Associations of healthy lifestyle and socioeconomic status with mortality and incident cardiovascular disease: two prospective cohort studies

Study population

US NHANES recruited a representative sample of civilian, community dwelling members of the US population using a complex, multistage probability design. The survey was conducted periodically before 1999 and continuously thereafter. Details of the study design and data collection have been previously described.9 Although US NHANES has released cross sectional questionnaire, examination, and laboratory data up to 2018, mortality data were updated to 31 December 2015. Accordingly, the current analysis included 61 202 participants who were aged 20 years and older and not pregnant at baseline in US NHANES III (1988-94) and continuous NHANES (1999-2014) surveys. Those with missing information on socioeconomic factors (n=6939), lifestyle factors (n=8156), other covariates (n=1619), and deaths (n=26) were excluded from the analysis. Overall, 44 462 participants from US NHANES were included (supplementary fig 1).

UK Biobank recruited more than 500 000 participants aged 37 to 73 years from 22 assessment centers across England, Scotland, and Wales between 2007 and 2010. Details of the study design and data collection have been described previously.10 Among the 502 492 participants, we excluded those with missing information on socioeconomic factors (n=77 962), lifestyle factors (n=20 029), and other covariates (n=4964). Overall, 399 537 participants were included (supplementary fig 1). For the analysis of incident CVD, we only included those without prevalent CVD (n=324 517) at baseline.

Assessment of SES

In US NHANES, self-reported family income level, occupation, education level, and health insurance were used to measure SES according to previous studies,711 and each factor was divided into three levels (low, medium, and high) with consideration of practical interpretation and sample size within levels. The family income level was operationalized using the family poverty to income ratio, which reflected the annual family income relative to the federal poverty level and was comparable across surveys since income thresholds were updated for inflation and family size each year.12 According to a published study and the Patient Protection and Affordable Care Act, we grouped participants according to the poverty to income ratio: low (≤1), middle (1-4), and high (≥4).12 Education was categorized into less than high school diploma, high school graduate or equivalent, and college or above.13 Occupation was classified based on the widely used socioeconomic index in the US,14 and each occupation was rated according to the employees’ earnings, education level, and prestige. The socioeconomic index ranged between 13.98 and 90.45,15 and occupation was categorized into upper (socioeconomic index ≥50), lower (socioeconomic index <50, including retirees16 and students), and unemployment. Health insurance was categorized into private health insurance (including any private health insurance, Medi-Gap, or single-service plan), public health insurance only (including Medicare, Medicaid, State Children’s Healthcare Plan, military healthcare, Indian Health Service, State Sponsored Health Plan, or other government programme), and no health insurance.17 An overall SES variable was created using latent class analysis based on family income level, occupation, education level, and health insurance (each factor had three levels).7 The latent class analysis, which uses multiple observed categorical variables to generate an unmeasured variable (ie, latent variable) with a set of mutually exclusive latent classes, was conducted using PROC LCA, a new SAS procedure.18 Three latent classes were identified, which respectively represented a high, medium, and low SES according to the item-response probabilities.19 The supplementary file describes the data collection and latent class analysis in US NHANES.

In UK Biobank, total household income before tax was obtained through questionnaires, and participants could choose an option from <₤18 000 ($25 000; €21 000), ₤18 000-£30 999, ₤31 000-£51 999, ₤52 000-£100 000, >₤100 000, do not know, or prefer not to answer. A total of 14.3% of participants chose the last two options and were excluded from the main analyses as missing values; however, we included them in sensitivity analyses when evaluating single socioeconomic factors, consistent with a previous study,20 on the basis that these participants might be more likely to have lower SES. Participants reported their education qualifications as college or university degree; A levels, AS levels, or equivalent; O levels, GCSEs, or equivalent; CSEs or equivalent; NVQ, HND, HNC, or equivalent; other professional qualifications; none of the above (equivalent to less than high school diploma); or prefer not to answer (which was excluded from our analyses as missing values). As UK Biobank only acquired employment status instead of information on specific occupation at baseline, we regrouped participants into two groups: employed (including those in paid employment or self-employed, retired, doing unpaid or voluntary work, or being full or part time students) and unemployed. Because the National Health Service, a publicly funded healthcare system aiming to provide comprehensive, universal and free services, is implemented in the UK,21 we did not consider health insurance as a component of SES in UK Biobank. An overall SES variable was created using latent class analysis based on three individual socioeconomic factors (household income, education level, and employment status). We did not regroup household income and education level into three groups as we did in US NHANES because of the larger sample size in UK Biobank and failure of model convergence owing to fewer observed groups if the two variables were regrouped. Three latent classes were identified, which respectively represented a high, medium, and low SES according to the item-response probabilities. Details are reported in the supplementary file.

