Nonlinear relationship between sleep midpoint and depression symptoms: a cross-sectional study of US adults

The participants in this study were obtained from the NHANES, a major program conducted by the Centers for Disease Control and Prevention (CDC) to assess the health and nutritional status of adults and children in the US. NHANES is a cross-sectional survey, conducted by the National Center for Health Statistics (NCHS), that includes a nationally representative sample of noninstitutionalized, resident US civilians of all ages. The NHANES comprises demographic, socioeconomic, health-related, and medical information. The National Center for Health Statistics Research Ethics Review Board authorized the NHANES study protocols. Since 1999, the NHANES has been conducted every two years and has used a stratified, multistage probability sampling approach. Further information regarding the NHANES is available on the CDC website [22]. The data of this study were collected during the NHANES survey cycle from 2015 to 2020.

Participants in our study were screened according to the following exclusion criteria: 1) Younger than 18 years old; 2) incomplete Patient Health Questionnaire-9 (PHQ-9); and 3) missing information on sleep midpoint.

The sleep midpoint was calculated based on self-reported waking and initial sleep times; the midpoint between going to sleep and the time of awakening was defined as the sleep midpoint. Information on waking and initial sleep times was acquired based on answers to the following questions: “What time do you usually go to sleep on weekdays or workdays?”; “What time do you usually wake up on weekdays or workdays?” This access method is very commonly used when studying the sleep midpoint [19, 20, 23].

The PHQ-9, which is a questionnaire based on self-reporting by patients, was used to assess depression symptoms. It contains nine items (depressed mood, appetite problems, fatigue, sleep difficulties, psychomotor retardation or agitation, concentration problems, lack of interest, feelings of worthlessness, and suicidal ideation) that assess the frequency of depressive symptoms and is well accepted as an accurate and reliable method for depression symptoms screening [24‐27]. The total PHQ-9 score varies from 0 to 27; a score of 10 or greater was defined as indicative of depression symptoms.

Adjustment covariates were chosen based on previous literature [28‐32]. Age, body mass index (BMI), and sleep duration were continuous variables. BMI was measured as weight (kg) divided by height (m) squared. Sleep duration was determined by answers to the question, “How much sleep do you usually get at night on weekdays or workdays?” Sex (male/female), race/ethnicity (non-Hispanic white, Mexican American, non-Hispanic black, other Hispanic or other race/multiple races), education level (< high school/completed high school/ > high school), marital status (married/living with a partner or never married/widowed /divorced/separated), physical activity (inactive/moderate/vigorous/both moderate and vigorous), family income, smoking status, drinking status, and comorbid condition were used as categorical variables. Family income was categorized into the following 3 levels based on the family poverty income ratio: low income (≤ 1.3), medium income (> 1.3 to 3.5), and high income (> 3.5). Drinking status was categorized as never (had < 12 drinks in a lifetime), former (had ≥ 12 drinks in 1 year and did not drink last year, or did not drink last year but drank ≥ 12 drinks in a lifetime), current light/moderate drinker (≤ 1 drink per day for females or ≤ 2 drinks per day for males on average over the past year), or current heavier drinker (> 1 drink per day for females or > 2 drinks per day for males on average over the past year). Smoking status was categorized into “never smoker,” “former smoker,” or “current smoker.” Participants were defined as a "never smoker" if they answered "No" to "Ever smoked a cigarette even one time" or "Smoked at least 100 cigarettes in life". If participants answered "Yes" to "Ever smoked a cigarette even one time," then they were excluded from the category of "never smoker." Participants were defined as a "former smoker" if they were not currently smoking but answered "Yes" to "Ever smoked a cigarette even one time" or "Smoked at least 100 cigarettes in life." Participants were defined as a "current smoker" if they answered "Every day" or "Some days" to "Do you now smoke cigarettes?". Physical activity was assessed as vigorous physical activity (high-intensity activities, fitness, and sports such as running or basketball) and moderate physical activity (e.g., brisk walking, swimming, bicycling at a regular pace), reported by participants. Participants were defined as having the comorbid condition if they reported at least one of the following medical conditions: diabetes, kidney failure, kidney stones, heart failure, stroke, hepatopathy, rheumatoid arthritis, and cancer. Because of the non-orthogonal relationship between sleep midpoint and depression symptoms, we adjusted the sleep item of the PHQ-9 questionnaire as a covariate. In this item, participants reported the frequency of sleep disturbance (Not at all/Several days/More than half the days/Nearly every day).

