Study population
Between 1993 and 1997, the Danish Diet, Cancer, and Health Study recruited 57,053 participants of the greater areas of Copenhagen and Aarhus. Of these, 56,468 completed a food-frequency questionnaire (FFQ) at baseline and were without a cancer diagnosis prior to enrolment. The following databases were cross-linked to the cohort: The Civil Registration System [15], The Integrated Database for Labor Market Research [16], The Danish National Patient Register (DNPR)[17] and The Danish National Prescription Registry [18].
Participants were excluded if they had any prevalent CVD (n = 2933), self-reported or recorded diagnosis in the DNPR. Furthermore, participants were excluded if they had missing or implausible values for any covariates (n = 192), or implausible energy intakes [< 2092 kJ/day (< 500 kcal/day) and > 20,920 kJ/day (> 5000 kcal/day), n = 193] Supplemental Figure 1.
This study was approved by the Danish Data Protection Agency (Ref No. 2012-58-0004 I-Suite nr: 6357, VD-2018-117).
Exposures
Prior to their first visit, participants completed a 192-item validated FFQ [19]. Participants reported their average intakes of different food and beverage items over the previous 12 months.
Vegetable nitrate
Vegetable nitrate intake was calculated from vegetables assessed using the FFQ and quantified in grams/day. A recently published comprehensive database, with nitrate data for 178 vegetables from 255 publications, was applied to assess nitrate levels for each vegetable [3]. The nitrate content of vegetables varies according to country of cultivation; therefore, the following strategy was employed. For each vegetable, if 3 or more entries were available in the database for Northern Europe (Denmark, Estonia, Finland, Iceland, Ireland, Latvia, Lithuania, Norway, Sweden and United Kingdom), the median of these values was used. If there were less than 3 entries in the database for Northern Europe, the median of values for all European countries was used. If there were less than 3 entries available for European countries, the median of values for all countries in the database was used. Vegetable nitrate intake (mg/day) was calculated by multiplying the quantity reportedly consumed (g/day) by the median nitrate content (mg/g) of that vegetable. Based on information in the nitrate content of vegetables database [3] and nitrate loss estimates in another cohort [20], for those vegetables not consumed raw, a 50% reduction in the assigned nitrate value was applied to take into account the effect of cooking. The nitrate values from each individual vegetable were added to obtain the sum of daily vegetable nitrate values.
Non-vegetable dietary nitrate
Nitrate intake from non-vegetable items, excluding water, was calculated using values from Inoue-Choi et al. [21]. Nitrate intake was calculated by multiplying the reported quantity of consumption for each food item (g/day) by its assigned mean nitrate value (mg/g). Total non-vegetable nitrate intake (mg/day) was determined by calculating the sum of nitrate values from all items included in the FFQ, excluding vegetables.
Study outcomes
Blood pressure
Baseline SBP and DBP was measured by trained staff members using automated oscillometric sphygmomanometers (model UA 751 or UA-743; Takeda Pharmaceutical Co. Ltd., Osaka, Japan). A measurement was taken on the right arm, with participants lying in a supine position after at least 5 min rest and a minimum of 30 min after tobacco smoking and intake of food, tea, or coffee. Where a participants’ SBP was ≥ 160 mmHg, or DBP was ≥ 95 mmHg, the measurement was repeated after 3 min, with the lower of the two measurements used.
Incident CVD and CVD subtypes
For the time-to-event analysis, the primary outcome was a combined endpoint of first-time incident CVD. Incident CVD was defined as a hospitalization with a primary or secondary diagnosis of ischemic heart disease (IHD), ischemic stroke, hemorrhagic stroke, peripheral artery disease (PAD), heart failure, or inpatient or outpatient diagnosis of atrial fibrillation (AF). Secondary outcomes were: IHD, ischemic stroke, hemorrhagic stroke, PAD, heart failure, and AF discretely. All outcomes were identified by ICD-10 codes using the DNPR. These ICD codes have previously been validated for research purposes in the DNPR with positive predictive values (PPVs) of 92–97% for IHD [22], 81 – 85% for ischemic stroke [23, 24], 88% for hemorrhagic stroke [25], 81% for PAD [26], 76% for heart failure [22] and 95% for AF [22]. We did not include a diagnosis of unstable angina or transient ischemic attack as these in the DNPR lack validity for research purposes.
Validated case analysis
Using only medically reviewed and validated cases of certain CVD diagnoses [myocardial infarction (MI), ischemic stroke, and PAD] during follow-up, we investigated associations to verify our results based on the ICD-10 codes for outcomes. Patients with an ICD-10 discharge code for ischemic stroke, PAD or MI up until 2009, registered as a primary or secondary diagnosis, were considered possible cases to be reviewed (Supplemental Table 1). Due to a prior diagnosis of validated CVD, a further seven participants were excluded in this analysis (n remaining in analysis = 53,143). The methods of case validation have been published previously [24, 26, 27].