In UK Biobank, Townsend deprivation index was available as an area level SES variable derived from national census data according to postcodes of residence, which considered car ownership, household overcrowding, owner occupation, and unemployment.22 A higher Townsend deprivation index denotes lower area level SES.22

Assessment of lifestyle factors and other covariates

Since multiple lifestyle factors are interrelated and are associated with mortality and morbidity, we constructed a healthy lifestyle score including cigarette smoking, alcohol consumption, physical activity, and diet according to a previous US NHANES study23 and that coincided with recommendations from the World Health Organization.24 All lifestyle factors were obtained through structured questionnaires and 24 hour dietary recalls. Never smoking was considered as a healthy level, which was defined in the questionnaire as smoking fewer than 100 cigarettes in life. Frequency and volume of current alcohol consumption were self-reported, and a healthy level was defined as daily consumption of one drink or fewer for women and two drinks or fewer for men, according to the dietary guidelines in the US and UK (one drink contains 14 g of ethanol in the US and 8 g in the UK).2526 For physical activity, different assessment questions were used between the US and UK studies, and questionnaires also varied in different survey years in US NHANES. Nevertheless, weekly metabolic equivalent hours of leisure time physical activity were calculated in US NHANES 1999-2014 and UK Biobank, whereas monthly frequency of leisure time physical activity was calculated in US NHANES 1988-94. To harmonize the data, we further classified the participants into thirds and defined the top third as a healthy level of physical activity.

In US NHANES, dietary quality was obtained from 24 hour dietary recalls and was assessed by healthy eating index (HEI) scores. The HEI-2015 was calculated for the 1999-2014 survey cycles, which aligns with the 2015-20 Dietary Guidelines for Americans.27 However, because food codes used in the 1988-94 cycles could not match those used in the 1999-2014 cycles, we used HEI-1995 for the 1988-94 cycles and the variable was directly provided by the original dataset. HEI-1995 aligns with the food guide pyramid released by the US Department of Agriculture in 1992.27 Supplementary table 1 provides details of constructions of HEI-1995 and HEI-2015, and both scores reflected the overall dietary quality according to the contemporary dietary guidelines. A healthy diet was defined as the health eating index in the top two fifths of distribution.28 In UK Biobank, dietary information was obtained through questionnaires and did not contain energy or salt intakes, thus we could not calculate the HEI scores. Instead, according to a previous UK Biobank study,29 we evaluated dietary quality using a more recent dietary recommendation for cardiovascular health, which considered adequate consumption of fruit, vegetables, whole grains, fish, shellfish, dairy products, and vegetable oils and reduced consumption of refined grains, processed meats, unprocessed meats, and sugar sweetened beverages. We defined a healthy diet as meeting at least five items of the recommendations (see supplementary table 2).

For each lifestyle factor, we assigned 1 point for a healthy level and 0 points for an unhealthy level. Thus, the healthy lifestyle score was the sum of the points and ranged between 0 and 4, with higher scores indicating healthier lifestyles. Although this simple additive method has been used widely,303132 the underlying assumption is that the associations between different lifestyle factors and the outcome were identical, which might not be true. Thus we also constructed a weighted lifestyle score, where each lifestyle factor was weighted by its association with the outcome. Body mass index (BMI) was not included in the lifestyle score given the concern that it could be an intermediate factor between behavioral factors and health outcomes. In addition, the obesity paradox is a concern,33 and overweight and obesity might not be strongly associated with mortality in older people.13 Nevertheless, we also included baseline BMI in the lifestyle score in a sensitivity analysis, and healthy bodyweight was defined as a BMI of 18.5-24.9.28

