Factors associated with habitual sleep duration in US adults with hypertension: a cross-sectional study of the 2015–2018 National Health and Nutrition Examination Survey

This cross-sectional study used data collected from US adults aged 18 and above in the 2015–2016 and 2017–2018 waves of the National Health and Nutrition Examination Survey (NHANES). The NHANES is a cross-sectional survey whose aim is to estimate the health and nutritional status of civilian, noninstitutionalized US adults and children of all ages. The survey uses complex stratified multistage sampling to get a nationally representative sample. The data is collected through interviews done in a study participant’s home and physical examinations completed at a designated mobile examination center (MEC) [17].

The 2015–2016 and 2017–2018 NHANES cycles had 19,225 study participants [18]. Of these, there were 11,848 adults aged 18 and older. The primary inclusion criteria for the present study was the presence of hypertension in an adult aged 18 years and above. In the present study, average systolic BP and diastolic BP were calculated from all the recorded BP measurements taken at the MEC. The NHANES participants had three consecutive BP readings taken after 5 min of rest [19]. Based on the 2017 ACC/AHA guidelines [5] and past studies [20, 21], hypertension is defined as the current use of antihypertensive medication in those with a history of hypertension (n = 3215), or an average systolic BP ≥130 mmHg (n = 3744) or diastolic BP ≥80 mmHg (n = 2323). The number of adults who met the criteria of hypertension was 5851.

Out of the remaining 5851, 189 participants were excluded for the following reasons: 7 were pregnant (self-reported pregnant status or a positive urine pregnancy test), 50 were missing data on sleep duration, and 139 only participated in household interviews. Thus, the final study sample of adults with hypertension after exclusions was 5660.

The self-reported habitual sleep duration assessment was based on the amount of sleep usually obtained during weekdays or workdays during the main sleep period (day or night). In the 2015–2018 NHANES cycles, the sleep duration was calculated from responses to two survey questions, “What time do you usually fall asleep on weekdays or workdays?” and “What time do you usually wake up on weekdays or workdays?”. In the present study, we categorized habitual sleep duration into short (< 7 h), adequate (7–9 h), and long (> 9 h) based on guidelines by AASM/SRS and NSF [1, 2] and past studies [14, 15].

Previous studies [9, 13, 14, 16] and Grandner’s Socio-ecological Model of Sleep and Health [22] were used as a guide in identifying covariates to include in the study. Grandner’s model outlines factors at the “individual, social and societal levels” related to sleep [22]. The current study focused on factors at the individual (e.g., health behaviors and other health-related characteristics) and social level (e.g., socioeconomic status, employment status, and race/ethnicity) of Grandner’s model to identify potential determinants of habitual sleep duration, all measured at the individual level of analysis [22].

The “individual-level” covariates included in this study were gender (male, female), age, body mass index, help-seeking for sleeping difficulty, depressive symptoms, chronic health conditions, cigarette smoking, alcohol intake, and physical activity. Age was categorized as 18–44 years (young adults), 45–64 years (middle-aged adults), and ≥ 65 years (older adults). Body mass index (BMI) was categorized as < 25 (no overweight or obesity), 25 to < 30 (overweight), 30 to < 35 (obesity class I), 35 to < 40 (obesity class II), and ≥ 40 kg/m² (obesity class III), based on the World Health Organization (WHO) guidelines [23]. Help-seeking for sleeping difficulty was categorized as yes or no based on the response to the question “Have you ever told a doctor or other health professional that you have trouble sleeping?” Depressive symptoms were screened using the 9-item Patient Health Questionnaire (PHQ-9, possible range of 0–27), categorized based on the established criteria of minimal or none (0–4), mild (5–9), moderate (10–14), and moderately severe to severe (≥15) depressive symptoms [24].

