRB No. IEC107014) with a waiver of additional informed consent.
RESULTS
Sample characteristics
The analytic sample comprised 3,709 Ansan residents (men 1,959, 52.8%; women 1,750, 47.2%). Men and women differed on age, anthropometry, blood pressure, smoking, drinking, occupational profile, pulmonary function, and the fatigue-related symptom score (Table 2). Sleep duration distribution differed between sexes (p<0.001): women were more likely than men to be short sleepers. The long-sleeper stratum (≥9 hr) was uncommon in both sexes (n=53 total, 1.4%), yielding a correspondingly small cell count for long-end interaction inference.
Primary interaction results
Under the primary M2 adjustment with the three-category sleep scheme (Figure 3), none of the four primary outcomes showed a significant sleep×sex interaction on the 2-df LR test (Table 1): BMI, LR χ2=0.15, p=0.930; fasting glucose, LR χ2=1.89, p=0.388; fatigue-related symptom score, LR χ2=4.53, p=0.104; FEV1/FVC ratio, LR χ2=0.31, p=0.855. Results were materially unchanged under M1 (all interaction p>0.05). Sex-stratified β estimates (Figure 2) showed directionally concordant and quantitatively similar short-sleep and long-sleep contrasts in men and women for BMI, glucose, and FEV1/FVC, consistent with the null interaction.
Sex-stratified adjusted means with 95% CIs for the four primary outcomes across three sleep-duration categories. A: Body mass index. B: Fasting plasma glucose. C: Fatigue-related symptom score. D: FEV1/FVC ratio. Adjusted means were estimated using the fully adjusted model (M2). Sample sizes were 2,393, 1,263, and 53 for the short-sleep, reference-sleep, and long-sleep groups, respectively. Among men, the corresponding sample sizes were 1,176, 749, and 34; among women, they were 1,217, 514, and 19. BMI, body mass index; FEV1, forced expiratory volume in 1 second; FVC, forced vital capacity; ref, reference; CI, confidence interval; M2, model 2.
1-df short-vs.-reference contrast
Because the 2-df test distributes power across short and long contrasts and the long cell is sparse (n=53), we additionally report the 1-df short-vs.-reference sex-interaction contrast. The contrast was clearly null for BMI (LR χ2=0.01, p=0.919, n=3,600), fasting glucose (LR χ2=0.70, p=0.404, n=3,376), and FEV1/FVC (LR χ2=0.01, p=0.910, n=3,490). For fatigue-related symptom score the 1-df contrast was LR χ2=3.94, p=0.047 (n=3,649), nominally significant and directionally concordant with the subsequent age-stratified result (post hoc age stratification).
Minimum detectable effect
At observed stratum Ns, outcome SDs, α=0.05, and 80% power (Supplementary Table 2), the minimum sex-difference in the short-vs.-reference contrast that our design can detect is 0.57 kg/m2 for BMI (approximately 0.20 SD), 3.51 mg/dL for fasting glucose (approximately 0.21 SD), 3.55 points for fatigue-related symptom score (approximately 0.20 SD on the observed 30–120 range), and 1.23 percentage points for FEV1/FVC (approximately 0.20 SD). For the long-vs.-reference contrast, the minimum detectable sexdifference is approximately 0.82–0.84 SD for all outcomes, reflecting the sparse long-sleep tail. The study is adequately powered to rule out sex-differences >0.20 SD at the short end but cannot rule out clinically meaningful sex-differences at the long end.
Secondary outcomes
Across the 20 secondary outcomes (Supplementary Table 1), no interaction term survived Benjamini–Hochberg FDR control at q<0.10 under either M1 or M2. Nominal p-values flagged several candidate domains (WHOQOL-BREF items; select lipid measures) but these did not retain significance after correction and are therefore regarded as hypothesis-generating only.
Sensitivity analyses
The six pre-specified sensitivity scenarios (S1–S6; Supplementary Table 3) broadly supported the robustness of the primary null for BMI, glucose, and FEV1/FVC; interaction p-values remained >0.30 for all three outcomes under scenarios S1–S5 (weekday-only sleep, widened plausible range, excluding HTN/diabetes mellitus [DM] cases, +BMI covariate, and simple imputation). Two findings warrant explicit comment: 1) Fatigue interaction sharpens under sensitivity S3 (excluding HTN/DM). Excluding participants with prevalent hypertension or fasting diabetes strengthened the 2-df fatigue-related symptom score×sex interaction from p=0.104 to p=0.030, suggesting that chronic-disease burden may dampen—rather than drive—the sex-dependent fatigue-related symptom score signal. 2) BMI sensitivity to sleep binning (sensitivity S6). Under the finer 5-category scheme (reference 7 to <8 hr), the BMI×sex interaction reached p=0.049—a fragile finding that depends on binning choice and should not be over-interpreted.
