The National Health and Nutrition Examination Survey (NHANES) is a series of cross-sectional data with a complex, multi-stage, probability sampling design to obtain a nationally representative sample of the civilian, noninstitutionalized US population. Participants were linked to death certificate records from the National Death Index for prospective follow-up until December 31, 2019. This study analyzed NHANES data from the 1999 – 2016 and 2017 – 2020 March (pre-pandemic) cycles (cross-sectional) and their mortality data (prospective).

Among 107519 eligible participants with paid jobs, those who reported pregnancy (N = 1784), had history of CVDs (coronary heart disease, myocardial infarction, congestive heart failure, heart attack, angina/angina pectoris, or stroke) (N = 6802) or cancers (N = 4144), or without reporting working hours (N = 63976) were excluded, leaving a sample of 30813 workers. Further, workers with missing dietary data (N = 3329), extreme total energy intake (< 500 or > 8000 kcal/day) (N = 200), or missing survey weights (N = 3) were excluded. Based on each of the cardiometabolic outcomes, we obtained subsamples of obesity (N = 27114), hypertension (N = 26409), diabetes (N = 27281), Framingham CVD risk score (FRS) (N = 17325) as illustrated in S Figure 1.

All NHANES protocols were approved by the National Center for Health Statistics’s Research Ethics Review Board. This study followed the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines.

CVD mortality was the primary mortality outcome, available only until December 31, 2014, as deaths from cerebrovascular diseases were not followed afterward. Secondary outcomes, including all-cause and heart disease mortality, were available for participants from NHANES 1999–2018, followed until December 31, 2019. In addition to the previous exclusion criteria, participants younger than 18 years (no records of death) or not eligible for the mortality linkage were further excluded, leaving a subsample of 19682 for CVD mortality outcome and 24630 for all-cause mortality and heart disease mortality outcomes (S Figure 1b). Cause-specific mortality was identified using the International Classification of Diseases (ICD-10) codes, with heart disease mortality defined by I00–I09, I11, I13, I20–I51, and CVD mortality further including I60–I69.

Cardiometabolic outcomes were based on in-person health examinations at the Mobile Examination Center (MEC), laboratory results, or self-reported conditions, including obesity, diabetes, hypertension, and FRS.

Hypertension was determined if previously diagnosed by a physician, or systolic blood pressure (SBP) ≥ 130 mm Hg and/or diastolic blood pressure (DBP) ≥ 80 mm Hg, or use of any blood pressure-lowering treatment¹⁷. All available blood pressure readings were utilized to calculate the mean blood pressure for each participant¹⁸.

Weight and height were directly measured at the MEC and were used to calculate body mass index (BMI) (kg/m²). Obesity was defined as a BMI of 30 kg/m² or higher. A further checking using lower BMI cutoffs (≥ 27.5 kg/m²)¹⁹ for participants self-reported as non-Hispanic Asian (year 2011 onward) showed negligible differences in obesity prevalence (0.4% higher, data not shown). Thus, the universal cutoff of ≥ 30 kg/m² was retained for all participants in this study.

Diabetes was classified as diagnosed (i.e., self-reported previous diagnosis or use of any glucose-lowering medications) and undiagnosed (i.e., glycohemoglobin ≥ 6.5%, or fasting glucose ≥ 7 mmol/L, or 2-hour oral glucose tolerance test ≥ 11.1 mmol/L)²⁰.

FRS were calculated separately for males and females based on age, total cholesterol, high-density lipoprotein cholesterol, SBP, treatment for hypertension, diabetes, and smoking status²¹. Then, the scores were dichotomized as low CVD risk (score ≤ 20%) versus high CVD risk (score > 20%) as commonly applied in primary care²².

Working hours were self-reported and obtained from the question, “How many hours did you work at all jobs or businesses last week”.

Assessment of dietary intake in NHANES has been described previously²³. Briefly, a validated 24-hour dietary recall was applied during the in-person interview from year 1999 through 2002. Since the year 2003, a second dietary interview was added via telephone 3–10 days after the initial in-person interview. Individual food intakes were averaged based on two-day dietary data for the year 2003 and after.

Adherence to the EAT-Lancet was evaluated by calculating an EAT-Lancet diet score based on the consumption of 14 components, including seven emphasized components (vegetables, fruits, unsaturated oils, legumes, nuts, whole grains, fish) and seven limited components (beef and lamb, pork, poultry, eggs, dairy, potatoes, added sugar)¹⁶. A total score was calculated by summing the 14 component scores, ranging from 0 (worst) to 42 (best), with higher scores indicating higher diet quality.