Chronic health conditions were assessed based on either self-reported history (being informed by a doctor or other health professional of having the specific health condition) or laboratory test findings done at the mobile examination centers. Stroke, heart disease (coronary heart disease, angina, heart attack, or congestive heart failure), arthritis (including gout), and chronic obstructive pulmonary disease (including chronic bronchitis and emphysema) were self-reported. Those who answered in the affirmative to both “Has a doctor or health professional ever told you that you have asthma?” and “Do you still have asthma?” were classified as having asthma. Diabetes was defined as either a self-reported history of diabetes or blood glycohemoglobin ≥6.5%. Chronic kidney disease (CKD) was defined as self-reported history of undergoing dialysis in the previous 12 months or an estimated glomerular filtration rate (eGFR) less than 60 mL/min/1.73 m² computed using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation [25, 26].

Smoking status was categorized as never smoker (never smoked at least 100 cigarettes in their lifetime), former smoker (smoked at least 100 cigarettes in their lifetime but not smoking currently), and current smoker (smokes cigarettes some days or every day). Alcohol intake was based on the self-reported average number of alcoholic drinks in a lifetime and the past year. The equivalent for one alcohol drink used in NHANES was a 12-oz beer, a 5-oz glass of wine, or one-and-half ounces of liquor. Based on the 2015–2020 US dietary guidelines, alcohol intake was categorized as none (never had at least 12 alcoholic drinks in a lifetime or any alcohol in the past year), moderate (one drink for women and not more than two drinks for men in a day), and heavy (more than one or two drinks in a day for women and men, respectively) [27].

Physical activity levels were assessed by calculating the metabolic equivalent of task (MET) minutes per week from the self-reported minutes of moderate-to-vigorous intensity physical activity related to leisure-time, work, and transportation (commuting by walking or bicycling). MET scores are used to estimate an activity’s intensity, and one MET is defined as the energy expenditure at rest [28]. The MET scores proposed by NHANES for moderate-intensity, vigorous-intensity, and transportation-related physical activity are four, eight, and four, respectively [29]. MET-minutes per week were computed by multiplying the minutes per week of each of the three categories of physical activity (moderate-intensity, vigorous-intensity, and transportation-related physical activity) with their respective MET scores. The MET minutes for the three activities were then summed to compute the total MET minutes per week. The US guidelines on physical activity recommend that adults get at least 75 to 150 min of vigorous-intensity or 150 to 300 min of moderate-intensity physical activity weekly [30]. The physical activity level was categorized as none, low, sufficient, and high to correspond to zero, < 600, 600–1200, and > 1200 MET-minutes a week, respectively [28].

The “social-level” covariates were all measured at the individual level of analysis and included race/ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic, non-Hispanic Asian, Other), nativity status (US-born, not US-born), marital status (married or living with a partner, unmarried [i.e., never married, separated, divorced, or widowed]), education level (less than high school, high school graduate, some college, college graduate), health insurance (insured, uninsured), employment status, and family income to poverty ratio. Employment status was assessed based on responses to the following survey items: (1) the type of work done the previous week; (2) the number of hours worked in all jobs or businesses the previous week; (3) whether one usually worked 35 h per week; and (4) the main reason one did not work the previous week. The categories of employment status were: works < 35 h/week, works 35–44 h/week, works ≥45 h/week, not working because they are retired, not working due to health reasons, and not working due to other reasons [for example, taking care of family, attending school, or laid off from work]). Family income to poverty ratio was categorized as < 1.00, 1.00–1.99, 2.00–3.99, and ≥ 4.00.

In the final analytic sample (n = 5660), 20.1% of the observations had missing data in one or more covariates. The degree of missingness (unweighted number and weighted percentages) was as follows: marital status (n = 39, 0.4%), education level (n = 9, 0.1%), health insurance status (n = 10, 0.2%), cigarette smoking status (n = 6, 0.1%), alcohol intake (n = 403, 5.7%), BMI (n = 90, 1.3%), depressive symptoms (n = 357, 5.2%), and income to poverty ratio (n = 689, 9.8%). Imputation of the missing data was completed using multiple imputation by chained equations (MICE) [31].

