EPIC-InterAct study
The EPIC-InterAct study is a case-cohort study nested within the prospective EPIC study [22]. In brief, EPIC includes 521,448 adults aged 25–79 years who were recruited between 1991 and 2000 at 23 centres in ten European countries participating in EPIC [23–25]. In the majority of the EPIC study centres, participants were recruited from the general population, with some exceptions [24]: the French cohort was based on members of a health insurance scheme for teachers; the Italian and Spanish cohorts included blood donors; participants from Utrecht (the Netherlands) and Florence (Italy) were recruited via a breast cancer screening programme; in Oxford (UK) half of the cohort were vegans, lacto-ovo vegetarians or fish eaters, and in France, Norway, Utrecht (the Netherlands) and Naples (Italy) only women were recruited [24]. Each EPIC centre obtained individual written informed consent and local ethics approval.
Within the InterAct project, incident cases of type 2 diabetes occurring in the EPIC cohort were ascertained and verified. All EPIC countries except Norway and Greece contributed to EPIC-InterAct (n = 455,680). Individuals without stored blood (n = 109,625) or without information on reported diabetes status (n = 5,821) were excluded, leaving 340,234 participants eligible for inclusion in EPIC-InterAct (corresponding to 3.99 million person-years follow-up).
Case-cohort construction and case ascertainment
A centre-stratified, random subcohort of 16,835 individuals was selected. After exclusion of 548 individuals with prevalent diabetes and 133 with uncertain diabetes status, the subcohort included 16,154 individuals for analysis. Because of random selection, this subcohort also included a random set of 778 individuals who had developed incident type 2 diabetes during follow-up (Fig. 1).
Ascertainment of incident type 2 diabetes involved a review of the existing EPIC datasets at each centre using multiple sources of evidence including self-report, linkage to primary-care registers, secondary-care registers, medication use (drug registers), hospital admissions and mortality data. Information from any follow-up visit or external evidence with a date later than the baseline visit was used. Rather than self-report, cases in Denmark and Sweden were identified via local and national diabetes and pharmaceutical registers [26] (www.ssi.dk/Sundhedsdataogit/Registre/Diabetesregisteret.aspx, accessed 11 October 2013) and hence all ascertained cases were considered to be verified. Some cases in centres other than Denmark and Sweden were based on only one source of information. To increase the specificity of the definition for these cases, we sought further evidence including review of individual medical records in some centres. Follow-up was censored at the date of diagnosis, 31 December 2007 or the date of death, whichever occurred first. Altogether, 12,403 verified incident cases were identified [22]. As stated earlier, 778 of these 12,403 incident cases were also subcohort members, due to the random selection of the subcohort. Thus, the EPIC-InterAct study involves 27,779 participants (16,154 subcohort members; 12,403 incident cases including 778 cases within the subcohort; Fig. 1).
Study population for the present analysis
Of these 27,779 participants, we excluded those from study centres in Italy and Umeå (Sweden) (n = 5,238) because these centres did not obtain specific intake data on diet soft drinks, breakfast cereals and dressing sauces (Italy) or diet soft drinks and cabbages (Umeå), which are important dietary components of the selected dietary-pattern scores. Specifically for analyses on DASH, the UK centres were excluded due to the unavailability of intake data on vegetable oils (n = 1,857). We further excluded participants with missing data on diet or covariates (n = 925), resulting in a final study population of 21,616 (9,682 cases; 12,595 subcohort members with an overlap of 661 subcohort members who had developed incident type 2 diabetes; Fig. 1). The excluded participants were more likely to be slightly older, women, slightly less overweight, less physically active, less educated and a current or former smoker and they were less likely to have a family history of diabetes. The proportion of participants with HbA1c ≥ 6.5% (47.5 mmol/mol) was slightly higher among excluded participants.
Dietary assessment and selection of dietary-pattern scores
Usual food intake during the past 12 months was assessed at baseline with the use of quantitative or semi-quantitative dietary questionnaires, which were developed and validated locally [24, 27]. The reproducibility of these questionnaires was generally good in the EPIC centres, while the relative validity ranged from moderate to good as also observed in other validity studies conducted by independent research groups [28, 29]. Individual food items were classified into food groups based on nutrient composition. Definitions and contents of the food groups considered for the present analysis are shown in electronic supplementary material (ESM) Table 1. Intakes of specific nutrients and total energy were derived with the standardised EPIC Nutrient Database [30].
