The finding followed the Strengthening the Reporting of Observational Studies in Epidemiology guideline [15] (STROBE-NUT).

The NHANES is a large periodic study that investigated the health and nutritional data from the US population. Details of the NHANES repository have been described elsewhere [16]. The study populations and all data were freely obtained from NHANES. The research ethics review board of the National Centre for Health Statistics Research approved the NHANES study, and all involved participants provided informed consent.

Participants who had one valid dietary questionnaire at baseline were included (n = 109,653). We excluded 33,018 individuals whose dietary recalls were unreliable and did not meet the minimum criteria for NHANES, 32,134 individuals with aged <20 years, 596 individuals who reported an energy intake of >5000 kcal/d or <500 kcal/d, 243 individuals who refused to answer and/or had missing mortality events, 784 individuals who lacked complete information on dietary intake, 1134 individuals who were pregnant at baseline. Overall, our sample consisted of 41,744 participants (Supplementary Fig. 1).

Baseline dietary intake from 2002 to 2018 was obtained from the first 24-h dietary recall interview. The first recall was executed through personal interaction by proficient personnel at the NHANES mobile examination facilities. Standardized protocols and measuring tools were utilized to facilitate the evaluation of food volume and dimensions. During the interview, the participants were requested to provide information regarding the consumption quantity and time of each food and beverage item. The nutrient values were determined by utilizing the Food and Nutrient Database for Dietary Studies (FNDDS). The dietary intake of NHANES participants was integrated into 37 major groups of MyPyramid, as per the USDA’s Food Patterns Equivalents Database 2017–2018 (FPED, 2017–2018). The dietary supplements were obtained through the administration of a dietary supplement questionnaire. To assess diet quality, we calculated the Healthy Eating Index-2015 (HEI-2015), which is a measure for evaluating the alignment of foods with the American Dietary Guidelines for 2015–2020 Dietary Guidelines for Americans [17]. Dietary intake (without dietary supplement) was adjusted for total energy intake using the residual method.

The main exposures were timing, frequency, and food quality of night eating. The night eating was defined as food consumption between 22:00 to 4:00 based on natural light cycle rhythm in this study. Timing of night eating was categorized into seven groups: “no eating“, “22:00 to 23:00“, “23:00 to 00:00“, “00:00 to 1:00“, “1:00 to 2:00“, “2:00 to 3:00“, and “3:00 to 4:00“; where if an individual had night eating more than twice a day, choose the timing of night eating based on the latest time point. Frequency of night eating was categorized into three groups: “no eating”, “one time”, and “two times and over”.

Food parameters of energy intake, low energy density foods intake (fruits, vegetables, whole grains, dairies, and protein foods), and high energy density foods intake (refined grains, add sugars, oils, solid fats, and alcoholic drinking) were used to generate a food quality for night eating. Each food parameter was preliminarily categorized by two levels and assigned into category 1 or 2 levels, respectively; where energy intake <200 kcal was assigned into 1, energy intake ≥200 kcal was assigned into 2; where food intakes <the median values of all foods intakes was assigned into 1, where food intakes ≥the median values of all foods intakes was assigned into 2. Then, based on above categories, food quality of night eating was defined using latent class analysis. A reasonable model was selected by analyzing latent classes with different numbers of latent classes. Akaike information criterion (AIC) and bayesian information criterion (BIC) were computed for the model selection (Supplementary Fig. 2) and four classes was selected (latent class 1, 2, 3, and 4). The item-response probabilities in models from four latent classes were shown in Supplementary Table 1.

Further, the characteristics of the four latent classes were assessed according to food intake in different latent class (Supplementary Table 2). Latent class 1 was characterized by very low energy intake, low intakes from low energy density foods, and very low intakes from high energy density foods, which could be labeled “very low dietary-energy-density intake” (VL-energy intake); latent class 2 was characterized by low energy intake, very low intakes from low energy density foods, and moderate intakes from high energy density foods, which could be labeled “low dietary-energy-density intake” (L-energy intake); latent class 3 was characterized by moderate energy intake, moderate intakes from low energy density foods, and moderate intakes from high energy density foods, which could be labeled “moderate dietary-energy-density intake” (M-energy intake); latent class 4 was characterized by high energy intake, high intakes from low energy density foods, and high intakes from high energy density foods, which could be labeled “high dietary-energy-density intake” (H-energy intake).

The outcomes were all-cause, cancer, and diabetes mortality that transpired subsequent to the survey participation date and prior to December 31, 2019. The National Death Index (NDI) was utilized to obtain death information. The NDI was publicly distributed by centers for disease control and prevention, Public-use Linked Mortality Files from National Center for Health Statistics, which are available for NHANES (National Center for Health Statistics. Office of Analysis and Epidemiology, Public-use Linked Mortality File, 2015. Hyattsville, Maryland. https://www.cdc.gov/nchs/data-linkage/mortality-public.htm). The International Classification of Diseases 10th Revision (ICD-10) was adopted to classify cause specific mortality; cancer mortality was identified by the ICD-10 codes C00-C9, and diabetes mortality was identified by the ICD-10 codes E10–E14. In total, 6066 deaths were recorded; of them, 1381 deaths were due to cancer and 206 deaths were due to diabetes.