Post-hoc age stratification (sensitivity S7)
Motivated by the 1-df fatigue-related symptom score signal (p=0.047) and by the well-documented age-dependence of both sleep architecture and HPA-axis reactivity [5,8], we conducted an explicitly post hoc age-stratified analysis (sensitivity S7). The <50-year stratum comprised 2,097 participants with complete fatigue-related symptom score and covariate data (women short= 658, reference=282, long=11; men short=690, reference=439, long=17); the ≥50-year stratum comprised n=1,603. This age-stratified analysis was not pre-specified and should be interpreted as exploratory and hypothesis-generating. Within adults <50 years, the fatigue-related symptom score×sex interaction was strong (2-df LR p=0.001), whereas in those ≥50 years it was essentially null (p=0.832). No equivalent age-conditional pattern emerged for BMI, glucose, or FEV1/FVC (all p>0.1 in both strata).
Direction and magnitude
Within the <50 stratum, short sleep (<7 hr) relative to the 7 to <9 hours reference was associated with a +4.75-point higher adjusted fatigue-related symptom score in women (95% CI +2.33 to +7.16, p<0.001) but essentially no change in men (β=-0.50, 95% CI -2.52 to +1.52, p=0.630). The women-vs.-men difference in the short-sleep contrast was +5.22 points (95% CI +2.11 to +8.34, p=0.001) on the sum-score’s observed 30–120 range, corresponding to approximately 0.29 SD of the outcome.
Multiplicity
With 8 post hoc tests (2 strata×4 outcomes), a Bonferroni-adjusted threshold of α/8=0.00625 applies. The <50 fatigue-related symptom score result (nominal p=0.001) survives this threshold.
Robustness
The <50 fatigue-related symptom score sex-interaction retains its direction and significance under the 5-category binning scheme (sensitivity S6) applied within the <50 stratum (4-df LR p=0.002; data not shown). The long-sleep cells in the <50 stratum are extremely sparse (n=11 women, n=17 men), so the <50 fatigue-related symptom score finding is driven almost entirely by the short-vs.-reference contrast, consistent with the 1-df result above.
DISCUSSION
In this a priori analysis of 3,709 adults in the Ansan site of KoGES Phase 2, the association between three-category sleep duration and each of four primary health outcomes—BMI, fasting glucose, a fatigue-related symptom score, and FEV1/FVC—did not differ significantly between men and women on the 2-df LR interaction test (all p≥0.104). Sex-stratified estimates were directionally concordant for BMI, fasting glucose, and FEV1/FVC, and no secondary outcome survived FDR correction. Three findings tempered this broad null. First, a 1-df short-vs.-reference sex-interaction contrast for fatigue-related symptom score was nominally significant (p=0.047), suggesting that the 2-df test was power-limited by the sparse long-sleep cell rather than by a true absence of signal at the short end. Second, the fatigue-related symptom score×sex interaction strengthened after excluding prevalent hypertension and diabetes cases (sensitivity S3, p=0.030). Third, in a post hoc age stratification, a fatigue-related symptom score× sex interaction concentrated in adults under 50 emerged (p=0.001, surviving the Bonferroni threshold for 8 post hoc tests), with short-sleeping women showing a +4.75-point higher adjusted fatigue-related symptom score (approximately 0.29 SD) and short-sleeping men showing essentially no change. Taken together, the data suggest that sex may modify the sleep–health relationship most visibly in the subjective neurobehavioral domain and primarily in younger adulthood, although this finding is explicitly exploratory and requires independent replication, rather than across cardiometabolic or pulmonary indices in middle age.
The cardiometabolic null is compatible with—but more specific than—prior sex-aggregated Korean estimates [11]. It is also consistent with Mendelian randomization analyses in UK Biobank, which found no evidence of a causal effect of sleep duration on BMI [29]. In our data the short-vs.-reference BMI contrast is +0.12 kg/m2 in women vs. +0.13 kg/m2 in men (women minus men, -0.02 kg/m2; 95% CI -0.42 to +0.38), well below our minimum detectable effect of 0.57 kg/m2. A similar pattern obtains for fasting glucose: the sex-difference in the short-sleep contrast in our data (approximately 0.9 mg/dL) is smaller than our minimum detectable effect of 3.5 mg/dL. In short, the null findings for these outcomes are plausibly bounded by statistical power at our sample size, and the minimum-detectable-effect calculation quantifies that boundary. Given the small size of the long-sleep stratum (n=53, 1.4%), we refrain from substantive interpretation of long-sleep contrasts; the study was not powered for this comparison (minimum detectable effect approximately 0.82–0.84 SD).