Dietary patterns considered in this study were selected from the literature. Criteria for selection were availability of the necessary intake data to construct the dietary patterns in the EPIC-InterAct study and presence of scientific evidence indicating that the dietary pattern had a potential relevance for diabetes risk. We have selected two widely used diet-quality scores, the aHEI [5] and the DASH diet [31, 32]. The relation of the Mediterranean dietary pattern to diabetes in EPIC-InterAct has been specifically addressed previously [9] and hence not investigated here. We could not evaluate the HEI and the ONQI as it was not possible to appropriately reflect these indices with the EPIC-InterAct dietary data. We selected three RRR-derived dietary patterns: RRR1 was derived in the American Nurses' Health Study (NHS) using six inflammatory markers as responses [20]; RRR2 was identified in the German EPIC-Potsdam study with HbA1c, HDL-cholesterol, C-reactive protein (CRP) and adiponectin as responses [18]; RRR3 was identified in the British Whitehall II study with the HOMA-IR index as response [19]. An RRR dietary pattern derived with BMI as response along with fasting glucose, triacylglycerols, HDL-cholesterol and hypertension [21] was not considered because we aimed to assess the association of dietary patterns with diabetes independent of body size. Tables 1 and 2 show the individual dietary components of the dietary-pattern scores used in this study and their weighting in the calculation of the scores, respectively. A detailed description of the construction of the dietary-pattern scores in EPIC-InterAct is given in ESM Methods.
Table 1 Individual dietary components of the aHEI and the DASH dietary patterns considered in the analysis, EPIC-InterAct study
Table 2 Individual dietary components of the RRR dietary patterns considered in the analysis, EPIC-InterAct study
Assessment of other covariates
Standardised questionnaires were used at baseline to collect information on sociodemographic characteristics and lifestyle including age, education level, smoking status, occupational and leisure-time physical activity and history of previous illness. Height, weight and waist circumference of participants were obtained by trained staff during the baseline examination using standardised protocols [33]. However, for participants from France and some participants from Oxford (UK), self-reported anthropometric data were collected (4% of EPIC-InterAct study).
Statistical analysis
All dietary-pattern scores were transformed to z scores, based on subcohort distributions. Median dietary-pattern scores by country were computed to quantify country-specific adherence to the dietary patterns. The UK cohorts from Norfolk (population-based) and Oxford (high proportion of vegans, vegetarians and health-conscious individuals) were considered separately. We performed Cox proportional hazards analysis, weighted according to the Prentice method [34], to study the association between the dietary-pattern scores and the hazard of type 2 diabetes. Age was used as underlying time scale. Four models were applied, all stratified by study centre and integers of age (years), but with different levels of adjustment. Model 1 was adjusted for sex. Model 2 included further adjustment for physical activity (classified into ‘inactive’, ‘moderately inactive’, ‘moderately active’ and ‘active’ according to the validated Cambridge Physical Activity Index [35]), smoking status (never, former, current), educational level (none, primary, technical/professional, secondary, university) and total energy intake (continuous). We also applied additional adjustment for BMI (model 3) and BMI and waist circumference (model 4, both continuous).
Heterogeneity among countries in the association of the dietary-pattern scores with diabetes risk was studied by computing country-specific risk estimates and pooling these with random-effects meta-analyses. The two UK cohorts from Norfolk and Oxford were considered separately in the meta-analyses. As our aim was to verify associations of dietary patterns with diabetes, which should be done in independent cohorts, we did not use the Potsdam cohort in the meta-analysis for RRR2 because this pattern was derived in this cohort [18]. To explore potential sources of heterogeneity, country-specific mean age and BMI were related to the log-transformed HRs in subsequent meta-regression analyses [36].
Several sensitivity analyses were performed. To minimise reverse causality caused by a change in diet due to a prediabetic condition or chronic disease, we excluded participants with baseline HbA1c ≥ 6.5% (47.5 mmol/mol; 1.5% of the study population were missing values for HbA1c), incident cases diagnosed with diabetes within the first 2 years of follow-up and participants with baseline cardiovascular disease (myocardial infarction, stroke) or self-reported hypertension or hyperlipidaemia. To investigate potential effects of misreporting, we excluded participants in the top or bottom 1% of the energy intake/energy requirement ratio. Possible confounding by diabetes family history was addressed by further adjusting for history of diabetes in a first-degree relative (information not available in the Spanish centres, Oxford [UK] or Heidelberg [Germany]).
We investigated the importance of individual components of the dietary patterns for diabetes risk by sequentially subtracting components from the score. The change in estimate (CIE) was calculated as the difference between the HRs divided by the HR for the original score and multiplied by 100 (%).
Statistical analyses were performed with SAS (Version 9.2, Enterprise Guide 4.3; SAS Institute, Cary, NC, USA), except for meta-analyses and meta-regressions, which were conducted using Stata 12 (StataCorp, College Station, TX, USA).