Other covariates were obtained through questionnaires, including age; sex; marital status (US NHANES only); assessment centers (UK Biobank only); self-reported race; an acculturation score based on the country of birth, length of time in the US or UK, and language spoken at home (see supplementary file);34 history of hypertension, diabetes, CVD, or cancer; and history of chronic bronchitis, emphysema, or chronic obstructive pulmonary disease (UK Biobank only). Diagnoses of CVD and cancer were also obtained through linked hospital admissions data and cancer registry in the UK Biobank. Bodyweight and height were measured at baseline, with BMI calculated as weight (kg)/(height (m)²).

Outcome ascertainment

Outcomes were classified using ICD-9 and ICD-10 (international classification of diseases, ninth and 10th revisions, respectively) codes. The primary outcomes included all cause mortality, CVD mortality, and incident CVD. In US NHANES, deaths were obtained through the National Death Index to 31 December 2015.35 In UK Biobank, deaths were obtained through death certificates held within the NHS Information Centre (England and Wales) and the NHS Central Register (Scotland) to 30 April 2020.36 CVD diagnoses, including myocardial infarction (ICD-9 codes 410-412 and 429.79; ICD-10 codes I21-I23, I24.1, and I25.2) and stroke diagnoses (ICD-9 codes 430, 431, 434, and 436; ICD-10 codes I60, I61, I63, and I64), were obtained through linked hospital admissions data including Hospital Episode Statistics-Admitted Patient Care (England), Scottish Morbidity Records-General/Acute Inpatient and Day Case Admissions (Scotland), and Patient Episode Database for Wales as well as death register data to 31 January 2018.3738 Secondary outcomes were mortality from heart disease (ICD-10 codes I00-I09, I11, I13, and I20-I51 in US NHANES), coronary heart disease (ICD-10 codes I20-I25 in UK Biobank), and stroke (ICD-10 codes I60, I61, I63, and I64 in UK Biobank), as well as incident myocardial infarction and stroke. Mortality from cerebrovascular disease or total CVD was not considered in NHANES because the US National Death Index matched mortality dataset stopped updating data on deaths from cerebrovascular diseases after 31 December 2011.

Statistical analysis

To estimate appropriate variance and statistics representative of US adults, our analysis in US NHANES considered the oversampling, stratification, and clustering according to the NHANES statistical analysis guideline.39 Baseline characteristics were described across different levels of SES, and differences among groups were tested by analysis of variance adjusted for sampling weights for continuous variables and Rao-Scott χ² test for categorical variables in US NHANES, and by analysis of variance and χ² test in UK Biobank.

We used Cox proportional hazard regression models to estimate the hazard ratios and 95% confidence intervals of outcomes associated with SES and lifestyle score. The proportional hazards assumption was examined by creating a product term of follow-up time and SES, and we found no significant deviation from the assumption.40 Person years were calculated from baseline until the date of death or diagnosis (for the incident CVD analysis), or end of follow-up, whichever occurred first. Based on previous researches,2023 model 1 included SES; age; sex; self-reported race; marital status (US NHANES only); assessment centers (UK Biobank only); acculturation; BMI; and history of hypertension, diabetes, CVD, cancer, chronic bronchitis, emphysema, or chronic obstructive pulmonary disease. Model 2 additionally included the healthy lifestyle score. We used the difference method to calculate the mediation proportion by the mediator (overall lifestyle) for the association between SES and each outcome—that is, comparing estimates from models with and without the hypothesized mediator.41 We additionally calculated the C statistics of the two models to compare the predictions with versus without the healthy lifestyle score.