The following mechanistic discussion is exploratory and hypothesis-generating; none of the pathways described below were directly tested in the present data. Three interlocking mechanisms make the <50-year fatigue-related symptom score×sex interaction biologically plausible. First, reproductive-hormone effects on sleep architecture are age-graded: estradiol and progesterone fluctuations across the menstrual cycle produce changes in slow-wave sleep and REM fraction that affect next-day subjective fatigue; these effects attenuate after menopause [5,6]. A fatigue-related symptom score×short-sleep interaction present in women <50 and absent in women ≥50 is the pattern a reproductive-hormone hypothesis would predict. Second, sex differences in autonomic regulation during sleep are most pronounced in the reproductive years, and sleep restriction may produce a larger next-day fatigue burden in premenopausal women. Third, experimental studies have documented sex differences in the inflammatory response to sleep deprivation, which is a proximate driver of subjective fatigue. However, these mechanistic pathways were not directly tested in the present data, and the interpretation remains speculative.
We were unable to test the reproductive-hormone interpretation directly by stratifying women <50 by pre-vs.-postmenopausal status, but the menopause/menstrual items in KoGES Phase 2 are not present in the derived analytic file and require re-extraction from raw source. The menopausal-transition analysis is therefore a specific pre-registered component of the planned longitudinal companion paper rather than a result of the present study. We label the reproductive-hormone interpretation as speculative in the present paper, whereas the inflammatory and autonomic pathways are better corroborated in the mechanistic literature cited above.
Prior Korean population analyses of sleep duration have largely focused on cross-sectional sex-aggregated associations or on insomnia symptom prevalence [9-11]. A smaller line of KoGES-based work has examined stress–insomnia–quality-of-life pathways [13] and sleep–metabolic associations without formal interaction testing. The present paper is to our knowledge the first in this cohort to pre-specify sleep×sex interaction tests simultaneously across cardiometabolic, pulmonary, and neurobehavioral domains. The broad null for cardiometabolic and pulmonary outcomes suggests that, in middle-aged Korean adults, sex-specific sleep-duration counseling on those endpoints is not supported by the present data; the <50-year fatigue-related symptom score finding, if replicated, could motivate sex-specific screening for subjective fatigue in short-sleeping younger women.
Strengths include a priori specification of the primary hypothesis, adjustment sets, and sensitivity battery; use of a LR interaction test that is invariant to reference coding [14]; FDR control within the secondary tier; an explicit 1-df contrast and minimum-detectable-effect calculation that allow the reader to judge what the data can and cannot rule out; and transparent disclosure of the post hoc age stratification with Bonferroni-adjusted multiplicity control.
Limitations are substantial. First, the long-sleeper stratum was small (n=53), which reduces power for the long-vs.-reference contrast; Supplementary Table 2 makes this boundary explicit. Second, sleep duration was self-reported; systematic sex differences in reporting style cannot be excluded, and accelerometer-assisted Phase-3+ data will be needed to address this directly. Recent work from our group has demonstrated the feasibility of wearable-derived digital phenotyping for sleep disorder assessment, including machine learning approaches for obstructive sleep apnea screening [30] and insomnia classification using heart rate circadian phase and hyperarousal markers [31], suggesting that objective sleep measurement in future KoGES waves is technically achievable. Third and most critically, the cross-sectional design precludes causal inference. For the fatigue-related symptom score finding this is not a minor caveat: chronic fatigue may drive both sleep complaints and reported sleep duration, and the <50-year signal may therefore reflect reverse causation (fatigued younger women sleep shorter) rather than the forward causal direction (short sleep makes younger women more fatigued). A longitudinal follow-up examining KoGES Phase 2 to Phase 4 change in BMI, fasting glucose, and fatigue-related symptom score by baseline sleep×sex—with an a priori analysis plan to be registered on Open Science Framework before data lock—is therefore essential and is planned as the next manuscript. Fourth, the fatigue-related symptom score is not a validated fatigue instrument; inferences drawn from it should be interpreted accordingly and cross-cohort comparability is limited. Moreover, smartphone-based digital phenotyping has recently been applied to multidimensional insomnia assessment [32], highlighting both the potential and the current gap in objective fatigue measurement. Fifth, Cronbach’s α for the sum-score cannot be computed from the derived analytic file and must be reported in the longitudinal follow-up. Sixth, exclusion of the Ansung (rural) sample—a conservative internal-comparability choice—limits generalizability to rural Korean adults; the pooled Ansan+Ansung site-adjusted analysis is a planned component of the longitudinal follow-up. Seventh, HbA1c was measured only in a subset without fasting glucose data (n=227), and fasting diabetes was defined by FPG alone; HbA1c-based estimates and diabetes prevalence should therefore be interpreted with caution.