All analyses accounted for the NHANES complex sampling design to derive nationally representative estimates [33]. The characteristics of participants are described as means (95% CIs) for continuous variables and percentage frequencies (95% CIs) for categorical variables. Continuous data were compared using analysis of variance, and the χ² test compared categorical data. The nonlinear relationship between sleep midpoint and depression symptoms was tested using restricted cubic splines (RCS) (Fig. 1). A P value for nonlinear < 0.05 indicates a nonlinear relationship [34‐37]. Multivariable logistic regression analysis was performed to evaluate the associations between sleep midpoint and depression symptoms. We used three levels of adjustment. In Model 1, age, sex, and race/ethnicity were chosen as covariates because they are basic demographic information and have been shown to be associated with sleep behaviors and depression symptoms. Model 2 was adjusted for the variables in Model 1 plus education level, marital status, physical activity, and family income, because these are sociological factors that are commonly adjusted for when exploring depression symptoms and sleep behaviors. These variables can influence an individual's lifestyle, social support, and access to healthcare resources, which could affect their sleep quality and depressive conditions. In addition to variables in Models 1 and 2, Model 3 was adjusted for BMI, smoking status, drinking status, physical activity, comorbid condition, sleep duration, and sleep disturbance. These variables were included because lifestyle habits and health status are known to impact both sleep quality and depression symptoms. Covariates that showed a nonlinear relationship with the outcome in the univariate analysis were adjusted as RCSs. Imputation of missing data was conducted using the missForest R package, which is a random forest-based technique that is highly computationally efficient for high-dimensional data consisting of both categorical and continuous predictors [38]. Missing covariates are listed in Table S1.

We further examined the threshold effect of sleep midpoint on depression symptoms. A likelihood ratio test comparing the one-line model with a piecewise regression model was used to determine whether a threshold exists. The threshold level of the sleep midpoint was determined using a recurrence method, including selecting the turning point along a predefined interval and choosing the turning point that yielded the maximum likelihood model. To further assess the robustness of the association between sleep midpoint and depression symptoms, we conducted stratified RCS.

A total of 18,280 participants were included in this study. A flow chart of the included/excluded participants is shown in Figure S1. Table 1 shows the characteristics of the participants (Q1: > 6:30 PM to ≤ 1:30 AM the next day; Q2: > 1:30 AM to ≤ 2:30 AM; Q3: > 2:30 AM to < 3:25 AM; Q4: ≥ 3:25 AM to ≤ 6:30 AM). Participants in Group 4 had the highest prevalence of depressive symptoms and more frequent sleep disturbances. Compared to participants in other groups, those in Group 4 had the lowest average age and BMI, and the longest sleep duration. Additionally, Group 4 had the highest proportion of participants who were physically inactive and had low income. Furthermore, the lowest proportion of participants who were married or living with a partner among all participants was in Group 4 (all P < 0.05).

A comparison of characteristics between participants included and excluded (excluded for the absence of sleep midpoint data or incomplete PHQ-9) from the analysis is shown in Table S2. In addition, we used univariate analysis to observe the association between covariates and depression symptoms (Table S3).

As shown in Fig. 1, RCSs suggested that the relationship between sleep midpoint and depression symptoms was nonlinear. Table 2 shows the results of the multivariable logistic regression analysis. Sleep midpoint was converted into a categorical variable; after adjusting for all covariates, compared with participants in Group 2, participants in Group 4 (odds ratio [OR] = 1.25, 95% confidence interval [CI]: 1.02–1.53), had an increased prevalence of depression symptoms; participants in Group 1 (OR = 0.93, 95%CI: 0.71–1.22) and Group 3 (OR = 0.83, 95%CI: 0.54–1.28) had no significant difference in the prevalence of depression symptoms.

According to the threshold effect analysis, after adjusting for all covariates, in participants with a sleep midpoint from 2:18 AM to 6:30 AM, the prevalence of depression symptoms increased by 0.2 times (unadjusted OR = 1.55, 95% CI: 1.43–1.67; adjusted OR = 1.20, 95% CI: 1.08–1.33) per 1-h increment in sleep midpoint compared to the reference point of 2:18 AM. For participants with a sleep midpoint after 6:30 AM and before 2:18 AM the next day, the relationship between sleep midpoint and depression symptoms was not significant after adjusting for all covariates (unadjusted OR = 0.94, 95% CI: 0.92–0.95; adjusted OR = 1.01, 95% CI: 0.99–1.03) (Table 3). Using imputation and setting dummy variables as methods for handling missing covariates gave consistent results in multivariable logistic regressions and threshold effect analyses (Tables 2, 3, and S4).

Figure 2 illustrates the relationship between sleep midpoint and depression symptoms in different subgroups. In this study, depression symptoms were assessed using the PHQ-9. To examine the robustness of the results, we also analyzed the PHQ-9 scores as a continuous variable. We plotted the RCS of the sleep midpoint and the PHQ-9 score (with the sleep subitem removed) in sensitivity analyses (Fig. S2). RCS and threshold effect analysis showed that the relationship between sleep midpoint and PHQ-9 score was still nonlinear (Table S5).