Sociodemographic information in NHANES was collected during in-person interviews, including age, sex, race/ethnicity, education, marital status, and ratio of family income to poverty. Smoking status was self-reported based on lifetime and current cigarette use. Alcohol drinking status was defined according to the self-reported number of drinks in a lifetime and in a week²⁴. Total physical activity was measured using the Global Physical Activity Questionnaire and leisure time physical activity was defined as vigorous and moderate recreational activity. Metabolic equivalent (MET) minutes were calculated to quantify the weekly energy expenditure for leisure time physical activity²⁵.

To account for the complex survey design, sampling weights were applied to all analyses per NHANES guidelines²⁶. Weighted means (standard errors (SE)) were estimated for continuous variables, and counts (weighted percentages) were calculated for categorical variables.

In primary analysis, LWH (i.e., ≥ 55 hours/week) was defined according to WHO report³, and was dichotomized into two categories (yes vs. no). The EAT-Lancet diet score was grouped as high vs. low using a median cut-off of 24. Multivariable logistic regression models were used to assess the cross-sectional associations with cardiometabolic outcomes, adjusting for age (years), sex, race/ethnicity (Mexican American, other Hispanic, non-Hispanic white, non-Hispanic black, other race including multi-racial), education (less than high school, some high school, high school or more), poverty income ratio (<1 or ≥ 1), marital status (married/living with a partner, widowed/divorced/separated, never married), smoking status (non-smoker, former smoker, current smoker), alcohol drinking (non-drinker, light-to-moderate drinker, heavy drinker), leisure time physical activity (< or ≥ 500 MET min/week), and total energy intake (kcal/day). Multivariable Cox proportional hazards models were used to assess the prospective associations with mortality outcomes, adjusting for the same covariates. Stratified analyses for the association with CVD mortality were performed by baseline FRS, as previous studies have shown that the impact of LWH may be more greater in individuals with preexisting CVDs²⁷. Models assessing LWH were adjusted for diet scores, and models assessing diet scores were adjusted for working hours, both in original continuous form to optimize confounding control and statistical precision. Joint associations of LWH (no vs. yes) and EAT-Lancet diet (high vs. low) were analyzed using four categories: (i) no+ high (reference), (ii) no + low, (iii) yes + high, (iv) yes + low.

Multiplicative interaction was evaluated by adding an exposure interaction term in the model. The relative excess risk due to interaction (RERI) was used to examine additive interaction, calculated as RERIOR=OR11−OR10−OR01+1, where subscripts 1/0 indicate exposure presence/absence: OR11 (both present), OR10 and OR01 (each alone). RERI > 0 indicates a synergistic interaction, 0 indicates additivity, and < 0 indicates antagonism²⁸.

Multiple imputation was carried out to impute covariates with missing in the primary analyses. Several sensitivity analyses were performed. First, participants working < 35h/week were excluded because they were assumed to have health issues with reduced work ability for full-time jobs and higher risk of sequent health outcomes. Second, self-reported occupations, which were only available before NHANES 2015, were further adjusted in the multivariable models. Third, missingness in covariates was coded as a separate category to repeat analyses. Lastly, repeated regression analyses were conducted to further adjust for or excluding shift work.

Multiple imputations and forest plots were conducted in R (version 4.2.2), and all other analyses were performed using SAS 9.4 (SAS Institute Inc., Cary NC).

Among all participants, the weighted mean (SE) age was 39.5 (0.2) years, 55% were male, 87% had high school or more education, and 89% had income above the poverty level. In total, 14% worked long hours (i.e., ≥ 55 h/wk). LWH workers were more likely to be male, married or living with partners, smokers, light-to-moderate drinkers, higher educated, and had higher incomes (Table 1). Similar characteristics were observed in the CVD mortality subsample.

Overall, the weighted mean (SE) of the EAT-Lancet score was 24 (0.06), ranging from 8 to 42. Workers with low diet score were more likely to be younger, male, never married, smokers, heavy drinkers, less educated, less active and had lower income. Among all participants, 6.8% were in the LWH + low diet score group, 8.1% in the LWH + high diet score group, 49.8% in the working hours < 55 h/wk + high diet score group, and 35.3% in the working hours < 55 h/wk + low diet score group.

Compared to those working < 55h/wk, LWH was associated with higher odds of obesity in both crude (OR = 1.20, 95%CI = 1.08, 1.34) and adjusted models (OR = 1.20, 95%CI = 1.07, 1.34). LWH was also associated with higher odds of hypertension only in crude model (OR = 1.20 (1.08, 1.34), but not the adjusted model (OR = 1.01, 95%CI = 0.90, 1.13). No associations were found between LWH and diabetes, or FRS (Table 2).