Two practical implications follow. Clinically, sleep-duration counseling for cardiometabolic risk in middle-aged Korean adults does not require routine sex stratification based on the present cross-sectional findings, although individualization remains appropriate; screening for subjective fatigue in short-sleeping younger women may be warranted pending replication. Methodologically, the contrast between the null 2-df test and the nominally significant 1-df contrast for fatigue-related symptom score illustrates a point of practical importance: a 2-df omnibus test that distributes power across a dense short-sleep cell and a sparse long-sleep cell can mask a true short-end signal, and investigators should pre-specify a 1-df complementary contrast where the outcome distribution is uneven. Future work should 1) replicate the <50-year fatigue finding using validated fatigue instruments (e.g., Fatigue Severity Scale) with pre-registered analysis plans; 2) extend the interaction test longitudinally using KoGES Phase 2 to Phase 4 follow-up data; 3) pool Ansan and Ansung sites with site as a covariate to gain power at the long-sleep tail and test urban/rural interaction; and 4) examine the menopausal transition within women directly to interrogate the reproductive-hormone mechanism suggested here.
Conclusion
In an a priori analysis of 3,709 adults in the KoGES Ansan Phase 2 cohort, we found no evidence that sex modified the cardiometabolic or pulmonary sleep–health associations on the primary 2-df test (BMI p=0.930; fasting glucose p=0.388; fatigue-related symptom score p=0.104; FEV1/FVC p=0.855). A complementary 1-df short-vs.-reference contrast detected a nominally significant fatigue-related symptom score×sex interaction (p=0.047). A post hoc age-stratified analysis concentrated the fatigue-related symptom score×sex signal in adults under 50 (p=0.001), with short-sleeping women showing a +4.75-point higher adjusted fatigue-related symptom score (95% CI +2.33 to +7.16) and short-sleeping men essentially no change (β=-0.50, 95% CI -2.52 to +1.52)—a women-vs.-men difference of +5.22 points, approximately 0.29 SD of the outcome. The finding should be regarded as hypothesis-generating pending replication in an independent sample using validated fatigue instruments and a longitudinal design.
Supplementary Materials
The Supplement is available with this article at https://doi.org/10.33069/cim.2026.0023.
Sex-stratified associations between sleep duration and secondary outcomes
Minimum detectable sex-difference (Δβ) by outcome for short-vs.-reference and long-vs.-ref contrasts (α=0.05, 80% power)
Sensitivity analyses (S1–S6) and post-hoc age stratification (S7): interaction p-values for four primary outcomes under M2
Notes
Heon-Jeong Lee, the Editor-in-Chief of Chronobiology in Medicine, is a co-author of this article. He was not involved in the peer review process or the editorial decision for this article, and the review was handled independently by another member of the editorial board. The other authors have no potential conflicts of interest to disclose.
Availability of Data and Material
KoGES data are available to qualified researchers upon request from the Korea Disease Control and Prevention Agency (https://nih.go.kr). Analysis code is available from the corresponding author on reasonable request.
Author Contributions
Conceptualization: Chul-Hyun Cho, Youngjae Yoo. Data curation: Chul-Hyun Cho, Heon-Jeong Lee. Formal analysis: Chul- Hyun Cho. Funding acquisition: Chul-Hyun Cho. Methodology: Chul-Hyun Cho, Youngjae Yoo. Supervision: Heon-Jeong Lee. Writing—original draft: Chul-Hyun Cho. Writing—review & editing: Chul-Hyun Cho.
Funding Statement
This study was supported by the National Research Foundation (NRF) of Korea (grant number: NRF-2021R1A5A8032895, NRF-2022M3C1B6080866, RS-2026-25523612, and RS-2026-25471696). The funders had no role in study design, data collection, analysis, interpretation, manuscript preparation, or the decision to submit the article for publication.
Acknowledgments
Data for this study were provided by the Korean Genome and Epidemiology Study (KoGES; 4845-301), Division of Population Health Research, Korea National Institute of Health, Korea Disease Control and Prevention Agency, Republic of Korea. The authors thank the KoGES participants, investigators, and staff.