All variables in the study (covariates and outcome variable) and NHANES design variables (strata, cluster, and survey weights) were included in the imputation model to avoid generating biased estimates in the analysis model [30, 32]. The literature on multiple imputation also recommends having in the imputation model other variables associated with missingness of data to make multiple imputations under the assumptions of missing at random assumptions more plausible [30, 33]. The auxiliary variables included in the imputation model included the reference person’s age and education level, household size, number of rooms in the household, and homeownership. Stata’s mi impute chained command was used to implement the multiple imputation. A total of 40 imputed datasets were generated, and Stata’s mi estimate command was used to analyze and pool the results [34].

We did a sensitivity analysis to check for differences between the observed (pre-imputation) data and the complete imputed data (post-imputation) by comparing the subgroup proportions across the variables with missing data. There was minimal or no change in the distribution of pre-imputed and post-imputed data across marital status, education level, health insurance status, BMI, smoking status, depressive symptoms, and alcohol intake. Changes were noted in the income to poverty ratio variable. The proportion of those in the ≥4.00 category decreased from 36.6 to 35.6% after imputation. The proportion of adults in the < 1.00 group increased from 12.5 to 13.1%, and the 1.00–1.99 group increased from 21.1 to 21.7% (Additional file 1: Supplemental Table 1). The results presented in this report are based on analyses of the complete imputed data.

Survey analysis procedures incorporating the NHANES sample’s MEC examination weights and strata and cluster variables were used to account for NHANES’s sampling procedures and non-response to obtain results representative of the sampled population [17]. Descriptive analysis was conducted on all covariates for the total study sample. In the bivariate analyses, the distribution of all covariates across the categories of habitual sleep duration was assessed using the χ² test.

Since the outcome variable (habitual sleep duration) categories are ordered, we examined the ordinal logistic regression model for model fit and violation of statistical assumptions. The approximate likelihood-ratio test of proportionality of odds across response categories and parallel regression assumption test statistics were significant indicating a violation of proportionality of odds assumption. Consequently, a multinomial logistic regression model was used in the analyses. We fit the model using the complete imputed data with habitual sleep duration as the outcome of interest. The multinomial logistic regression models (unadjusted and adjusted) were fitted with all covariates to examine their associations with short (< 7 h) and long sleep duration (> 9 h) using adequate sleep duration (7–9 h) as the reference category. The multinomial logistic regression model results were compared with complete case analysis results. All data were analyzed using STATA 1C Version 15 [35]. The level of significance for all analyses was set at a p-value < 0.05.

The average age of the participants was 56.1 years; 52.1% were male; 64.1% were non-Hispanic White, 13.3% non-Hispanic Black and 12.9% Hispanic; 33.2% had some college education, and 26.9% were college graduates. The prevalence of self-reported adequate sleep duration was 65.7%, while short sleep duration was 23.6%, and long sleep duration was 10.7%. Over a third (37.2%) had a history of ever telling a doctor or other health professional that they had trouble sleeping. Most (89.8%) had health insurance. Over half (51.3%) had a BMI of 30 kg/m² and above; 48.3% had high self-reported physical activity levels, while 26.2% were physically inactive. The prevalence of heavy alcohol intake was 32.0% (Tables 1 and 2).

In bivariate descriptive comparisons, participants with short sleep duration were younger and more likely to be male, non-Hispanic Black, have no health insurance, and work ≥45 h weekly. Conversely, individuals with long sleep duration were older and more likely to be female, not employed, have a high school or lower level of education, have comorbidity or moderate to severe depressive symptoms, and be non-drinkers and physically inactive (Tables 1 and 2).

The multinomial logistic analysis results shown in Tables 3 were derived from the complete data (n = 5660) created using multiple imputations, and all associations are presented relative to adequate sleep duration. In the unadjusted multinomial logistic regression results, all covariates except BMI and nativity status showed significant associations with habitual sleep duration (Table 3).

In the fully adjusted multinomial logistic model, age, sex, history of help-seeking for sleeping difficulty, depressive symptoms, and CKD were associated with habitual sleep duration. Older adults (≥ 65 years old) were less likely than young adults (18–44 years) to have short sleep duration (RRR, 0.63; 95% CI, 0.45–0.91). No age differences were noted in odds for being a long sleeper. Compared to men, women were less likely to be short sleepers (RRR, 0.70; 95% CI, 0.56–0.88) and more likely to be long sleepers (RRR, 1.24; 95% CI, 1.001–1.54). Those reporting a history of ever telling a doctor or other healthcare professional they had trouble sleeping were more likely to have short sleep duration (RRR, 1.25; 95% CI, 1.02–1.53).

