Study design and participants
Data came from the Seniors-ENRICA-1 study, a representative cohort of the non-institutionalized persons aged ≥ 60 years in Spain (ClinicalTrials.gov identifier NCT01133093) [18, 19]. The study participants were recruited between March 2008 and September 2010 by stratified cluster sampling. First, the sample was stratified by province and size of the municipality. Second, clusters were selected randomly in 2 stages: municipalities and census sections. Finally, the households within each section were selected by random telephone dialing. Subjects in the households were selected proportionally to the distribution of the population of Spain by sex and age group.
A detailed diet history, a comprehensive set of physical measurements, and a blood test were collected at home visits by trained personnel, whereas data on sociodemographic and lifestyle variables, morbidity, and health services use were gathered through telephone interviews [18]. Study participants were contacted again between February and November 2012 to update information on diet and other study variables and were followed through January 2020 to ascertain vital status. All subjects gave written informed consent, and the Clinical Research Ethics Committee of the “La Paz” University Hospital in Madrid approved the research protocol.
From the 3483 subjects recruited at baseline, we excluded 318 with inadequate data (13 had implausible energy intakes, 239 had incomplete information on diet, and 254 on potential confounders; note that one individual may lack data in more than one variable). Hence, the analytical sample comprised 3165 individuals (Additional file 1: Fig. S1). From these, the 3-year follow-up food consumption was available in 2000 individuals, whereas in 1165, only baseline food consumption was available.
Study variables
Diet
The main exposure variable was 3-year cumulative adherence to the SEAD. Food consumption was obtained in the 2008–2010 and the 2012 visits with a validated electronic diet history [18, 20]. Subjects could report up to 861 foods and recipes habitually consumed in Spain. Portion sizes were estimated with the help of 127 digitized photographs and household measures. Nutrient and energy intake were estimated with Spanish food composition tables [20]. A previous validation study comparing the results of the diet history against seven 24-h recalls over 1 year showed a mean correlation coefficient of 0.53 for all 15 food groups considered, of 0.76 for energy, and of 0.55 for all 41 nutrients studied [20].
To estimate the adherence to the SEAD, we used the food components and scorings proposed by Oliveira et al. [4], which have been used by most of the subsequent studies on this dietary pattern [6,7,8,9,10]. We first calculated the habitual consumption (g/day) of each of the 9 components of this dietary pattern: fresh fish (excluding cod), cod, red meat and pork products, dairy, legumes and vegetables (excluding those consumed in soup), vegetable soup, potatoes regardless of the cooking method, whole-grain bread, and wine. For those study participants who were followed up at 3 years, we averaged the baseline and follow-up food consumption; for those who were not, we used the baseline food consumption. Second, we computed every food component—except wine—as g/1000 kcal/day and calculated their respective sex-specific medians. The subjects who were above the median consumption were scored 1 point, whereas those who were at or below it scored 0 points. As regards wine consumption, men who drank > 0 and ≤ 2 glasses/day and women who drank > 0 and ≤ 1 glasses/day were given 1 point, whereas no points were given for > 2 glasses/day in men, > 1 glass/day in women, or 0 glasses/day. We finally obtained the adherence to the SEAD as the sum of these 9 food group scores; it ranged from 0 to 9, with higher values indicating better adherence.
To place the SEAD into context, we calculated the scores of two other healthy eating patterns: the MEDAS index [21], which reflects the adherence to the Mediterranean diet, and the Alternate Healthy Eating Index (AHEI) [22], whose components were selected based on its association with chronic disease risk. To do this, we first calculated the consumption of the food components of the dietary pattern (14 for the MEDAS and 11 for the AHEI). Second, we scored these according to established cutoff points (scores of 0 or 1 for the MEDAS and 1 to 10 for the AHEI). Third, we summed all components to obtain the final score, which ranges from 0 to 14 for MEDAS and from 0 to 110 for the AHEI; higher values indicate better adherence in both scores (Additional file 1: Table S1).
Mortality
On one hand, vital status was ascertained with the National Death Index of Spain, an information system that collects the personal data of every demise recorded on civil registries nationwide [23]. Subjects were matched to the index with combinations of first and last names, birthdates, and national identity card numbers. Hence, the main outcome variable was death from any cause on or before January 31, 2020. Time to death was calculated as the difference between the date of death and the baseline telephone interview.
On the other hand, information on causes of death on or before December 31, 2018, was taken from the National Institute of Statistics of Spain [24]. These data are based on the death certificates of the deceased Spanish residents. Causes of death were classified and grouped according to the International Classification of Diseases (ICD), 10th revision. We considered ICD codes ranging from I00 to I99 to be cardiovascular deaths and those from C00 to D48 to be cancer deaths.