Covariates included age (years), sex (male/female), race/ethnicity (Mexican American/non-Hispanic Black/non-Hispanic White/other Hispanic/other), education (less than 9th grade/9–11th grade/college graduate or above/high school graduate or GED or equivalent/some college or AA degree), income ($0–$19,999/$20,000–$44,999/$45,000–$74,999/$75,000–$99,999/ ≥ $100,000), smoking status (never smoker/past smoker/current smoker), drinking status (never drinker/past drinker/current drinker), body mass index (kg/m2), physical activity (metabolic equivalent hours per week (METs-h/week), sleep hours (hours/day), dietary energy intake (kcal), adherence to HEI-2015, dietary supplement use (%), glycohemoglobin (%), triglycerides (mmol/L), fasting glucose (mmol/L), total cholesterol (TCHO, mg/dL), oral glucose tolerance test (OGTT, mg/dL), hypertension, hyperlipidemia, cardiovascular disease (CVD), diabetes, and cancer. Drinking status was defined as never drinker (drank <12 drinks lifetime), past drinker (drank ≥12 drink lifetime and nondrinker over the past 12 months, and current drinker (drank ≥12 drink and currently a drinker). Smoking status was defined as never smoker (smoked <100 cigarettes lifetime), past smoker (smoked >100 cigarettes lifetime and currently did not smoke), and current smoker (smoked >100 cigarettes and currently a smoker). Timing of blood draw for biochemical variables detection was in the morning recorded by NHANES. Diabetes was defined as self-reported, diagnosed diabetes, hemoglobin A1c (HbA1c) ≥ 6.5%, or fasting plasma glucose ≥7.0 mmol/L. Hypertension was defined as diagnosed hypertension, taking antihypertensive drugs, systolic blood pressure ≥140 mm Hg, or diastolic blood pressure ≥90 mm Hg. Hyperlipidemia was defined as taking antihyperlipidemic drugs, triglycerides ≥150 mg/dL, total cholesterol ≥200 mg/dL, low density lipoprotein cholesterol ≥130 mg/dL, or high-density lipoprotein cholesterol <40 mg/dL for male and <50 mg/dL for female [18].

Analyses were performed according to NHANES analytic guidelines, including sample weights, stratification, and clustering. Data analyses were conducted by R version 4.2.3 (the R Core Team). A two-sided P < 0.05 indicated statistical significance. The baseline characteristics were expressed as the means ± SD or numbers (percentages).

Weighted Cox proportional hazards (CPH) regression models were applied to evaluate the associations of the timing, frequency and food quality of night eating with all-cause, cancer, and diabetes mortality (no night eating as a reference). Results were expressed as adjusted hazard ratios [aHR] with 95% confidence intervals [CI]. We adjusted for baseline age and sex in model 1. We further adjusted for baseline education, race/ethnicity, family income, and body mass index in mode 2. Finally, we additionally adjusted for baseline dietary energy intake, alcohol consumption per day, smoking status, physical activity, histories of diabetes, hypertension, CVD, cancer, hyperlipidemia, adherence to the HEI-2015 score, and dietary supplement use in model 3. Percentage of missing values from covariates was less than 10% except for sleep hours (20.2%) (Supplementary Table 3). Chained equations (multivariate imputation) were used to impute missing values.

Subgroup analyses were performed in CPH models, categorized by baseline age (<65 and ≥65 years), sex, body mass index (>25, 25–29, and ≥30 kg/m2), smoking status (never smoker, past smoker, and current smoker), drinking status (never drinker, past drinker, and current drinker), HEI-2015 score (<70 and ≥70), sleep hours (<6 and ≥6 hours).

Linear regression analysis was used to investigate the correlation among timing, frequency and quality of night eating and biochemical variables (HbA1c, triglycerides, total cholesterol, OGTT, fasting glucose). Results were expressed as β value with 95% CIs. Three models (models 1, 2, and 3) were adjusted as described above except that model 3 was further adjusted for total length of fasting time. The fasting time is defined as the time (in hours) from not eating or drinking (except water) to venipuncture [19].

Sensitivity analyses were performed to assay the robustness of the results. Firstly, we excluded 810 participants who had over 50% energy intake from night to reduce the influence of night eating syndrome on the results. Secondly, we added eating time (9:00–10:00) and re-ran CPH analyses. Moreover, we repeated the CPH analysis after further adjusting for sleep times.