Long sleep duration was significantly associated with moderate depressive symptoms (RRR, 1.62; 95% CI, 1.08–2.44) and moderately severe to severe depressive symptoms (RRR, 1.89; 95% CI, 1.05–3.43) compared to minimal or no depressive symptoms. In addition, those with CKD were more likely to have a long sleep duration than those with no CKD (RRR, 1.48; 95% CI, 1.14–1.92). Other comorbid conditions (diabetes mellitus, COPD or current asthma, heart disease, stroke, and arthritis) were not independently associated with habitual sleep duration (Table 3). Similarly, cigarette smoking status, alcohol intake, BMI, and total physical activity level were not independently associated with habitual sleep duration. The physical activity and sleep duration relationship findings remained the same when total physical activity level was substituted with physical activity derived from leisure-time and transport-related activities in the multinomial logistic model.

Non-Hispanic Black adults with hypertension were twice as likely as their non-Hispanic White counterparts to have short sleep duration (RRR, 2.08; 95% CI, 1.61–2.67). The odds of being a short or long sleeper in Hispanic, non-Hispanic Asian, and Other race/ethnic groups were not significantly different from that of the non-Hispanic White race/ethnic group (Table 3).

In the association between employment status and habitual sleep duration, individuals working for ≥45 h/week were more likely to have short sleep duration than those working 35–44 h/week (RRR, 1.81; 95% CI, 1.32–2.48). Those retired were more likely to be short sleepers (RRR, 1.41; 95% CI, 1.01–1.98) or long sleepers (RRR, 3.46; 95% CI, 2.18–5.49) than those working 35–44 h/week. Long sleep duration was also more likely in those not working due to health reasons (RRR, 4.87; 95% CI, 2.89–8.22) and those not working due to other reasons such as attending school and taking care of the family (RRR, 3.29; 95% CI, 1.84–5.88) compared to those working 35–44 h/week. Marital status, education level, and health insurance did not show independent associations with habitual sleep duration.

The adjusted multinomial logistic regression results from the complete imputed data model (n = 5660) were compared to those of complete case analysis (n = 4520). The direction and significance of the relationship between the covariates and habitual sleep duration in the two models were similar to except in employment status, help-seeking for sleeping difficulty, and depressive symptoms covariates where the level of significance differed (Additional file 2: Supplemental Table 2). In employment status, retirement was associated with short sleep in the imputed model (RRR, 1.41; 95% CI, 1.01–1.98) but not in complete case analysis (RRR, 1.39; 95% CI, 0.94–2.06). Help-seeking for sleeping difficulty was associated with short sleep in the imputed model (RRR,1.25; 95% CI, 1.02–1.53) but not in complete case analysis (RRR,1.20; 95% CI, 0.97–1.50). Finally, moderate depressive symptoms were associated with short sleep duration in the complete case analysis (RRR, 1.87; 95% CI, 1.23–2.85) but not in the imputed model (RRR, 1.61; 95% CI, 0.99–2.61).

Results from the complete case analyses showed no significant associations between depressive symptoms and long sleep duration, but in the complete imputed data model, long sleep duration was associated with both moderate (RRR,1.62; 95% CI, 1.08–2.44 versus RRR, 1.52; 95% CI, 0.98–2.37 in the complete case analysis) and moderately severe to severe depressive symptoms (RRR,1.89; 95% CI, 1.05–3.43 vs. RRR, 1.78; 95% CI, 0.84–3.76 in the complete case analysis). The results show that the differences noted were on the significance of the relationship between the specified variables and habitual sleep duration but not on the direction of the relationships. The complete case analysis model included 4520 observations (compared to 5660 observations for complete imputed data), translating to a 20% loss in information. This loss of information may bias results in the complete case analysis and may have contributed to the differences noted between the two models. Some of the differences observed could also be due to an increase in power in the imputed data model.