Covariates
Information on age, sex, education, smoking habits, alcohol consumption, and daily activity was obtained from self-administered lifestyle questionnaires completed by participants. Dietary data were obtained from the semi-quantitative FFQ described above. Anthropometric measurements were taken, and SBP, DBP, and cholesterol levels were measured at the study centers. Household annual income after taxation and interest for the value of the Danish currency was averaged over the 5-years immediately prior to study enrolment in 2015. ICD-8 and ICD-10 codes were used for diagnosis of chronic kidney disease, chronic obstructive pulmonary disease, and cancers. For diabetes mellitus, only self-reported data were used due to the low validity of ICD-codes in DNPR [28]. For treatment of diabetes mellitus, both self-reported data and data on filled prescription for insulin and non-insulin medication were used. Use of antihypertensive medication was obtained from self-reported use or the use of two or more antihypertensive medications within 180 days prior to study enrolment. Presence of hypertension was defined by a combination of the use of two or more antihypertensive medications within 180 days prior to study enrolment or self-reported hypertension, which has a PPV of 80.0% and a specificity of 94.7% to predict hypertension [28, 29]. Statin use at study enrolment was obtained from a combination of self-reported use and filled prescriptions. Hypercholesterolaemia was defined by self-reported hypercholesterolaemia or self-reported statin-use. Prescriptions were identified by ATC codes in the Danish National Prescription Registry if they were claimed within 180 days prior to enrolment in the study (from 1994 onwards). All ATC codes used are presented in Supplemental Table 2.
Statistical analysis
Baseline characteristics of the cohort are presented overall, and across quintiles of vegetable nitrate intake. First, a cross-sectional linear regression analysis was performed to investigate the association between vegetable and non-vegetable nitrate intake in quintiles and baseline SBP and DBP, both in the whole population and stratified by use of antihypertensive medication. Second, participants’ time-to-events were based on a maximum of 23 years of follow-up from the date of enrollment until the date of death, emigration, event of interest, or end of follow-up (August 2017), whichever came first. Nonlinear relationships were examined using restricted cubic splines, with hazard ratios (HRs) based on Cox proportional hazards models. All HRs and 95% confidence intervals (CIs) were obtained from the model with the exposure fitted as a continuous variable through a restricted cubic spline; HR estimates are reported for the median intake in each quintile with the first quintile median as the reference point and are graphed over a fine grid of x values. Cox proportional hazards assumptions were tested using log–log plots of the survival function versus time and assessed for parallel appearance, with no violation found. Our main models of adjustment were: Model 1a adjusting for age and sex; Model 1b adjusting for all covariates in Model 1a plus BMI, smoking status (current/former/never), physical activity (total daily metabolic equivalent), social economic status (income), marital status, education, pure alcohol intake (g/day), hypercholesterolemia (yes/no), and prevalent disease (diabetes, chronic obstructive pulmonary disease, chronic kidney disease, and cancer); entered into the model separately); Model 2 adjusting for all covariates in Model 1b plus energy; and Model 3: adjusting for all covariates in Model 2 plus intakes (g/day) of fish, red meat, polyunsaturated fatty acids, monounsaturated fatty acids, saturated fatty acids, and all fruit. Covariates were chosen a priori using expert and prior knowledge of potential confounders of nitrate intake and CVD. Additionally, standard logistic regression models were used to obtain the 20-year absolute risk estimates of incident CVD and CVD subtypes.
Analyses were further stratified by sex, BMI (above/below 30 kg/m2), alcohol intake (above/below 20 g/day), smoking status (ever/never), and by tertiles of total vegetable intake. When stratifying by alcohol intake and BMI, all participants with an alcohol intake of zero (n = 1180) and a BMI < 18.5 (n = 426) were excluded due to potential underlying pathologies or habits that may increase CVD risk. Stratification cut-off points of 20 g alcohol/day and a BMI of 30 kg/m2 were chosen as risk of disease is higher beyond these levels [30, 31]. We tested for interactions using chi-squared tests comparing nested standard cox proportional hazards models, using quintiles of vegetable nitrate intake.
The extent to which the association between vegetable nitrate intake and CVD was mediated by baseline SBP was quantified through natural direct and indirect effects [32], using a Cox proportional hazards model in the medflex package for R [33].
Analyses were undertaken using STATA/IC 14.2 (StataCorp LLC) and R statistics (R Core Team, 2019)[34].