Compared to the high diet score group, a low EAT-Lancet diet score was associated with a higher odds of obesity (crude: OR = 1.29, 95%CI = 1.20, 1.39; adjusted OR = 1.34, 95%CI=1.21, 1.42) and diabetes (crude OR = 1.17, 95%CI = 1.04, 1.32); adjusted OR = 1.34, 95%CI=1.17, 1.53) in both crude and adjusted models, but only crude model for FRS (OR = 1.20, 95%CI = 1.03, 1.40). A low diet score was also associated with higher odds of hypertension in adjusted model (adjusted OR = 1.12, 95%CI = 1.03, 1.20) and borderline statistically significant in crude model (crude OR = 1.07, 95%CI = 0.99, 1.15).

The joint associations of working hours and the EAT-Lancet diet score on cardiometabolic outcomes are presented in Figure 1(A–D). Compared to the group of working <55h/wk + high diet score, participants in the group of LWH + low diet score had increased odds of obesity (adjusted OR = 1.61, 95%CI = 1.36, 1.92), diabetes (adjusted OR = 1.33, 95%CI = 1.01, 1.76), and hypertension (adjusted OR = 1.16, 95%CI = 0.96, 1.39), adjusting for confounders. No association was observed for FRS. The RERI for obesity outcome was greater than 0 (RERI = 0.15, 95%CI = 0.10, 0.21), indicating a synergistic interaction between LWH and low diet scores. The RERIs for diabetes (RERI = −0.27, 95%CI = −0.56, 0.02), hypertension (RERI = 0.15, 95%CI = 0.10, 0.26), and FRS (RERI = −0.14, 95%CI = −0.28, 0.00) were more uncertain and included positive, zero and negative interaction between LWH and low diet scores.

The median (interquartile) follow-up time for all-cause and heart disease mortality was 10.3 (5.5–15.2) years, and 7.4 (3.6–11.5) years for CVD mortality. During the follow-up period, 1060 all-cause deaths, 221 heart disease deaths, and 257 CVD deaths were documented. The overall mortality rates were 43.0 per 1000 for all-cause, 9.0 per 1000 for heart disease, and 13.1 per 1000 for CVD mortality.

No association was observed between LWH and CVD mortality in crude (HR = 1.02, 95%CI = 0.63, 1.63) and adjusted models (adjusted HR = 0.97, 95%CI = 0.60, 1.56) among all workers. In workers with high CVD risk (i.e., high FRS scores) at baseline, LWH was associated with a 64% higher hazard of CVD mortality compared to those working < 55h/wk in the adjusted model (adjusted HR = 1.64, 95%CI = 0.79, 3.12), although it was not statistically significant. No association was found between LWH and CVD mortality in workers with low FRS scores (adjusted HR = 0.90, 95%CI = 0.44, 1.83). No associations were found for LWH with all-cause mortality (adjusted HR = 0.93, 95%CI = 0.69, 1.25), or heart disease mortality (adjusted HR = 0.86, 95%CI = 0.51, 1.46) among all workers (Table 3).

A low Eat-Lancet diet score was associated with higher hazard of CVD mortality in both crude (HR = 1.34, 95%CI= 0.91, 1.97) and adjusted models (adjusted HR = 1.45, 95%CI= 0.96, 2.19) among all workers, whereas increased CVD mortality risk only among participants with low FRS scores (adjusted HR = 1.98, 95%CI = 1.10, 3.57). Additionally, a low diet score was associated with higher hazard of all-cause mortality (adjusted HR = 1.25, 95%CI = 1.05, 1.50) and heart disease mortality (adjusted HR = 1.62, 95%CI = 0.98, 2.69) among all workers, after adjusting for confounders (Table 3).

The joint associations on mortality are shown in Figure 1(E–G). The LWH + low diet score group had higher hazards for CVD mortality (adjusted HR = 1.73, 95% CI = 0.60, 3.09), heart disease mortality (adjusted HR = 1.50, 95% CI = 0.65, 3.48), and all-cause mortality (adjusted HR = 1.13, 95% CI = 0.60, 1.73) compared to no LWH + high diet score group. The RERI for CVD mortality indicated a synergistic interaction (RERI = 0.21, 95% CI = 0.03, 0.39), while RERIs for all-cause (RERI = −0.08, 95% CI = −0.16, 0.01) and heart disease mortality (RERI = 0.23, 95% CI = −0.72, 1.18) suggested an additive interaction.

Results for several sensitivity analyses are presented in S Tables 1–4. Overall, the results remained unchanged when 1) excluding participants working < 35h/week, 2) including occupation types in the multivariable models, and 3) coding missing values as a separate category, 4) adjusting shift work status or excluding shift worker.