Potential confounders
We used data on several potential confounders of the association between the SEAD and mortality, specifically age, sex, educational level (primary or less, secondary, or university), energy intake (kcal/day), tobacco smoking (never, former, or current), recreational physical activity, time spent watching TV (h/day) (as a proxy of sedentary behavior) [25], BMI, and morbidity. Physical activity was assessed with the validated questionnaire developed in the EPIC study in Spain [26] and expressed as metabolic equivalents of task hours/week [27]. TV hours/day were assessed with the Nurses’ Health Study questionnaire validated in Spain [28]. BMI was calculated as weight (kg) divided by height (m) squared, both measured under standardized conditions [29]. As regards morbidity, we considered that a subject had diabetes if he/she either had blood glucose levels ≥ 126 mg/dl, was being treated with antidiabetic drugs, or reported that their doctor had given them a diabetes diagnosis. The medical diagnoses of cardiovascular disease (heart attack, stroke, or heart failure), chronic obstructive respiratory disease, musculoskeletal disease (osteoarthritis, arthritis, or hip fracture), cancer, and depression requiring medical treatment also were self-reported.
Statistical analyses
Differences in baseline characteristics and nutrient intakes of study participants across categories of the SEAD score were evaluated with Pearson’s chi-squared tests for discrete variables and analysis of variance for continuous variables.
The association between SEAD adherence and all-cause mortality was summarized with hazard ratios (HR), and their 95% confidence interval (CI), estimated with Cox proportional hazards regression. To control for potential confounding, three incrementally adjusted models were used: (1) adjusted for sociodemographic characteristics (age, sex, and educational level) and energy intake, (2) additionally adjusted for lifestyle variables and morbidity (tobacco smoking, physical activity, time watching TV, BMI, diabetes, cardiovascular disease, respiratory disease, musculoskeletal disease, cancer, and depression), and (3) further adjusted for the consumption of common foods not included in the SEAD (white meat, fruits, and refined grains). In the analyses, we used baseline values for categorical variables and averaged the baseline and 3-year follow-up values for continuous variables.
The adherence to the SEAD was modeled in the analyses as (1) a continuous variable (per 1-SD increment), (2) quartiles (using the lowest one as the reference), and (3) a restricted cubic spline (knots located at 2.5, 3.5, and 4.5 points). The adherence to the MEDAS and AHEI scores was modeled alike. When we examined the individual food components comprising the SEAD and the MEDAS, they were entered into the models as dichotomous variables (above or below specific food consumption thresholds). Conversely, components of the AHEI were modeled as continuous variables. Further details can be found in Additional file 1: Table S1.
We conducted several sensitivity analyses: First, we calculated the adherence to the 9 SEAD food components in two other different ways: (1) scoring 1, 2, 3, or 4 points if subjects were respectively in the 1st, 2nd, 3rd, or 4th sex-specific quartile of the consumption of the food component (in g/1000 kcal/day) and (2) scoring 1 point if subjects consumed < 1 serving/week of the food component (in g/week), 2 points for 1 to 7 servings/week, and 3 points for ≥ 1 serving/day. Second, we calculated the SEAD adherence with reverse scoring for the consumption of red meat and pork products and for potatoes, as higher consumption of these foods may have deleterious health effects [12,13,14,15]. For each of these two food components, subjects who were above the sex-specific median consumption were hereby scored 0 points, whereas those who were at or below it received 1 point. Third, to better understand the contribution of alcohol intake to the association between the SEAD and mortality, we further calculated the SEAD adherence without scoring wine consumption. Fourth, to reduce the potential residual confounding regarding morbidity, we adjusted the analyses for blood pressure- and lipid-lowering drugs, which are two of the most habitually used chronic medications. Fifth, to minimize the potential for reverse causation—health status influencing food consumption, rather than the opposite—we alternatively excluded from the analyses the subjects who died within the first 2 years of follow-up and those with prevalent morbidity (diabetes, cardiovascular disease, respiratory disease, musculoskeletal disease, cancer, or depression). And sixth, to test the consistency of our results with those for the main causes of death, we replicated the analyses for cardiovascular disease and cancer mortality.
Lastly, we conducted two additional analyses to address a potential methodological limitation: to minimize measurement error in SEAD adherence, we averaged the baseline and follow-up food consumption values for those participants who were followed at 3 years, but we used the baseline food consumption for those who were not, and who likely had worse health and higher mortality (for some of them had already died at this follow-up wave) than the subjects who remained in the cohort. To investigate how loss to follow-up may have affected our findings, we first compared the baseline characteristics between participants who were and were not followed at 3 years. Secondly, to test if the association of SEAD with mortality differed between the subjects who were and were not followed up at 3 years, we calculated the hazard ratio of the multiplicative interaction as follows: HRint = HR(SEAD+ Follow-up+)/[HR(SEAD+ Follow-up−) × HR(SEAD− Follow-up+)].
To assure that point estimates and their confidence intervals were representative of the Spanish population, descriptive data and regression analyses accounted for the complex sampling design, using the svy command in Stata® (StataCorp LLC), version 14.0.