We further conducted a stratified analysis by latent class of SES to investigate associations of the lifestyle score with health outcomes among adults in different socioeconomic subgroups. As only 838 (2.2%) and 3495 (9.4%) US adults had 0 and 4 points of healthy lifestyle score, and the corresponding numbers in the UK Biobank were 49 545 (12.4%) and 9841 (2.5%), we merged participants with 0 points and 1 point as well as those with 3 and 4 points to increase the statistical power. In this analysis, the reference group was set as the participants with unhealthy lifestyles (lifestyle scores of 0 or 1), and we examined whether adherence to healthy lifestyles was associated with protection against mortality and incident CVD across different SES subgroups. To quantify the additive and multiplicative interactions, we additionally included a product term of SES (low, medium, and high) and healthy lifestyle score (0 or 1; 2; and 3 or 4 points) in the model. The hazard ratio with its 95% confidence interval of the product term was the measure of interaction on the multiplicative scale. We used the relative excess risk due to interaction (RERI) and corresponding 95% confidence intervals as the measure of interaction on the additive scale, calculated using the coefficients and corresponding standard errors of the product term, SES, and lifestyle score, as well as covariance matrix.42

To assess the joint associations, we further classified participants into nine groups according to SES (low, medium, and high) and healthy lifestyle score (0 or 1; 2; and 3 or 4 points) and estimated hazard ratios of mortality and incident CVD in different groups compared with those with high SES and three or four healthy lifestyle factors.

To test the robustness and potential variations in different subgroups, we repeated all analyses stratified by sex (men and women), self-reported race (white and non-white participants), and age groups (<60, and ≥60, defined as elders by the World Health Organization43).

We conducted several sensitivity analyses. First, we repeated all analyses by substituting SES with each socioeconomic factor (ie, family income level, occupation or employment status, education level, and health insurance), and these factors were mutually adjusted in the models. Similarly, we also used the individual lifestyle factors instead of the score in the models to evaluate whether the estimated mediation proportion was similar to that of the main analysis. Second, a weighted healthy lifestyle score was constructed to account for varied magnitudes of the associations between different lifestyle factors and outcomes.44 Third, we constructed a lifestyle score including baseline BMI. Fourth, we excluded individuals with prevalent diabetes, CVD, cancer, chronic bronchitis, emphysema, or chronic obstructive pulmonary disease because both lifestyles and SES could be influenced by major chronic diseases. Fifth, we excluded events that occurred within the first three years of follow-up to reduce potential reverse causation. Sixth, we restricted the analysis to those aged 40 years or older in US NHANES to coincide with the age distribution in UK Biobank, and to reduce the concern that SES is prone to change and the risk of mortality due to lifestyles is relatively lower in younger adults. As only five participants in UK Biobank were aged less than 40 years, this sensitivity analysis was not performed in UK Biobank. Seventh, we used multiple imputation to impute all missing independent variables to test the influence of missing variables.45 Eighth, we assigned 0, 1, and 2 points to each low, medium, and high level socioeconomic factor (for employment status in UK Biobank, only 0 and 2 points were assigned for unemployed and employed status) and added the scores to get a socioeconomic score (range 0-8 in US NHANES and 0-6 in UK Biobank). As only 756 (0.9%) participants in US NHANES and 5000 (1.3%) in UK Biobank had a score of 0, we merged those with 0 or 1 point. The socioeconomic score was then used in all analyses instead of the latent class derived SES variable. Ninth, in the final model in UK Biobank we further included the Townsend deprivation index, a variable reflecting the area level SES, for two purposes: to evaluate whether the association between individual level SES and health outcomes remained robust when controlling for area level SES, and to repeat all the main analysis using Townsend deprivation index as the SES variable, instead of the individual level SES variable. Tenth, we additionally included quadratic terms of age in the models to consider the possible non-linear associations of age with health outcomes.

All analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC). We considered two sided P values <0.05 to be significant.

Patient and public involvement

The analyses were based on existing data of two cohort studies in general populations, US NHANES and UK Biobank, and we did not participate in the participant recruitment. To our knowledge, no patients were involved in the design, recruitment, or conduct of the studies. The research question and outcome measures of the present study were proposed by systematically reviewing the evidence of the associations between lifestyles and non-communicable diseases, and no patients were involved in the process. Participants from the two cohorts were deidentified, and thus we could not disseminate the results to each participant; however, the results will be disseminated to the public through broadcasts and popular science articles.