The data sets used in this research were retrieved from the 2007–2018 National Health and Nutrition Examination Survey (NHANES). NHANES is an ongoing, biennial, nationally representative series of surveys, which adopt a complex, multistage, probability sampling design to monitor the health and nutritional status of adults and children in the United States. Detailed information on the NHANES could be assessed at https://www.cdc.gov/nchs/nhanes/index.htm.

The protocols for NHANES were approved by the National Center for Health Statistics (NCHS) Research Ethics Review Board, and informed consent was obtained from all participants, available online at https://cdc.gov/nchs/nhanes/irba98.htm. According to 45 CFR Part 46, ethical approval and informed consent were not required for the current study as the data sets were all publicly available from NHANES.

Participants included in this study need to satisfy the following inclusion criteria: adults (aged ≥18 years old), with complete information on their sleep duration and two 24-h dietary carotenoid intakes. Those with missing data on any covariates including demographic, behavioral, and health characteristics would be excluded.

Sleep habits and disorders related questions were asked, in the home, by trained interviewers using the Computer-Assisted Personal Interviewing (CAPI) system, where consistency checks were built in to reduce data entry errors for quality assurance and control. Sleep duration data reflected self-reported usual sleep were asked “How much sleep do you usually get at night on weekdays or workdays?”. We categorized the participants into three groups: short sleep duration (< 7 h/night), optimal sleep duration (7–8 h/night), and long sleep duration (> 8 h/night) according to the previous studies [4,5,6], while the answers of “Do not know” and “Refused” were considered missing and omitted.

Dietary intake information from NHANES participants was obtained from the two 24-h dietary interviews, which were conducted by trained dietary interviewers. The first dietary recall interview was collected in-person in the Mobile Examination Center (MEC) and the second was collected via telephone approximately 3 to 10 days after the first interview. Dietary carotenoid intakes used in this study including α-carotene (mcg/day), β-carotene (mcg/day), β-cryptoxanthin (mcg/day), lycopene (mcg/day), and lutein + zeaxanthin (mcg/day) were retrieved from the two 24-hour dietary recall interviews, and divided into three categories based on the quartiles of average amount from the two recalls. The cut-off values for each can be found in Supplementary Table S1.

The covariates of three dimensions, including sociodemographic, behavioral, and health characteristics, were regarded as potential confounding factors a priori.

Sociodemographic characteristics comprised age groups (18–39, 40–59, and ≥ 60 years old), sex (Female and Male), race (Non-Hispanic White, Mexican American, Non-Hispanic Black, and Other/multiracial), highest education degree (Less than high school graduate, High school graduate or GED, and Some college or above), and family income level (0–130, 130–350%, and > 350% PIR, PIR refers to the ratio of family income to poverty threshold).

Behavioral variables consisted of smoking status (Never, former, and current), drinking (Yes or No), physical activity [Inactive (< 600 metabolic equivalents of task (MET) per week) and active (≥ 600 MET/week)], and the amount of caffeine consumption per day [<Q1 (< 41.5 mg/day), Q1-Q3 (41.5–240 mg/day), and > Q3 (> 240 mg/day)].

Health characteristics included body mass index (BMI) categories [underweight/normal (≤ 24.9 kg/m2), overweight (25.0–29.9 kg/m2), and obese (≥ 30.0 kg/m2)], hypertension (Yes or No), diabetes (Yes or No), and depression (Yes or No).

As suggested by the analytic guidelines of NHANES, primary sampling units (SDMVPSU), stratification (SDMVSTRA), and sampling weight (WTMEC2YR, full sample 2-year MEC exam weight) were incorporated in all analyses to generate nationally representative estimates.

Dietary carotenoid intake and potential confounding factors were summarized according to the sleep duration groups and described as the frequency with weighted percentages. Survey design-based χ2 tests were used to examine the associations between these variables and sleep duration phenotypes.

Since the participants were categorized into optimal, short, and long sleep duration groups, multinominal logistic regression was constructed to calculate the odds ratio (OR) and 95% confidence interval (CI) of dietary carotenoid intakes with risks among different sleep duration groups. With the people with optimal sleep duration as the reference group, the multinominal logistic regression basically worked in the same way as binary logistic regression, where the analysis broke down the sleep duration groups into two comparisons: short sleep duration vs. optimal sleep duration, and long sleep duration vs. optimal sleep duration. Additionally, to assess the confounding effects from the aforementioned three different dimensional covariates, these covariates were gradually adjusted: Model I was adjusted for sociodemographic characteristics, Model II was further adjusted for behavioral variables, and health factors were additionally added in Model III. Trend tests (p for trend) were performed by entering the dietary carotenoid intake (quartile-categorical) as a continuous variable and rerunning the corresponding regression models.

Furthermore, the restricted cubic spline (RCS) models were utilized to examine the dose-response relationships between dietary carotenoid intakes and sleep duration, with three knots located at the 5th, 50th, and 95th percentiles of the distributions [25, 26]. For more visual purposes, we illustrated the levels of dietary carotenoid intakes in participants with different sleep durations, with sleep duration (continuous, h/night) on the X-axis [5].

Additionally, the weighted quantile sum (WQS) regression model was used to estimate the overall mixed effects associated with five carotenoid subclasses and identify the predominant carotenoid types. Individual weight for each carotenoid was estimated using bootstrap sampling (n = 100), where the data were randomly split into the training set (80%) and the validation set (20%). Detailed information on the WQS regression model could be obtained from the previous literature [27].

Statistical analyses were performed in the R software 4.2.3 (R Foundation for Statistical Computing) and Stata/MP 17.0 (StataCorp, Texas, USA). All statistical tests were two-sided, and α = 0.05 was considered as the significance level.